Multiperspective Fusion (MPF) is a novel posttraining alignment framework for large language models (LLMs) developed in response to the growing need for easy bias mitigation. Built on top of the SAGED pipeline, an automated system for constructing bias benchmarks and extracting interpretable baseline distributions, MPF leverages multiperspective generations to expose and align biases in LLM outputs with nuanced, humanlike baselines. By decomposing baseline, such as sentiment distributions from HR professionals, into interpretable perspective components, MPF guides generation through sampling and balancing of responses, weighted by the probabilities obtained in the decomposition. Empirically, we demonstrate its ability to align LLM sentiment distributions with both counterfactual baselines (absolute equality) and the Human Resource baseline (biased for Top Univeristy), resulting in small KL divergence, reduction of calibration error and generalization to unseen questions. This shows that MPF offers a scalable and interpretable method for alignment and bias mitigation, compatible with deployed LLMs and requiring no extensive prompt engineering or finetuning.

Due to the rapid growth of social media platforms, these tools have become essential for monitoring information during ongoing disaster events. However, extracting valuable insights requires real-time processing of vast amounts of data. A major challenge in existing systems is their exposure to event-related biases, which negatively affects their ability to generalize to emerging events. While recent advancements in debiasing and causal learning offer promising solutions, they remain underexplored in the disaster event domain. In this work, we approach bias mitigation through a causal lens and propose a method to reduce event- and domain-related biases, enhancing generalization to future events. Our approach outperforms multiple baselines by up to +1.9% F1 and significantly improves a PLM-based classifier across three disaster classification tasks.

Modern large language models (LLMs) exhibit critical vulnerabilities to poison pill attacks—localized data poisoning that alters specific factual knowledge while preserving overall model utility. We systematically demonstrate these attacks exploit inherent architectural properties of LLMs, achieving 54.6% increased retrieval inaccuracy on long-tail knowledge versus dominant topics and up to 25.5% increase retrieval inaccuracy on compressed models versus original architectures. Through controlled mutations (e.g. temporal/spatial/entity alterations) and , our method induces localized memorization deterioration with negligible impact on models’ performance on regular standard benchmarks (e.g., <2% performance drop on MMLU/GPQA), leading to potential detection evasion. Our findings suggest: (1) Disproportionate vulnerability in long-tail knowledge may result from reduced parameter redundancy ; (2) Model compression may increase attack surfaces, with pruned/distilled models requiring 30% fewer poison samples for equivalent damage; (3) Associative memory enables both spread of collateral damage to related concepts and amplification of damage from simultaneous attack, particularly for dominant topics. These findings raise concerns over current scaling paradigms since attack costs are lowering while defense complexity is rising. Our work establishes poison pills as both a security threat and diagnostic tool, revealing critical security-efficiency trade-offs in language model compression that challenge prevailing safety assumptions.

On July 13, 2024, an assassination attempt was made on Republican presidential candidate Donald Trump during a rally in Pennsylvania. This event triggered widespread discourses on social media platforms. In this study, we analyze posts from X (formerly Twitter) collected during the week preceding and following the incident to examine the short-term impact of this political shock on public opinion and discourse. Our investigation is guided by three central research questions. First (RQ1), we assess how public stance toward Donald Trump evolved over time and varied across geographic regions. Second (RQ2), we apply causal inference methods to determine whether the assassination attempt itself significantly influenced public attitudes, independent of pre-existing political alignments. Third (RQ3), we conduct topic modeling to identify shifts in dominant themes of online discussions before and after the event. Integrating large language model-based stance detection, difference-in-differences estimation, and topic modeling, our findings reveal a marked surge in sympathetic responses toward Trump in the immediate aftermath of the attempt, suggesting a unifying effect that temporarily transcended ideological and regional divides.

Multi-source Opinion Summarization (M-OS) extends beyond traditional opinion summarization by incorporating additional sources of product metadata such as descriptions, key features, specifications, and ratings, alongside reviews. This integration results in comprehensive summaries that capture both subjective opinions and objective product attributes essential for informed decision-making. While Large Language Models (LLMs) have shown significant success in various Natural Language Processing (NLP) tasks, their potential in M-OS remains largely unexplored. Additionally, the lack of evaluation datasets for this task has impeded further advancements. To bridge this gap, we introduce M-OS-EVAL, a benchmark dataset for evaluating multi-source opinion summaries across seven key dimensions: fluency, coherence, relevance, faithfulness, aspect coverage, sentiment consistency, and specificity. Our results demonstrate that M-OS significantly enhances user engagement, as evidenced by a user study in which, on average, 87% of participants preferred M-OS over opinion summaries. Our experiments demonstrate that factually enriched summaries enhance user engagement. Notably, M-OS-PROMPTS exhibit stronger alignment with human judgment, achieving an average Spearman correlation of ρ = 0.74, which surpasses the performance of previous methodologies.

To address the instability of unguided reasoning paths in standard Chain-of-Thought prompting, recent methods guide large language models (LLMs) by first eliciting a single reasoning strategy. However, relying on just one strategy for each question can still limit performance across diverse tasks. We propose Diverge-to-Induce Prompting (DIP), a framework that first prompts an LLM to generate multiple diverse high-level rationales for each question. Each rationale is then elaborated into a detailed, step-by-step draft plan. Finally, these draft plans are induced into a final plan. DIP enhances zero-shot reasoning accuracy without reliance on resource-intensive sampling. Experiments show that DIP outperforms single-strategy prompting, demonstrating the effectiveness of multi-plan induction for prompt-based reasoning.

Reinforcement learning from AI feedback (RLAIF) is widely used for customizing the safety policies of large language models (LLMs) at scale. However, standard aligned LLMs are poorly suited in this setting, as their fixed alignment prevents adaptation to new policies. To address this, prior works have employed Helpful-Only LLMs (HOLLMs). Despite their effectiveness, no public framework exists for training or evaluating HOLLMs. In this paper, we present a comprehensive framework for developing HOLLMs that enable custom safety alignment. We first define the key attributes of a HOLLM and then propose Refusal-Avoidant Instruction Learning (RAIL), a novel training method that constructs HOLLMs from open-source datasets. We also introduce a comprehensive evaluation framework including a new benchmark: Helpfulness Evaluation without Limitations from Policies (HELP). Experiments show that the HOLLM achieves a 30.28% reduction in refusal rate over the strongest refusal-optimized baseline without compromising general capabilities. The HOLLM also achieves a 29.25% higher accuracy on HELP compared to the best-performing baseline. These results demonstrate that RAIL effectively cultivates the key attributes required of a HOLLM.

Translating natural language into formal language such as First-Order Logic (FOL) is a foundational challenge in NLP with wide-ranging applications in automated reasoning, misinformation tracking, and knowledge validation. In this paper, we introduce Natural Language to First-Order Logic (NL2FOL), a framework to autoformalize natural language to FOL step-by-step using Large Language Models (LLMs). Our approach addresses key challenges in this translation process, including the integration of implicit background knowledge. By leveraging structured representations generated by NL2FOL, we use Satisfiability Modulo Theory (SMT) solvers to reason about the logical validity of natural language statements. We present logical fallacy detection as a case study to evaluate the efficacy of NL2FOL. Being neurosymbolic, our approach also provides interpretable insights into the reasoning process and demonstrates robustness without requiring model fine-tuning or labeled training data. Our framework achieves good performance on multiple datasets–on the Logic dataset, NL2FOL achieves an F1-score of 78%, while generalizing effectively to the LogicClimate dataset with an F1-score of 80%.

Large language models (LLMs) often suffer from hallucinations, posing significant challenges for real-world applications. Confidence calibration, as an effective indicator of hallucination, is thus essential to enhance the trustworthiness of LLMs. Prior work mainly focuses on short-form tasks using a single response-level score (macro calibration), which is insufficient for long-form outputs that may contain both accurate and inaccurate claims. In this work, we systematically study atomic calibration, which evaluates factuality calibration at a fine-grained level by decomposing long responses into atomic claims. We further categorize existing confidence elicitation methods into discriminative and generative types, and propose two new confidence fusion strategies to improve calibration. Our experiments demonstrate that LLMs exhibit poorer calibration at the atomic level during long-form generation. More importantly, atomic calibration uncovers insightful patterns regarding the alignment of confidence methods and the changes of confidence throughout generation. This sheds light on future research directions for confidence estimation in long-form generation.

Quantifying the effects of textual interventions in social systems, such as reducing anger in social media posts to see its impact on engagement, is challenging. Real-world interventions are often infeasible, necessitating reliance on observational data. Traditional causal inference methods, typically designed for binary or discrete treatments, are inadequate for handling the complex, high-dimensional textual data. This paper addresses these challenges by proposing CausalDANN, a novel approach to estimate causal effects using text transformations facilitated by large language models (LLMs). Unlike existing methods, our approach accommodates arbitrary textual interventions and leverages text-level classifiers with domain adaptation ability to produce robust effect estimates against domain shifts, even when only the control group is observed. This flexibility in handling various text interventions is a key advancement in causal estimation for textual data, offering opportunities to better understand human behaviors and develop effective interventions within social systems.

Anomaly detection (AD) is essential in areas such as fraud detection, network monitoring, and scientific research. However, the diversity of data modalities and the increasing number of specialized AD libraries pose challenges for non-expert users who lack in-depth library-specific knowledge and advanced programming skills. To tackle this, we present AD-AGENT, an LLM-driven multi-agent framework that turns natural-language instructions into fully executable AD pipelines. AD-AGENT coordinates specialized agents for intent parsing, data preparation, library and model selection, documentation mining, and iterative code generation and debugging. Using a shared short-term workspace and a long-term cache, the agents integrate popular AD libraries like PyOD, PyGOD, and TSLib into a unified workflow. Experiments demonstrate that AD-AGENT produces reliable scripts and recommends competitive models across libraries. The system is open-sourced to support further research and practical applications in AD.

The growing adoption of Large Language Models (LLMs) has influenced the development of Small Language Models (SLMs) for on-device deployment across smartphones and edge devices, offering enhanced privacy, reduced latency, server-free functionality, and improved user experience. However, due to on-device resource constraints, SLMs undergo size optimization through compression techniques like quantization, which inadvertently introduce fairness, ethical and privacy risks. Critically, quantized SLMs may respond to harmful queries directly, without requiring adversarial manipulation, raising significant safety and trust concerns. To address this, we propose LiteLMGuard, an on-device guardrail that provides real-time, prompt-level defense for quantized SLMs. Additionally, our guardrail is designed to be model-agnostic such that it can be seamlessly integrated with any SLM, operating independently of underlying architectures. Our LiteLMGuard formalizes deep learning (DL)-based prompt filtering by leveraging semantic understanding to classify prompt answerability for SLMs. Built on our curated Answerable-or-Not dataset, LiteLMGuard employs ELECTRA as the candidate model with 97.75% answerability classification accuracy. The on-device deployment of LiteLMGuard enabled real-time offline filtering with over 85% defense-rate against harmful prompts (including jailbreak attacks), 94% filtering accuracy and ~135 ms average latency. These results demonstrate LiteLMGuard as a lightweight robust defense mechanism for effectively and efficiently securing on-device SLMs against Open Knowledge Attacks.

The rapid growth of social media platforms has led to concerns about radicalization, filter bubbles, and content bias. Existing approaches to classifying ideology are limited in that they require extensive human effort, the labeling of large datasets, and are not able to adapt to evolving ideological contexts. This paper explores the potential of Large Language Models (LLMs) for classifying the political ideology of online content through in-context learning (ICL). Our extensive experiments involving demonstration selection in label-balanced fashion, conducted on three datasets comprising news articles and YouTube videos, reveal that our approach significantly outperforms zero-shot and traditional supervised methods. Additionally, we evaluate the influence of metadata (e.g., content source and descriptions) on ideological classification and discuss its implications. Finally, we show how providing the source for political and non-political content influences the LLM’s classification.

Evaluating Theory of Mind (ToM) in Large Language Models (LLMs) is an important area of research for understanding the social intelligence of AI. Recent ToM benchmarks have made significant strides in enhancing the complexity, comprehensiveness, and practicality of evaluation. However, while the focus has been on constructing “more difficult” or “more comprehensive” tasks, there has been insufficient systematic analysis of the structural factors that inherently determine the difficulty of ToM reasoning—that is, “what” makes reasoning difficult. To address this challenge, we propose a new dataset generation framework for ToM evaluation named AnaToM. To realize an “Anatomy of Difficulty” in ToM reasoning, AnaToM strictly controls structural parameters such as the number of entities and the timeline in a story. This parameter control enables the isolation and identification of factors affecting the ToM of LLMs, allowing for a more precise examination of their reasoning mechanisms. The proposed framework provides a systematic methodology for diagnosing the limits of LLM reasoning abilities and offers new guidelines for future benchmark design.

We propose an information-theoretic formalization of the distinction between two fundamental AI safety failure modes: deceptive alignment and goal drift. While both can lead to systems that appear misaligned, we demonstrate that they represent distinct forms of information divergence occurring at different interfaces in the human-AI system. Deceptive alignment creates entropy between an agent’s true goals and its observable behavior, while goal drift, or confusion, creates entropy between the intended human goal and the agent’s actual goal. Though often observationally equivalent, these failures necessitate different interventions. We present a formal model and an illustrative thought experiment to clarify this distinction. We offer a formal language for re-examining prominent alignment challenges observed in Large Language Models (LLMs), offering novel perspectives on their underlying causes.

India, a country with a large population, possesses two official and twenty-two scheduled languages, making it the most linguistically diverse nation. Despite being one of the scheduled languages, Santali remains a low-resource language. Although Ol Chiki is recognized as the official script for Santali, many continue to use Bengali, Devanagari, Odia, and Roman scripts. In tribute to the upcoming centennial anniversary of the Ol Chiki script, we present an Automatic Speech Recognition for Santali in the Ol Chiki script. Our approach involves cross-lingual transfer learning by utilizing the Whisper framework pre-trained in Bengali and Hindi on the Santali language, using Ol Chiki script transcriptions. With the adoption of the Bengali pre-trained framework, we achieved a Word Error Rate (WER) score of 28.47%, whereas the adaptation of the Hindi pre-trained framework resulted in a score of 34.50% WER. These outcomes were obtained using the Whisper Small framework.

Large language models (LLMs) are increasingly used in everyday tools and applications, raising concerns about their potential influence on political views. While prior research has shown that LLMs often exhibit measurable political biases-frequently skewing toward liberal or progressive positions-key gaps remain. Most existing studies evaluate only a narrow set of models and languages, leaving open questions about the generalizability of political biases across architectures, scales, and multilingual settings. Moreover, few works examine whether these biases can be actively controlled.In this work, we address these gaps through a large-scale study of political orientation in modern open-source instruction-tuned LLMs. We evaluate seven models, including LLaMA-3.1, Qwen-3, and Aya-Expanse, across 14 languages using the Political Compass Test with 11 semantically equivalent paraphrases per statement to ensure robust measurement. Our results reveal that larger models consistently shift toward libertarian-left positions, with significant variations across languages and model families. To test the manipulability of political stances, we utilize a simple center-of-mass activation intervention technique and show that it reliably steers model responses toward alternative ideological positions across multiple languages. Our code is publicly available at https://github.com/d-gurgurov/Political-Ideologies-LLMs.

Rapid advancements in Visual Language Models (VLMs) have transformed multimodal understanding but are often constrained by generating English responses regardless of the input language. This phenomenon has been termed as Image-induced Fidelity Loss (IFL) and stems from limited multimodal multilingual training data. To address this, we propose a continuous multilingual integration strategy that injects text-only multilingual data during visual instruction tuning, preserving the language model’s original multilingual capabilities. Extensive evaluations demonstrate that our approach significantly improves linguistic fidelity across languages without degradation in visual performance. We also explore model merging, which improves language fidelity but comes at the cost of visual performance. In contrast, our core method achieves robust multilingual alignment without trade-offs, offering a scalable and effective path to mitigating IFL for global VLM adoption.

We propose XL-DURel, a finetuned, multilingual Sentence Transformer model optimized for ordinal Word-in-Context classification. We test several loss functions for regression and ranking tasks managing to outperform previous models on ordinal and binary data with a ranking objective based on angular distance in complex space. We further show that binary WiC can be treated as a special case of ordinal WiC and that optimizing models for the general ordinal task improves performance on the more specific binary task. This paves the way for a unified treatment of WiC modeling across different task formulations.

Response consistency-based, reference-free hallucination detection (RFHD) methods do not depend on internal model states, such as generation probabilities or gradients, which Grey-box models typically rely on but are inaccessible in closed-source LLMs. However, their inability to capture query-response alignment patterns often results in lower detection accuracy. Additionally, the lack of large-scale benchmark datasets spanning diverse domains remains a challenge, as most existing datasets are limited in size and scope. To this end, we propose HalluCounter, a novel reference-free hallucination detection method that utilizes both response-response and query-response consistency and alignment patterns. This enables the training of a classifier that detects hallucinations and provides a confidence score and an optimal response for user queries. Furthermore, we introduce HalluCounterEval, a benchmark dataset comprising both synthetically generated and human-curated samples across multiple domains. Our method outperforms state-of-the-art approaches by a significant margin, achieving over 90% average confidence in hallucination detection across datasets.

The social biases inherent in language models necessitate a critical analysis of their social influence in many linguistic situations because of their extensive use. This study investigates gender bias in Bengali language models by highlighting the unique linguistic challenges posed by its complex morphology, dialectical variations, and distinctions between formal and informal language versions. While prior research on social bias in Bengali has provided foundational insights, it has not adequately addressed the nuances arising from these variations. This research extends to measuring intrinsic gender bias in both formal and informal Bengali, analyzing the impact of context lengths on bias detection, and proposing modifications to existing techniques to enhance their applicability to Bengali. Addressing these, the study aims to contribute to developing more inclusive and representative bias measurement methodologies for underrepresented languages. We open the source code and data at https://github.com/kraritt/b-bias-ctext.

The development of large language models (LLMs) has resulted in significant transformations in the field of chemistry, with potential applications in molecular science. Traditionally, the exploration of methods to enhance pre-trained general-purpose LLMs has focused on techniques like supervised fine-tuning (SFT) and retrieval-augmented generation (RAG), to improve model performance and tailor them to specific applications. General purpose extended approaches are being researched, but their adaptation within the chemical domain has not progressed significantly. This study advances the application of LLMs in molecular science by exploring SFT of LLMs, and developing RAG and multimodal models, incorporating molecular embeddings derived from molecular fingerprints and other properties. Experimental results show that a multimodal model with fingerprint inputs to the LLM achieved the highest overall performance. For molecular representation based on SMILES notation, fingerprints effectively capture the structural information of molecular compounds, demonstrating the applicability of LLMs in drug discovery research.

Large language models can be quantized to reduce inference time latency, model size, and energy consumption, thereby delivering a better user experience at lower cost. A challenge exists to deliver quantized models with minimal loss of accuracy in reasonable time, and in particular to do so without requiring mechanisms incompatible with specialized inference accelerators. Here, we demonstrate a simple, end-to-end quantization-aware training approach that, with an increase in total model training budget of less than 0.1%, outperforms the leading published quantization methods by large margins on several modern benchmarks, with both base and instruct model variants. The approach easily generalizes across different model architectures, can be applied to activations, cache, and weights, and requires the introduction of no additional operations to the model other than the quantization itself.

Transformer models often exhibit brittle extrapolation, failing on inputs that are longer or structurally more complex than those seen during training. We introduce Counter-Example–Driven Curricula (CEDC), an automated framework that improves model robustness by iteratively focusing on its own failures. At each step, CEDC uses the current model to generate a diverse set of candidate problems, employs a fast, executable verifier to identify incorrect predictions (counter‐examples), and then fine‐tunes the model on a dataset enriched with these discovered failures. We evaluate CEDC on a suite of algorithmic and natural language tasks, including integer addition, sorting, Dyck‐2 language recognition, and three text classification benchmarks. Compared to static training and standard curriculum learning baselines, CEDC achieves up to 30× greater length extrapolation, is 3.75× more computationally efficient than uniform data augmentation, and requires no manual difficulty heuristics. We provide a detailed analysis of the counter‐examples, showing how the curriculum naturally adapts to target progressively more complex error modes. Our findings establish verifier‐guided, failure‐driven learning as a simple, powerful, and efficient paradigm for enhancing the generalization capabilities of Transformer models.

As large language models (LLMs) continue to scale, the memory footprint of Key-Value (KV) caches during inference has become a significant bottleneck. Existing approaches primarily focus on compressing KV caches within a single prompt or reusing shared prefixes or frequently occurred text segments across prompts. However, such strategies are limited in scenarios where prompts are semantically similar but lexically different, which frequently occurs in tasks such as multi-document summarization and conversational agents. We propose SemShareKV, a KV cache sharing and compression framework that accelerates LLM inference by reusing KV cache in semantically similar prompts. Instead of relying on exact token matches, SemShareKV applies fuzzy token matching using Locality-Sensitive Hashing (LSH) on token embeddings and incorporates Rotary Position Embedding (RoPE) to better preserve positional information. By selectively reusing relevant KV pairs from a reference prompt’s cache, SemShareKV reduces redundant computation while maintaining output quality. Experiments on diverse summarization datasets show up to 6.25× speedup and 42% lower GPU memory usage with 5k tokens input, with negligible quality degradation. These results highlight the potential of semantic-aware cache sharing for efficient LLM inference.

Dominant sequence models like the Transformer represent structure implicitly through dense attention weights, incurring quadratic complexity. We propose RewriteNets, a novel neural architecture built on an alternative paradigm: explicit, parallel string rewriting. Each layer in a RewriteNet contains a set of learnable rules. For each position in an input sequence, the layer performs four operations: (1) fuzzy matching of rule patterns, (2) conflict resolution via a differentiable assignment operator to select non-overlapping rewrites, (3) application of the chosen rules to replace input segments with output segments of potentially different lengths, and (4) propagation of untouched tokens. While the discrete assignment of rules is non-differentiable, we employ a straight-through Gumbel-Sinkhorn estimator, enabling stable end-to-end training. We evaluate RewriteNets on algorithmic, compositional, and string manipulation tasks, comparing them against strong LSTM and Transformer baselines. Results show that RewriteNets excel at tasks requiring systematic generalization (achieving 98.7% accuracy on the SCAN benchmark’s length split) and are computationally more efficient than Transformers. We also provide an analysis of learned rules and an extensive ablation study, demonstrating that this architecture presents a promising direction for sequence modeling with explicit structural inductive biases.

Large Language Models (LLMs) have shown strong capabilities in Question Answering (QA), but their effectiveness in high-stakes, closed-domain settings is often constrained by hallucinations and limited handling of vague or underspecified queries. These challenges are especially pronounced in Vietnamese, a low-resource language with complex syntax and strong contextual dependence, where user questions are often short, informal, and ambiguous. We introduce the Unified Retrieval Agent-Based System (URASys), a QA framework that combines agent-based reasoning with dual retrieval under the Just Enough principle to address standard, ambiguous, and unanswerable questions in a unified manner. URASys performs lightweight query decomposition and integrates document retrieval with a question–answer layer via a two-phase indexing pipeline, engaging in interactive clarification when intent is uncertain and explicitly signaling unanswerable cases to avoid hallucination. We evaluate URASys on Vietnamese and English QA benchmarks spanning single-hop, multi-hop, and real-world academic advising tasks, and release new dual-language ambiguous subsets for benchmarking interactive clarification. Results show that URASys outperforms strong retrieval-based baselines in factual accuracy, improves unanswerable handling, and achieves statistically significant gains in human evaluations for clarity and trustworthiness.

Grammatical Error Correction (GEC) is a fundamental task in Natural Language Processing that focuses on automatically detecting and correcting grammatical errors in text. In this paper, we present a novel approach for GEC for Gujarati. Gujarati is an Indian language spoken by over 55 million people worldwide. Our approach combines a large language model with non-parametric memory modules to address the low-resource challenge. We have evaluated our system on human-annotated and synthetic datasets. The overall result indicates promising results for Gujarati. The proposed approach is generic enough to be adopted by other languages. Furthermore, we release a publicly available evaluation dataset for Gujarati GEC along with an adapted version of the ERRANT framework to enable error-type-wise evaluation in Gujarati.

Recent studies have combined Large Language Models (LLMs) with Knowledge Graphs (KGs) to enhance reasoning, improving inference accuracy without additional training while mitigating hallucination. However, existing frameworks still suffer two practical drawbacks: they must be re-tuned whenever the KG or reasoning task changes, and they depend on a single, high-capacity LLM for reliable (i.e., trustworthy) reasoning. To address this, we introduce R2-KG, a plug-and-play, dual-agent framework that separates reasoning into two roles: an Operator (a low-capacity LLM) that gathers evidence and a Supervisor (a high-capacity LLM) that makes final judgments. This design is cost-efficient for LLM inference while still maintaining strong reasoning accuracy. Additionally, R2-KG employs an Abstention mechanism, generating answers only when sufficient evidence is collected from KG, which significantly enhances reliability. Experiments across five diverse benchmarks show that R2-KG consistently outperforms baselines in both accuracy and reliability, regardless of the inherent capability of LLMs used as the operator. Further experiments reveal that the single-agent version of R2-KG, equipped with a strict self-consistency strategy, achieves significantly higher-than-baseline reliability with reduced inference cost but increased abstention rate in complex KGs. Our findings establish R2-KG as a flexible and cost-effective solution for KG-based reasoning, reducing reliance on high-capacity LLMs while ensuring trustworthy inference.

As an effective way to address data scarcity problem, data augmentation has received significant interest in low-resource neural machine translation, while the latter has the potential to reduce digital divide and benefit out of domain translation. However, the existing works mainly focus on how to generate the synthetic data, while the synthetic data quality and the way we use the synthetic data also matter. In this paper, we give a systematic analysis of data augmentation for low-resource neural machine translation that encompasses all the three aspects. We show that with careful control of the synthetic data quality and the way we use the synthetic data, the performance can be greatly boosted even with the same method to generate the synthetic data.

Business process modeling has traditionally depended on manual efforts or rigid rule-based techniques, limiting scalability and flexibility. Recent progress in Large Language Models (LLMs) enables automatic generation of process models from text, yet a systematic evaluation remains lacking. This paper explores the ability of LLMs to produce structurally and semantically valid business process workflows using five approaches: zero-shot, zero-shot CoT, few-shot, few-shot CoT, and fine-tuning. We assess performance under increasing control-flow complexity (e.g., nested gateways, parallel branches) using the MaD dataset, and introduce a masked-input setting to test semantic robustness. Results show that while fine-tuning achieves the best accuracy, few-shot CoT excels in handling complex logic and incomplete inputs. These findings reveal the strengths and limits of LLMs in process modeling and offer practical guidance for enterprise Business Process Management (BPM) automation.

Recent progress in Large Language Model (LLM) technology has changed our role in interacting with these models. Instead of primarily testing these models with questions we already know answers to, we are now using them for queries where the answers are unknown to us, driven by human curiosity. This shift highlights the growing need to understand curiosity-driven human questions – those that are more complex, open-ended, and reflective of real-world needs. To this end, we present Quriosity, a collection of 13K naturally occurring questions from three diverse sources: human-to-search-engine queries, human-to-human interactions, and human-to-LLM conversations. Our comprehensive collection enables a rich understanding of human curiosity across various domains and contexts. Our analysis reveals a significant presence of causal questions (up to 42%) in the dataset, for which we develop an iterative prompt improvement framework to identify all causal queries and examine their unique linguistic properties, cognitive complexity and source distribution. We also lay the groundwork for exploring efficient identifiers of causal questions, providing six efficient classification models.

Training effective text rerankers is crucial for information retrieval. Two strategies are widely used: contrastive learning (optimizing directly on ground-truth labels) and knowledge distillation (transferring knowledge from a larger reranker). While both have been studied extensively, a clear comparison of their effectiveness for training cross-encoder rerankers under practical conditions is needed.This paper empirically compares these strategies by training rerankers of different sizes (0.5B, 1.5B, 3B, 7B) and architectures (Transformer, Recurrent) using both methods on the same data, with a strong contrastive learning model acting as the distillation teacher. Our results show that knowledge distillation generally yields better in-domain and out-of-domain ranking performance than contrastive learning when distilling from a more performant teacher model. This finding is consistent across student model sizes and architectures. However, distilling from a teacher of the same capacity does not provide the same advantage, particularly for out-of-domain tasks. These findings offer practical guidance for choosing a training strategy based on available teacher models. We recommend using knowledge distillation to train smaller rerankers if a larger, more performant teacher is accessible; in its absence, contrastive learning remains a robust baseline. Our code implementation is made available to facilitate reproducbility.

Natural Language Processing is a branch of artificial intelligence that enables man- machine interactions through regional languages. In Kannada, there are two types of Sandhi: Kannada Sandhi and Sanskrit Sandhi. A morph-phonemic word “Sandhi” is created when two words or distinct morphemes are joined or combined. Conversely, Sandhi word splitting reverses this process. Rules governing Sandhi exist across all the Dravidian languages. A rule-based method has been developed to split Sanskrit Sandhi words into their components within Kannada sentences. Once the Sanskrit Sandhi (SS) words are split, the type of Sandhi is also identified, facilitating accurate translation of the Sanskrit Sandhi words into English. This paper discusses seven types of SS words: SavarNadeergha, YaN, GuNa, Vruddhi, Jatva, Shchutva and Anunasika Sandhi. The identified split points adhere precisely to Sandhi rules. A dataset of 4900 SanskritSandhi words found in Kannada sentences was used to evaluate the proposed method, which achieved an accuracy of 90.03% for Sanskrit Sandhi Identification and 85.87% for reliable English Translation. This work has potential applications in other Dravidian languages.

There is growing research interest in Questions Under Discussion (QUD), a linguistic framework for representing discourse in the form of natural language question-answer pairs, which are more easily understandable and have been found useful in several applications. Our goal in this work is to improve on the quality of automatic QUD generation. As a way to sidestep the paucity of data currently, we propose a reinforcement learning-based approach using the Group Relative Policy Optimisation (GRPO) objective for LLM post-training on the task. To get there, we: (i) carefully investigated five promising methods for reference-free automatic QUD evaluation, (ii) proposed a novel prompting strategy, SCRS, involving ranking and scoring with structured outputs that enables QUD evaluation close to the human upperbound, (iii) leveraged findings from (i) with (ii) for the knowledge distillation from a very large LLM to obtain a more resource-efficient reward model, and which (iv) we then used in the GRPO post-training for 3B LLMs on the QUD generation task. Our QUD generators give overall higher-quality QUDs compared to the SOTA which is based on supervised fine-tuning; all of these are achieved using only three annotated exemplars in the few-shot prompting for evaluation, and without the use of any other annotated questions for training the QUD generators. Our code, models, and annotated examples can be found at https://github.com/hankelvin/grpo_qud_generation.

Large language models (LLMs) are increasingly integrated into academic workflows, with many conferences and journals permitting their use for tasks such as language refinement and literature summarization. However, their use in peer review remains prohibited due to concerns around confidentiality breaches, hallucinated content, and inconsistent evaluations. As LLM-generated text becomes more indistinguishable from human writing, there is a growing need for reliable attribution mechanisms to preserve the integrity of the review process. In this work, we evaluate topic-based watermarking (TBW), a semantic-aware technique designed to embed detectable signals into LLM-generated text. We conduct a systematic assessment across multiple LLM configurations, including base, few-shot, and fine-tuned variants, using authentic peer review data from academic conferences. Our results show that TBW maintains review quality relative to non-watermarked outputs, while demonstrating robust detection performance under paraphrasing. These findings highlight the viability of TBW as a minimally intrusive and practical solution for LLM attribution in peer review settings.

In this work, we address the problem of fine-grained traceback of emotional and manipulation characteristics from synthetically manipu- lated speech. We hypothesize that combining semantic–prosodic cues captured by Speech Foundation Models (SFMs) with fine-grainedspectral dynamics from auditory representations can enable more precise tracing of both emotion and manipulation source. To validate this hypothesis, we introduce MiCuNet, a novel multitask framework for fine-grained tracing of emotional and manipulation attributes in synthetically generated speech. Our approach integrates SFM embeddings with spectrogram-based auditory features through a mixed-curvature projection mechanism that spans Hyperbolic, Euclidean, and Spherical spaces guided by a learnable temporal gating mechanism. Our proposed method adopts a multitask learning setup to simultaneously predict original emotions, manipulated emotions, and manipulation sources on the Emo-Fake dataset (EFD) across both English and Chinese subsets. MiCuNet yields consistent improvements, consistently surpassing conventional fusion strategies. To the best of our knowledge, this work presents the first study to explore a curvature-adaptive framework specifically tailored for multitask tracking in synthetic speech.

Probing the multilingual knowledge of linguistic structure in LLMs, often characterized as sequence labeling, faces challenges with maintaining output templates in current text-to-text prompting strategies. To solve this, we introduce a decomposed prompting approach for sequence labeling tasks. Diverging from the single text-to-text prompt, our prompt method generates for each token of the input sentence an individual prompt which asks for its linguistic label. We test our method on the Universal Dependencies part-of-speech tagging dataset for 38 languages, using both English-centric and multilingual LLMs. Our findings show that decomposed prompting surpasses the iterative prompting baseline in efficacy and efficiency under zero- and few-shot settings. Moreover, our analysis of multilingual performance of English-centric LLMs yields insights into the transferability of linguistic knowledge via multilingual prompting.

Although there has been growing interest in the self-correction capability of Large Language Models (LLMs), there are varying conclusions about its effectiveness.Prior research has largely concentrated on intrinsic self-correction, extrinsic self-correction, particularly the interplay between internal knowledge and external feedback, remains underexplored. In this paper, we aim to comprehensively investigate the underlying mechanism of moral self-correction by addressing a fundamental question: is moral self-correction an innate capability of LLMs? Specifically, we conduct: (1) a behavioral analysis of LLMs’ moral sensitivity based on a self-distinguishing task; and (2) a mechanistic analysis of the hidden states to examine how key components of self-correction, such as Chain-of-Thought (CoT) and external feedback, interact to facilitate moral self-correction. Drawing on empirical evidence from both behavioral and mechanistic analyses, we demonstrate that moral self-correction is not an inherent capability of LLMs, as they are neither morally sensitive nor able to effectively incorporate external feedback during the self-correction process.

Large language models (LLMs) hold promise for advancing patient–provider communication, yet a persistent gap remains between benchmark-driven model development and the realities of clinical practice. This work presents a systematic, clinically grounded review of text-based medical datasets for LLM training and evaluation. We propose a scenario-based taxonomy derived from established clinical frameworks to map major knowledge-based and conversation-based corpora against core communication scenarios. We further synthesize core communication skills from gold-standard clinical assessment instruments and meta-analyze state-of-the-art medical LLM performance, highlighting how dataset properties, fine-tuning strategies, and evaluation metrics shape both knowledge acquisition and communicative competence. To empirically validate these findings, we conducted controlled fine-tuning experiments across representative LLMs, demonstrating that data composition and scenario alignment critically affect model performance. Our findings highlight the urgent need for scenario-rich datasets and standardized, human-centered evaluation protocol to advance clinically relevant medical LLMs.

Recent advances in AI-generated video have shown strong performance on text-to-video tasks, particularly for short clips depicting a single scene. However, current models struggle to generate longer videos with coherent scene transitions, primarily because they cannot infer when a transition is needed from the prompt. Most open-source models are trained on datasets consisting of single-scene video clips, which limits their capacity to learn and respond to prompts requiring multiple scenes. Developing scene transition awareness is essential for multi-scene generation, as it allows models to identify and segment videos into distinct clips by accurately detecting transitions. To address this, we introduce the Transition-Aware Video (TAV) dataset with multi-scene clips and captions that explicitly state scene segmentation and transition structure. Our focus is on how prompt semantics and dataset annotations about temporal context affect text-to-video generation. Post-training on TAV improves alignment between the scene count implied by prompt and the scene count produced by the model, while preserving visual quality.

In this work, we propose DoubleDipper, a novel In-Context-Learning method that automatically generates few-shot examples for several QA tasks by _recycling_ contexts. Specifically, given an input context (1-3k tokens) and a query, we generate additional query-output pairs from the given context as few-shot examples, while introducing the context only once. This ensures that the demonstrations are leveraging the same context as the target query while only adding a small number of tokens to the prompt. We further enhance each demonstration by instructing the model to _explicitly_ identify the relevant paragraphs before the answer, which improves performance while providing fine-grained attribution to the answer source. We apply our method on multiple LLMs and obtain substantial improvements (+16 absolute points on average across models) on various QA datasets. Surprisingly, despite introducing only single-hop ICL examples, LLMs successfully generalize to multi-hop QA using our approach.

The discipline of physics stands as a cornerstone of human intellect, driving the evolution of technology and deepening our understanding of the fundamental principles of the cosmos. Contemporary literature includes some works centered on the task of solving physics problems—a crucial domain of natural language reasoning. In this paper, we evaluate the performance of frontier LLMs in solving physics problems, both mathematical and descriptive. We also employ a plethora of inference-time techniques and agentic frameworks to improve the performance of the models. This includes the verification of proposed solutions in a cumulative fashion by other, smaller LLM agents, and we perform a comparative analysis of the performance that the techniques entail. There are significant improvements when the multi-agent framework is applied to problems that the models initially perform poorly on. Furthermore, we introduce a new evaluation benchmark for physics problems, PhysicsEval, consisting of 19,609 problems sourced from various physics textbooks and their corresponding correct solutions scraped from physics forums and educational websites. Our code and data are publicly available at https://github.com/areebuzair/PhysicsEval.

Large language models (LLMs) can suggest missing elements from items listed in a prompt, which can be used for list completion or similar item recommendation. However, their performance degrades when they are exposed to too many items, as they start to suggest items already included in the input list. This occurs at around 100 items for mid-2024 flagship LLMs. We evaluate this phenomenon on both synthetic problems (e.g., finding missing numbers in a given range of shuffled integers) and realistic movie recommendation scenarios. We refer to this issue as “attention overflow”, as avoiding repetition requires attending to all items simultaneously. Although iterative loops can mitigate this problem, their costs increase with the repetition rate, affecting the language models’ ability to derive novelty from lengthy inputs.

Language and culture are deeply intertwined, yet it has been unclear how and where multilingual large language models encode culture. Here, we build on an established methodology for identifying language-specific neurons to localize and isolate culture-specific neurons, carefully disentangling their overlap and interaction with language-specific neurons. To facilitate our experiments, we introduce MUREL, a curated dataset of 85.2 million tokens spanning six different cultures. Our localization and intervention experiments show that LLMs encode different cultures in distinct neuron populations, predominantly in upper layers, and that these culture neurons can be modulated largely independently of language-specific neurons or those specific to other cultures. These findings suggest that cultural knowledge and propensities in multilingual language models can be selectively isolated and edited, with implications for fairness, inclusivity, and alignment. Code and data are available at https://github.com/namazifard/Culture_Neurons

Causal reasoning is central to language understanding, yet remains under-resourced in Bangla. In this paper, we introduce the first large-scale dataset for causal inference in Bangla, consisting of over 11663 sentences annotated for causal sentence types (explicit, implicit, non-causal) and token-level spans for causes, effects, and connectives. The dataset captures both simple and complex causal structures across diverse domains such as news, education, and health. We further benchmark a suite of state-of-the-art instruction-tuned large language models, including LLaMA 3.3 70B, Gemma 2 9B, Qwen 32B, and DeepSeek, under zero-shot and three-shot prompting conditions. Our analysis reveals that while LLMs demonstrate moderate success in explicit causality detection, their performance drops significantly on implicit and span-level extraction tasks. This work establishes a foundational resource for Bangla causal understanding and highlights key challenges in adapting multilingual LLMs for structured reasoning in low-resource languages.

International enterprises, organizations, and hospitals collect large amounts of multi-modal data stored in databases, text documents, images, and videos. While there has been recent progress in the separate fields of multi-modal data exploration as well as in database systems that automatically translate natural language questions to database query languages, the research challenge of querying both structured databases and unstructured modalities (e.g., texts, images) in natural language remains largely unexplored.In this paper, we propose M²EX, a system that enables multi-modal data exploration via language agents. Our approach is based on the following research contributions: (1) Our system is inspired by a real-world use case that enables users to explore multi-modal information systems. (2) M²EX leverages an LLM-based agentic AI framework to decompose a natural language question into subtasks such as text-to-SQL generation and image analysis and to orchestrate modality-specific experts in an efficient query plan. (3) Experimental results on multi-modal datasets, encompassing relational data, text, and images, demonstrate that our system outperforms state-of-the-art multi-modal exploration systems, excelling in both accuracy and various performance metrics, including query latency, API costs, and planning efficiency, thanks to the more effective utilization of the reasoning capabilities of LLMs.

Generation of Artificial Intelligence (AI) texts in important works has become a common practice that can be used to misuse and abuse AI at various levels. Traditional AI detectors often rely on document-level classification, which struggles to identify AI content in hybrid or slightly edited texts designed to avoid detection, leading to concerns about the model’s efficiency, which makes it hard to distinguish between human-written and AI-generated texts. A sentence-level sequence labeling model proposed to detect transitions between human- and AI-generated text, leveraging nuanced linguistic signals overlooked by document-level classifiers. By this method, detecting and segmenting AI and human-written text within a single document at the token-level granularity is achieved. Our model combines the state-of-the-art pre-trained Transformer models, incorporating Neural Networks (NN) and Conditional Random Fields (CRFs). This approach extends the power of transformers to extract semantic and syntactic patterns, and the neural network component to capture enhanced sequence-level representations, thereby improving the boundary predictions by the CRF layer, which enhances sequence recognition and further identification of the partition between Human- and AI-generated texts. The evaluation is performed on two publicly available benchmark datasets containing collaborative human and AI-generated texts. Our experimental comparisons are with zero-shot detectors and the existing state-of-the-art models, along with rigorous ablation studies to justify that this approach, in particular, can accurately detect the spans of AI texts in a completely collaborative text.

Full-duplex spoken dialogue models, which process audio input and output simultaneously, have been actively studied for their ability to naturally model turn-taking and non-verbal phenomena in addition to generating responses. Although these models enable natural conversational flow, they lack mechanisms for language understanding and dialogue management, making them difficult to apply to task-oriented dialogue systems. We propose a method for incorporating dialogue state tracking in task-oriented dialogue into Moshi, aiming to achieve a multi-channel, full-duplex task-oriented spoken dialogue model. We evaluated the proposed method on JMultiWOZ, a benchmark corpus for Japanese task-oriented dialogue, focusing on dialogue state tracking and response generation.

Identifying rhetorical roles in legal judgments is a foundational step for automating legal reasoning, summarization, and retrieval. In this paper, we propose a novel Sequential Tolerance-based Classifier (SeqTNS) for rhetorical role classification in Indian legal documents. The proposed classifier leverages semantic similarity and contextual dependencies by using label sequence aware BiLSTMs on top of word embeddings from finetuned InLegalBERT model. These enriched embeddings are clustered into tolerance classes via a tolerance relation using a cosine distance threshold,enabling the model to make flexible, similarity-based predictions. We evaluate SeqTNS on two benchmark datasets annotated with thirteen and seven rhetorical roles, respectively. The proposed method outperforms fine-tuned transformer baselines (LegalBERT, InLegalBERT) as well as the previously developed tolerance relation-based (TNS) model, achieving a weighted F1 score of 0.78 on thirteen class dataset and a macro F1 of 0.83 on the seven class dataset, while reducing training time by 39-40% compared to state of the art BiLSTM-CRF models. The larger of our two datasets is substantial, containing over 40,000 sentences and 1.3M tokens, and serves as a challenging real world benchmark. Additionally, we use LIME for explainability and t-SNE to validate the coherence of tolerance-based clusters.

Recent studies have shown that prompting large language models (LLMs) with role-playing personas can enhance their reasoning capabilities. While the benefits of role-playing personas in reasoning tasks are widely recognized, it remains uncertain whether a persona aligned with the given dataset can consistently achieve these improvements. In this work, we empirically investigate the potential drawbacks of using dataset-aligned personas (referred to as **coarsely aligned personas**) and introduce Jekyll & Hyde, a novel framework that enhances reasoning robustness by ensembling solutions from both role-playing and neutral (non-persona) prompts.Jekyll & Hyde first predicts an instance-specific persona tailored to each query using an LLM, then generates answers with both persona and neutral prompts, and finally selects the superior output through an LLM-based evaluator.Experimental results claim that across twelve widely used natural language reasoning datasets and three backbone large language models, Jekyll & Hyde consistently outperforms single-perspective LLMs, achieving an average accuracy gain of **9.98%** on GPT‐4.We further demonstrate that using instance‐aligned personas yields more accurate and stable performance than using dataset-aligned personas.

Large Language Models (LLMs) are predominantly assessed based on their common sense reasoning, language comprehension, and logical reasoning abilities. While models trained in specialized domains like mathematics or coding have demonstrated remarkable advancements in logical reasoning, there remains a significant gap in evaluating their code generation capabilities. Existing benchmark datasets fall short in pinpointing specific strengths and weaknesses, impeding targeted enhancements in models’ reasoning abilities to synthesize code. To bridge this gap, our paper introduces an innovative, pedagogical benchmarking method that mirrors the evaluation processes encountered in academic programming courses. We introduce CodeEval, a multi-dimensional benchmark dataset designed to rigorously evaluate LLMs across 24 distinct aspects of Python programming. The dataset covers three proficiency levels—beginner, intermediate, and advanced—and includes both class-based and function-based problem types with detailed problem specifications and comprehensive test suites. To facilitate widespread adoption, we also developed RunCodeEval, an open-source execution framework that provides researchers with a ready-to-use evaluation pipeline for CodeEval. RunCodeEval handles test execution, context setup, and metrics generation, enabling researchers to quickly obtain detailed insights into model strengths and weaknesses across complexity levels, problem types, and programming categories. This combination enables targeted evaluation and guides improvements in LLMs’ programming proficiencies.

Rapid information access is vital during wildfires, yet traditional data sources are slow and costly. Social media offers real-time updates, but extracting relevant insights remains a challenge. In this work, we focus on multimodal wildfire social media data, which, although existing in current datasets, is currently underrepresented in Canadian contexts. We present WildFireCan-MMD, a new multimodal dataset of X posts from recent Canadian wildfires, annotated across twelve key themes. We evaluate zero-shot vision-language models on this dataset and compare their results with those of custom-trained and baseline classifiers. We show that while baseline methods and zero-shot prompting offer quick deployment, custom-trained models outperform them when labelled data is available. Our best-performing custom model reaches 84.48±0.69% f-score, outperforming VLMs and baseline classifiers. We also demonstrate how this model can be used to uncover trends during wildfires, through the collection and analysis of a large unlabeled dataset. Our dataset facilitates future research in wildfire response, and our findings highlight the importance of tailored datasets and task-specific training. Importantly, such datasets should be localized, as disaster response requirements vary across regions and contexts.

Effective tool pre-selection via retrieval is essential for AI agents to select from a vast array of tools when identifying and planning actions in the context of complex user queries. Despite its central role in planning, this aspect remains underexplored in the literature. Traditional approaches rely primarily on similarities between user queries and tool descriptions, which significantly limits retrieval accuracy, specifically when handling multi-step user requests. To address these limitations, we propose a Knowledge Graph (KG)-based tool retrieval framework that captures the semantic relationships between tools and their functional dependencies. Our retrieval algorithm leverages ensembles of 1-hop ego tool graphs to model direct and indirect connections between tools, enabling more comprehensive and contextual tool selection for multi-step tasks. We evaluate our approach on a synthetically generated internal dataset across six defined user classes, extending previous work on coherent dialogue synthesis and tool retrieval benchmarks. Results demonstrate that our tool graph-based method achieves 91.85% tool coverage on the micro-average CompleteRecall metric, compared to 89.26% for re-ranked semantic-lexical hybrid retrieval, the strongest non-KG baseline in our experiments. These findings support our hypothesis that the structural information modeled in the graph provides complementary signals to pure similarity matching, particularly for queries requiring sequential tool composition.

Language-specific neurons, units in LLMs that strongly correlate with individual languages have been shown to influence model behavior by deactivating them. However, their role in amplification remains underexplored.This work investigates the effect of amplifying language-specific neurons through interventions across 18 languages, including low-resource ones, using three models primarily trained in different languages. We compare amplification factors by their effectiveness in steering to the target language using a proposed Language Steering Shift (LSS) evaluation score, then evaluate it on downstream tasks: commonsense reasoning (XCOPA, XWinograd), knowledge (Include), and translation (FLORES). The optimal amplification steering factors effectively steer output toward nearly all tested languages. Intervention using this factor on downstream tasks improves self-language performance in some cases but generally degrades cross-language results. These findings highlight the effect of language-specific neurons in multilingual behavior, where amplification can be beneficial especially for low-resource languages, but provides limited advantage for cross-lingual transfer.

Coreference resolution is a fundamental task in natural language processing that involves linking different references to the same entity within a text. However, existing models often struggle to reliably identify referential relationships in contexts with extensive length or complex modifiers. This study proposes a data augmentation technique adding adjective phrases and employing a prompt-based adversarial filtering pipeline to address these challenges. Specifically, we generated and inserted contextually appropriate adjective phrases through the interaction between GPT-4o-mini based Few-shot Prompting and a Discriminative Language Model. The grammatical and semantic consistency of these phrases was validated via human evaluation and inter-annotator agreement (IAA) procedures. The generated synthetic dataset was integrated with existing data, leading to enhanced model performance. On the LitBank dataset, the CoNLL-F1 score increased by up to 1.7%, while the synthetic dataset improved linguistic diversity and the complexity of referential structures. The proposed pipeline represents a significant step towards developing coreference resolution models capable of better capturing linguistic variety and demonstrating robustness under challenging conditions.

In the legal domain, Fact-based Judgment Prediction and Explanation (FJPE) aims to predict judicial outcomes and generate grounded explanations using only factual information, mirroring early-phase legal reasoning. Motivated by the overwhelming case backlog in the Indian judiciary, we introduce TathyaNyaya, the first large-scale, expert-annotated dataset for FJPE in the Indian context. Covering judgments from the Supreme Court and multiple High Courts, the dataset comprises four complementary components, NyayaFacts, NyayaScrape, NyayaSimplify, and NyayaFilter, that facilitate diverse factual modeling strategies. Alongside, we present FactLegalLlama, an instruction-tuned LLaMa-3-8B model fine-tuned to generate faithful, fact-grounded explanations. While FactLegalLlama trails transformer baselines in raw prediction accuracy, it excels in generating interpretable explanations, as validated by both automatic metrics and legal expert evaluation. Our findings show that fact-only inputs and preprocessing techniques like text simplification and fact filtering can improve both interpretability and predictive performance. Together, TathyaNyaya and FactLegalLlama establish a robust foundation for realistic, transparent, and trustworthy AI applications in the Indian legal system.

Large language models are increasingly adapted to downstream tasks through fine-tuning. Full supervised fine-tuning (SFT) and parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), are two dominant approaches. While PEFT methods are widely used for their computational efficiency, the implications of their configurations (e.g., rank) remain under-explored in downstream Q&A tasks and generalisation. In this work, we perform a comprehensive evaluation across multiple reasoning and recall datasets, conducting a rank sweep to quantify the trade-off between SFT and PEFT. We also compare the accuracy of PEFT and SFT models across in-domain and out-of-domain adaptation, highlighting distinct generalisation behaviour and task-specific forgetting. We demonstrate that LoRA achieves competitive and in some cases superior performance compared to SFT, particularly on reasoning tasks at specific rank values. Additionally, we analyze the internal representations via spectral features and layer-wise attention structures, offering insights into representational drift and structural changes in attention patterns.

Detoxification, the task of rewriting harmful language into non-toxic text, has become increasingly important amid the growing prevalence of toxic content online. However, high-quality parallel datasets for detoxification, especially for hate speech, remain scarce due to the cost and sensitivity of human annotation. In this paper, we propose a novel LLM-in-the-loop pipeline leveraging GPT-4o-mini for automated detoxification. We first replicate the ParaDetox pipeline by replacing human annotators with LLM and show that LLM performs comparably to the human annotation. Building on this, we construct ParaDeHate, a large-scale parallel dataset specifically for hate speech detoxification. We release ParaDeHate as a benchmark of over 8,000 hate/non-hate text pairs and evaluate a wide range of baseline methods. Experimental results show that models such as BART fine-tuned on ParaDeHate achieve better performance in style accuracy, content preservation, and fluency, demonstrating the effectiveness of LLM-generated detoxification text as a scalable alternative to human annotation.

Large Language Models (LLMs) exhibit strong linguistic capabilities, but little is known about how they encode psycholinguistic knowledge across languages. We investigate whether and how LLMs exhibit human-like psycholinguistic responses under different linguistic identities using two tasks: sound symbolism and word valence. We evaluate two models, Llama-3.3-70B-Instruct and Qwen2.5-72B-Instruct, under monolingual and bilingual prompting in English, Dutch, and Chinese. Behaviorally, both models adjust their outputs based on prompted language identity, with Qwen showing greater sensitivity and sharper distinctions between Dutch and Chinese. Probing analysis reveals that psycholinguistic signals become more decodable in deeper layers, with Chinese prompts yielding stronger and more stable valence representations than Dutch. Our results demonstrate that language identity conditions both output behavior and internal representations in LLMs, providing new insights into their application as models of cross-linguistic cognition.

Rising mental health issues among youth have increased interest in automated approaches for detecting early signs of psychological distress in digital text. One key focus is the identification of cognitive distortions, irrational thought patterns that have a role in aggravating mental distress. Early detection of these distortions may enable timely, low-cost interventions. While prior work has focused on English clinical data, we present the first in-depth study of cross-lingual and cross-register generalization of cognitive distortion detection, analyzing forum posts written by Dutch adolescents. Our findings show that while changes in language and writing style can significantly affect model performance, domain adaptation methods show the most promise

Traditional Retrieval-Augmented Generation (RAG) struggles with complex queries that lack strong signals to retrieve the most relevant context, forcing a trade-off between choosing a small context that misses key information and a large context that confuses the LLM. To address this, we propose Forward-Backward RAG (FB-RAG), a new training-free framework based on a simple yet powerful forward-looking strategy. FB-RAG employs a light-weight LLM to peek into potential future generations, using evidence from multiple sampled outputs to precisely identify the most relevant context for a final, more powerful generator. This improves performance without complex finetuning or Reinforcement Learning common in prior work. Across 9 datasets from LongBench and ∞Bench, FB-RAG consistently delivers strong results. Further, the performance gains can be achieved with reduced latency due to a shorter, more focused prompt for the powerful generator. On EN.QA dataset, FB-RAG matches the leading baseline with over 48% latency reduction or achieves an 8% performance improvement with a 10% latency reduction. Our analysis finds cases where even when the forward-looking LLM fails to generate correct answers, its attempts are sufficient to guide the final model to an accurate response, demonstrating how smaller LLMs can systematically improve the performance and efficiency of larger ones.

Dysarthria, a motor speech disorder affecting over 46 million individuals globally, impairs both intelligibility and emotional expression in communication. This work introduces a novel framework for emotion-aware sentence reconstruction from dysarthric speech using Large Language Models (LLMs) fine-tuned with QLoRA, namely LLaMA 3.1 and Mistral 8x7B. Our pipeline integrates direct emotion recognition from raw audio and conditions textual reconstruction on this emotional context to enhance both semantic and affective fidelity.We propose the Multimodal Communication Dysarthria Score (MCDS), a holistic evaluation metric combining BLEU, semantic similarity, emotion consistency, and human ratings:MCDS=αB+βE+γS+δHwhere 𝛼 + 𝛽 + 𝛾 + 𝛿 = 1.On our extended TORGO+ dataset, our emotion-aware LLM model achieves a MCDS of 0.87 and BLEU of 72.4%, significantly outperforming traditional pipelines like Kaldi GMM-HMM (MCDS: 0.52, BLEU: 38.1%) and Whisper-based models. It also surpasses baseline LLM systems by 0.09 MCDS. This sets a new benchmark in emotionally intelligent dysarthric speech reconstruction, with future directions including multilingual support and real-time deployment.

Difficult and unfamiliar concepts often hinder comprehension for lay audiences, especially in technical and educational domains. This motivates the usage of large language models (LLMs) for the process of text simplification (TS). In this work, we propose an iterative refinement framework that aims to simplify definitions by carefully handling complex terminology and domain-specific expressions. The obtained definition is reprocessed based on the critique, making refinements in successive iterations. We emphasize the use of small language models (SLMs) due to their faster response times and cost-efficient deployment. Human evaluations of the definitions produced at each refinement stage indicate consistent improvements in our specified evaluation criteria. We evaluate both LLM-as-a-judge score and human assessments along with automated metrics like BERTScore, BLEU-4, which provided supporting evidence for the effectiveness of our approach. Our work highlights the use of LLMs mimicking human-like feedback system in a TS task catering to a reader’s specific cognitive needs. Thus, we find that an iterative, critique-driven method can be an effective strategy for the simplification of dense or technical texts, particularly in domains where jargon impedes understanding.

Persuasion is a central feature of communication, widely used to influence beliefs, attitudes, and behaviors. In today’s digital landscape, across social media and online platforms, persuasive content is pervasive, appearing in political campaigns, marketing, fundraising appeals, and more. These strategies span a broad spectrum, from rational and ethical appeals to highly manipulative tactics, some of which pose significant risks to individuals and society. Despite the growing need to identify and differentiate safe from unsafe persuasion, empirical research in this area remains limited. To address this gap, we introduce SafePersuasion, a two-level taxonomy and annotated dataset that categorizes persuasive techniques based on their safety. We evaluate the baseline performance of three large language models in detecting manipulation and its subtypes, and report only moderate success in distinguishing manipulative content from rational persuasion. By releasing SafePersuasion, we aim to advance research on detecting unsafe persuasion and support the development of tools that promote ethical standards and transparency in persuasive communication online.

This work considers a black-box threat model in which adversaries attempt to propagate arbitrary non-relevant content in search. We show that retrievers, rerankers, and LLM relevance judges are all highly vulnerable to attacks that enable arbitrary content to be promoted to the top of search results and to be assigned perfect relevance scores. We investigate how attackers may achieve this via content injection, injecting arbitrary sentences into relevant passages or query terms into arbitrary passages. Our study analyzes how factors such as model class and size, the balance between relevant and non-relevant content, injection location, toxicity and severity of injected content, and the role of LLM-generated content influence attack success, yielding novel, concerning, and often counterintuitive results. Our results reveal a weakness in embedding models, LLM-based scoring models, and generative LLMs, raising concerns about the general robustness, safety, and trustworthiness of language models regardless of the type of model or the role in which they are employed. We also emphasize the challenges of robust defenses against these attacks. Classifiers and more carefully prompted LLM judges often fail to recognize passages with content injection, especially when considering diverse text topics and styles. Our findings highlight the need for further research into arbitrary content injection attacks. We release our code for further study: https://github.com/manveertamber/content_injection_attacks.

We show that human players’ gameplay in the game of Wordle is influenced by the semantics, orthography, and phonology of the player’s previous guesses. We compare actual human players’ guesses with near-optimal guesses using NLP techniques. We study human language use in the constrained environment of Wordle, which is situated between natural language use and the artificial word association task.

Large language models (LLMs) sometimes hallucinate facts. Recent studies have shown that use of non-factual LLMs (anti-expert) have the potential to improve the factuality of the base LLM. Anti-expert methods penalize the output probabilities of the base LLM with an anti-expert LLM. Anti-expert methods are effective in mitigating hallucinations, but require high computational costs because the two LLMs are run simultaneously. In this paper, we propose an efficient anti-expert method called in-model anti-expert. It mitigated the hallucination problem with a single LLM and intervening to change the internal representations in the direction of improving factuality. Experiments results showed that the proposed method is less costly than the conventional anti-expert method and outperformed existing methods except for the anti-expert method. We confirmed that the proposed method improved GPU memory usage from 2.2x to 1.2x and latency from 1.9x to 1.2x.

Preference optimization is a critical post-training technique used to align large language models (LLMs) with human preferences, typically by fine-tuning on ranked response pairs. While methods like Direct Preference Optimization (DPO) have proven effective in English, they often fail to generalize robustly to multilingual settings. We propose a simple yet effective alternative, Confidence-Aware Preference Optimization (CAPO), which replaces DPO’s fixed treatment of preference pairs with a dynamic loss scaling mechanism based on a relative reward. By modulating the learning signal according to the confidence in each preference pair, CAPO enhances robustness to noisy or low-margin comparisons, typically encountered in multilingual text. Empirically, CAPO outperforms existing preference optimization baselines by at least 16% in reward accuracy, and improves alignment by widening the gap between preferred and dispreferred responses across languages.

Test-time compute has emerged as a powerful paradigm in function-level code generation. However, previous proposed strategies have been viewed as disparate, thus lacking a fair apples-to-apples analysis enabling understanding of their operational mechanisms in execution-based benchmarks. Therefore, we present a mathematical framework that unifies generation and reranking with theoretical justifications through the lens of Minimum Bayes Risk (MBR) decoding. Our proposed framework leads to key research questions regarding the effectiveness of using parallel and/or iterative sampling, design choices of reranking signals and soft/hard MBR utility functions, and behaviors of the final selected program across different methods. Our empirical findings highlight the importance of the diversity of sampled candidates (over self-improvement), reranking with simple and high-quality signals, and the effectiveness of test-time compute to select programs that manifest general and edge test case robustness. We will open-source our analysis toolkit and implementation to enable reproducible research.We open-source our analysis toolkit and implementation to enable reproducible research.

The integration of large language models (LLMs) into clinical medicine represents a major advancement in natural language processing (NLP). We introduce BioMistral-Clinical 7B, a clinical LLM built on BioMistral-7B (Labrak et al., 2024), designed to support continual learning from unstructured clinical notes for real-world tasks such as clinical decision support. Using the augmented-clinical-notes dataset provided by Hugging Face (2024), we apply prompt engineering to transform unstructured text into structured JSON, capturing key clinical information (symptoms, diagnoses, treatments, outcomes). This enables efficient incremental training via self-supervised continual learning (SPeCiaL) (Caccia and Pineau, 2021). Evaluation on MedQA (Jin et al., 2021) and MedMCQA (Pal et al., 2022) shows that BioMistral-Clinical 7B improves accuracy on MedMCQA by nearly 10 points (37.4% vs. 28.0%) over the base model, while maintaining comparable performance on MedQA (34.8% vs. 36.5%). Building on this, we propose the BioMistral-Clinical System, which integrates Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) to enrich responses with relevant clinical cases retrieved from a structured vector database. The full system enhances clinical reasoning by combining domain-specific adaptation with contextual retrieval.

This work examines the potential of surprisal slope as a feature for identifying multi-word expressions (MWEs) in English, leveraging token-level surprisal estimates from the GPT-2 language model. Evaluations on the DiMSUM and SemEval-2022 datasets reveal that surprisal slope provides moderate yet meaningful discriminative power with a trade-off between specificity and coverage: while high recall indicates that surprisal slope captures many true MWEs, the slightly lower precision reflects false positives, particularly for non-MWEs that follow formulaic patterns (e.g., adjective-noun or verb-pronoun structures). The method performs particularly well for conventionalized expressions, such as idiomatic bigrams in the SemEval-2022 corpus. Both idiomatic and literal usages of these bigrams exhibit negative slopes, with idiomatic instances generally showing a more pronounced decrease.Overall, surprisal slope offers a cognitively motivated and interpretable signal that complements existing MWE identification methods, particularly for conventionalized expressions.

Large Language Models (LLMs) excel at providing information acquired during pretraining on large-scale corpora and following instructions through user prompts. However, recent studies suggest that LLMs exhibit biases favoring Western native English speakers over non-Western native speakers. Given English’s role as a global lingua franca and the diversity of its dialects, we extend this analysis to examine whether non-native English speakers also receive lower-quality or factually incorrect responses more frequently. We compare three groups—Western native, non-Western native, and non-native English speakers—across classification and generation tasks. Our results show that performance discrepancies occur when LLMs are prompted by the different groups for the classification tasks. Generative tasks, in contrast, are largely robust to nativeness bias, likely due to their longer context length and optimization for open-ended responses. Additionally, we find a strong anchoring effect when the model is made aware of the user’s nativeness for objective classification tasks, regardless of the correctness of this information. Our analysis is based on a newly collected dataset with over 12,000 unique annotations from 124 annotators, including information on their native language and English proficiency.

In this study, we evaluate the performance of large language models (LLMs) in detecting and correcting grammatical errors made by Chinese language learners. We find that incorporating various linguistic features—such as dependency structures, parts of speech, and pinyin transliteration—into the prompts can potentially enhance model performance. Among these features, parts of speech and pinyin prove to be the most effective across all tested models. Additionally, our findings show that the success of error correction also depends on the severity of the errors. When the intended meaning is preserved, LLMs tend to provide accurate revisions following the principle of minimal editing. However, when the meaning is obscured, LLMs are more likely to produce divergent outputs, both in comparison to reference corrections and to the responses of other models.

As large reasoning models (LRMs) grow more capable, chain-of-thought (CoT) reasoning introduces new safety challenges. Existing SFT-based safety alignment studies dominantly focused on filtering prompts with safe, high-quality responses, while overlooking hard prompts that always elicit harmful outputs. To fill this gap, we introduce UnsafeChain, a safety alignment dataset constructed from hard prompts with diverse sources, where unsafe completions are identified and explicitly corrected into safe responses. By exposing models to unsafe behaviors and guiding their correction, UnsafeChain enhances safety while preserving general reasoning ability. We fine-tune three LRMs on UnsafeChain and compare them against recent SafeChain and STAR-1 across six out-of-distribution and five in-distribution benchmarks. UnsafeChain consistently outperforms prior datasets (21 wins out of 36 settings), with even a small selected-1K subset matching or surpassing baseline performance, demonstrating the effectiveness and generalizability of correction-based supervision.We release our dataset and code at https://github.com/mbzuai-nlp/UnsafeChain.

Emotion recognition methods typically assign labels at the sentence level, obscuring the specific linguistic cues that signal emotion. This limits their utility in applications requiring targeted responses, such as empathetic dialogue and clinical support, which depend on knowing which language expresses emotion. The task of identifying emotion evidence – text spans conveying emotion – remains underexplored due to a lack of labeled data. Without span-level annotations, we cannot evaluate whether models truly localize emotion expression, nor can we diagnose the sources of emotion misclassification. We introduce the SEER (Span-based Emotion Evidence Retrieval) Benchmark to evaluate Large Language Models (LLMs) on this task. SEER evaluates single and multi-sentence span identification with new annotations on 1200 real-world sentences. We evaluate 14 LLMs and find that, on single-sentence inputs, the strongest models match the performance of average human annotators, but performance declines in multi-sentence contexts. Key failure modes include fixation on emotion keywords and false positives in neutral text.

Despite Telugu being spoken by over 80 million people, speech translation research for this morphologically rich language remains severely underexplored. We address this gap by developing a high-quality Telugu-English speech translation benchmark from 46 hours of manually verified CSTD corpus data (30h/8h/8h train/dev/test split). Our systematic comparison of cascaded versus end-to-end architectures shows that while IndicWhisper + IndicMT achieves the highest performance due to extensive Telugu-specific training data, fine-tuned SeamlessM4T models demonstrate remarkable competitiveness despite using significantly less Telugu-specific training data. This finding suggests that with careful hyperparameter tuning and sufficient parallel data (potentially less than 100 hours), end-to-end systems can achieve performance comparable to cascaded approaches in low-resource settings. While our metric reliability study evaluating BLEU, METEOR, ChrF++, ROUGE-L, TER, and BERTScore against human judgments reveals that traditional metrics provide better quality discrimination than BERTScore for Telugu–English translation. The work delivers three key contributions: a reproducible Telugu–English benchmark, empirical evidence of competitive end-to-end performance potential in low-resource scenarios, and practical guidance for automatic evaluation in morphologically complex language pairs.

While majority of current resources rely on formal text corrections, our work shifts the focus to non-native spoken Hindi error correction, which presents unique challenges due to its rich morphology, complex syntax, and distinct error patterns. To address the scarcity of authentic learner data, we introduce HiLearners, a dataset gathered from 2,500 real non-native Hindi speakers across three linguistic backgrounds (English, Bengali, Dravidian), capturing authentic error patterns including transfer errors, overgeneralization patterns, and contextual agreement issues. Furthermore, to overcome data resource limitations, we develop a methodical synthetic data augmentation technique, utilizing Large Language Models (LLMs) with a pattern analysis and controlled generation approach similar to Retrieval-Augmented Generation (RAG), yielding 5,500 carefully verified synthetic examples. Through extensive experiments on individual, mixed, and progressive curriculum-based configurations using multilingual models, we demonstrate that LLM-based synthetic data combined with three-phase curriculum learning significantly boosts performance, achieving a 76.92 GLEU score and surpassing human-only baselines. This work bridges the gap between native-centric error correction research and non-native Hindi learner needs, establishing a realistic assessment standard for advancing low-resource language processing.

The use of large language models (LLMs) for evaluating outputs is becoming an increasingly effective and scalable approach. However, it remains uncertain whether this capability extends beyond task-specific evaluations to more general assessments of text quality, particularly in multilingual contexts. In this study, we introduce – MTQ-Eval – a novel framework for multilingual text quality evaluation. We automatically generate text quality preference data and train open-source base LLMs to align with ratings of high- and low-quality text. Our comprehensive evaluation across 115 languages demonstrates the improved performance of the proposed model. Additionally, we explore whether this enhanced ability to distinguish between high- and low-quality text translates to better performance in downstream tasks.

The 3rd Generation Partnership Project (3GPP) produces complex technical specifications essential to global telecommunications, yet their hierarchical structure, dense formatting, and multi-modal content make them difficult to process. While Large Language Models (LLMs) show promise, existing approaches fall short in handling complex queries, visual information, and document interdependencies. We present TelcoAI, an agentic, multi-modal Retrieval-Augmented Generation (RAG) system tailored for 3GPP documentation. TelcoAI introduces section-aware chunking, structured query planning, metadata-guided retrieval, and multi-modal fusion of text and diagrams. Evaluated on multiple benchmarks—including expert-curated queries—our system achieves 87% recall, 83% claim recall, and 92% faithfulness, representing a 16% improvement over state-of-the-art baselines. These results demonstrate the effectiveness of agentic and multi-modal reasoning in technical document understanding, advancing practical solutions for real-world telecommunications research and engineering.

Discourse relation parsing plays a crucial role in uncovering the logical structure of text, yet existing corpora focus almost exclusively on general-domain genres, leaving specialized fields like engineering under-resourced. We introduce ENG‐DRB, the first PDTB‐style discourse relation corpus derived from transcripts of hands‐on engineering tutorial videos. ENG‐DRB comprises 11 tutorials spanning civil, mechanical, and electrical/electronics engineering (155 minutes total) with 1,215 annotated relations. Compared to general‐domain benchmarks, this dataset features a high proportion of explicit senses, dense causal and temporal relations, and frequent overlapping and embedded senses. Our benchmarking experiments underscore the dataset’s difficulty. A top parser (HITS) detects segment boundaries well (98.6% F1), but its relation classification is more than 11 F1 percentages lower than on the standard PDTB. In addition, state‐of‐the‐art LLMs (OpenAI o4‐mini, Claude 3.7, LLaMA‐3.1) achieve at best 41% F1 on explicit relations and less than 9% F1 on implicit relations, revealing systematic errors in temporal and causal sense detection. The dataset can be accessed at: https://doi.org/10.57967/hf/6895. Code to reproduce our results is available at: https://github.com/chengzhangedu/ENG-DRB.

We investigate the characteristics of location review texts written on the basis of actual visit experiences or without any visit experiences. Specifically, we formalize this as a binary classification task and propose a data construction framework that labels reviews as Visit or NotVisit by linking them with users’ GPS-based movement data. We train a logistic regression model on the dataset and evaluate it alongside human annotators and a large language model (LLM). The results show that the task is more challenging for humans and LLMs than for the simple trained model.

Multimodal sarcasm detection is challenging, especially in low-resource settings where subtle image-text contradictions are hard to learn due to scarce annotated data, which hinders the model’s performance. Parameter-efficient fine-tuning (PEFT) methods like adapters, LoRA, and prompt tuning reduce overfitting but struggle to reach optimal performance due to limited supervision from few-shot data. We propose PEKD, a unified framework that enhances PEFT methods via distillation from an expert model trained on large-scale sarcasm data, which acts as the teacher. To mitigate unreliable signals from the teacher, we introduce an entropy-aware gating mechanism that dynamically adjusts the distillation strength based on teacher confidence. Experiments on two public datasets demonstrate that our PEKD framework enables PEFT methods to outperform both prior parameter-efficient approaches and large multimodal models, achieving strong results in the few-shot scenario. The framework is modular and adaptable to a wide range of multimodal models and tasks.

The rise of generative AI has led to challenges in distinguishing AI-generated text from human-written content, raising concerns about misinformation and content authenticity. Detecting AI-generated text remains challenging, especially under various stylistic domains and paraphrased inputs. We introduce SGG-ATD, a novel detection framework that models structural and contextual relationships between LLM-predicted and original-input text. By masking parts of the input and reconstructing them using a language model, we capture implicit coherence patterns. These are encoded in a graph where cosine and contextual links between keywords guide classification via a Graph Convolutional Network (GCN). SGG-ATD achieves strong performance across diverse datasets and shows resilience to adversarial rephrasing and out-of-distribution inputs, outperforming competitive baselines.

Retrieval Augmented Generation (RAG) framework mitigates hallucinations in Large Language Models (LLMs) by integrating external knowledge, yet faces two critical challenges: (1) the distribution gap between user queries and knowledge bases in a specific domain, and (2) incomplete coverage of required knowledge for complex queries. Existing solutions either require task-specific annotations or neglect inherent connections among query, context, and missing knowledge interactions. We propose a reasoning-based missing knowledge RAG framework that synergistically resolves both issues through Chain-of-Thought reasoning. By leveraging open-source LLMs, our method generates structured missing knowledge queries in a single inference pass while aligning query knowledge distributions, and integrates reasoning traces into answer generation. Experiments on open-domain medical and general question answering (QA) datasets demonstrate significant improvements in context recall and answer accuracy. Our approach achieves effective knowledge supplementation without additional training, offering enhanced interpretability and robustness for real-world QA applications.

While recent work has begun to uncover the internal strategies that Large Language Models (LLMs) employ for simple arithmetic tasks, a unified understanding of their underlying mechanisms is still lacking. We extend recent findings showing that LLMs represent numbers in a digit-wise manner and present evidence for the existence of digit-position-specific circuits that LLMs use to perform simple arithmetic tasks, i.e. modular subgroups of MLP neurons that operate independently on different digit positions (units, tens, hundreds). Notably, such circuits exist independently of model size and of tokenization strategy, i.e. both for models that encode longer numbers digit-by-digit and as one token.Using Feature Importance and Causal Interventions, we identify and validate the digit-position-specific circuits, revealing a compositional and interpretable structure underlying the solving of arithmetic problems in LLMs. Our interventions selectively alter the model’s prediction at targeted digit positions, demonstrating the causal role of digit-position circuits in solving arithmetic tasks.

Despite their success, LLMs are too computationally expensive to replace task- or domain-specific NLP systems. However, the variety of corpus formats makes reusing these systems difficult. This underscores the importance of maintaining an interoperable NLP landscape. We address this challenge by pursuing two objectives: standardizing corpus formats and enabling massively parallel corpus processing. We present a unified conversion framework embedded in a massively parallel, microservice-based, programming language-independent NLP architecture designed for modularity and extensibility. It allows for the integration of external NLP conversion tools and supports the addition of new components that meet basic compatibility requirements. To evaluate our dual data- and process-oriented approach to standardization, we (1) benchmark its efficiency in terms of processing speed and memory usage, (2) demonstrate the benefits of standardized corpus formats for NLP downstream tasks, and (3) illustrate the advantages of incorporating custom formats into a corpus format ecosystem.

LLMs have emerged as a transformative technology, enabling a wide range of tasks such as text generation, summarization, question answering, and more. The use of RAG with LLM is on the rise to provide deeper knowledge bases of various domains. In the present study, we propose a RAG framework that employs weighted Rocchio mechanism for retrieval and LLM collaborative forum with supervision for generation. Our framework is evaluated in two downstream tasks: a biomedical question answering (BioASQ-QA) and a multilingual claim verification (e.g. in English, Hindi, and Bengali) to showcase its adaptability across various domains and languages. The proposed retriever is capable to achieve substantial improvement over BM25 of +8% (BioASQ-QA), +15% (English), +5% (Hindi), and +20% (Bengali) for Recall@5. In veracity classification, our framework achieves an average answer correctness of 0.78 on BioASQ-QA while achieving F1-score of 0.59, 0.56, and 0.41 for English, Hindi and Bengali languages, respectively. These results demonstrate the effectiveness and robustness of our framework for retrieval and generation in multilingual and multi-domain settings.

Despite the communicative importance of negation, its detection remains challenging. Previous approaches perform poorly in out-of-domain scenarios, and progress outside of English has been slow due to a lack of resources and robust models. To address this gap, we present D-Neg: a syntax-aware graph reasoning model based on a transformer that incorporates syntactic embeddings by attention-gating. D-Neg uses graph attention to represent syntactic structures, emulating the effectiveness of rule-based dependency approaches for negation detection. We train D-Neg using 7 English resources and their translations into 10 languages, all aligned at the annotation level. We conduct an evaluation of all these datasets in in-domain and out-of-domain settings. Our work represents a significant advance in negation detection, enabling more effective cross-lingual research.

Shaping inclusive representations that embrace diversity and ensure fair participation and reflections of values is at the core of many conversation-based models. However, many existing methods rely on surface inclusion using mention of user demographics or behavioral attributes of social groups. Such methods overlook the nuanced, implicit expression of opinion embedded in conversations. Furthermore, the over-reliance on overt cues can exacerbate misalignment and reinforce harmful or stereotypical representations in model outputs. Thus, we took a step back and recognized that equitable inclusion needs to account for the implicit expression of opinion and use the stance of responses to validate the normative alignment. This study aims to evaluate how opinions are represented in NLP or computational models by introducing an alignment evaluation framework that foregrounds implicit, often overlooked conversations and evaluates the normative social views and discourse. Our approach models the stance of responses as a proxy for the underlying opinion, enabling a considerate and reflective representation of diverse social viewpoints. We evaluate the framework using both (i) positive-unlabeled (PU) online learning with base classifiers, and (ii) instruction-tuned language models to assess post-training alignment. Through this, we provide a basis for understanding how implicit opinions are (mis)represented and offer a pathway toward more inclusive model behavior.

Large language models (LLMs) have demonstrated strong capabilities in translating natural language questions about relational databases into SQL queries. In particular, test-time scaling techniques such as Self-Consistency and Self-Correction can enhance SQL generation accuracy by increasing computational effort during inference. However, these methods have notable limitations: Self-Consistency may select suboptimal outputs despite majority votes, while Self-Correction typically addresses only syntactic errors. To leverage the strengths of both approaches, we propose CSC-SQL, a novel method that integrates Self-Consistency and Self-Correction. CSC-SQL selects the two most frequently occurring outputs from parallel sampling and feeds them into a merge revision model for correction. Additionally, we employ the Group Relative Policy Optimization (GRPO) algorithm to fine-tune both the SQL generation and revision models via reinforcement learning, significantly enhancing output quality. Experimental results confirm the effectiveness and generalizability of CSC-SQL. On the BIRD private test set, our 7B model achieves 71.72% execution accuracy, while the 32B model achieves 73.67%, outperforming other known methods using open source models. The code has been open sourced at https://github.com/CycloneBoy/csc_sql.

Large language models (LLMs) have demonstrated strong performance in translating natural language questions into SQL queries (Text-to-SQL). In contrast, small language models (SLMs) ranging from 0.5B to 1.5B parameters currently underperform on Text-to-SQL tasks due to their limited logical reasoning capabilities. However, SLMs offer inherent advantages in inference speed and suitability for edge deployment. To explore their potential in Text-to-SQL applications, we leverage recent advancements in post-training techniques. Specifically, we used the open-source SynSQL-2.5M dataset to construct two derived datasets: SynSQL-Think-916K for SQL generation and SynSQL-Merge-Think-310K for SQL merge revision. We then applied supervised fine-tuning and reinforcement learning-based post-training to the SLM, followed by inference using a corrective self-consistency approach. Experimental results validate the effectiveness and generalizability of our method, SLM-SQL. On the BIRD development set, the five evaluated models achieved an average improvement of 31.4 points. Notably, the 0.5B model reached 56.87% execution accuracy (EX), while the 1.5B model achieved 67.08% EX. On the BIRD private test set, our 0.5B model achieves 61.82% EX, while the 1.5B model achieves 70.49%. We will release our dataset, model, and code to github: https://github.com/CycloneBoy/slm_sql.

Evaluating free-form Question-Answering (QA) remains a challenge due to its diverse and open-ended nature. Traditional automatic metrics fail to capture semantic equivalence or accommodate the variability of open-ended responses. Leveraging Large Language Models (LLMs) as evaluators offers a promising alternative due to their strong language understanding and instruction-following capabilities. We propose the Consensus via Lightweight Efficient Voting (CLEV), which employs two primary LLMs as judges and engages a third judge only in cases of disagreement. This approach prioritizes evaluation reliability while reducing unnecessary computational demands. Through experiments, including human evaluation, we demonstrate CLEV’s ability to provide consistent, scalable, and resource-efficient assessments, establishing it as a robust framework for evaluating LLMs on free-form QA.

In this work, we aim to clarify and reconcile metrics for evaluating privacy protection in text through a systematic survey.Although text anonymization is essential for enabling NLP research and model development in domains with sensitive data, evaluating whether anonymization methods sufficiently protect privacy remains an open challenge. In manually reviewing 47 papers that report privacy metrics, we identify and compare six distinct privacy notions, and analyze how the associated metrics capture different aspects of privacy risk. We then assess how well these notions align with legal privacy standards (HIPAA and GDPR), as well as user-centered expectations grounded in HCI studies. Our analysis offers practical guidance on navigating the landscape of privacy evaluation approaches further and highlights gaps in current practices. Ultimately, we aim to facilitate more robust, comparable, and legally aware privacy evaluations in text anonymization.

Socio-demographic prompting (SDP) - prompting Large Language Models (LLMs) using demographic proxies to generate culturally aligned outputs - often shows LLM responses as stereotypical and biased. While effective in assessing LLMs’ cultural competency, SDP is prone to confounding factors such as prompt sensitivity, decoding parameters, and the inherent difficulty of generation over discrimination tasks due to larger output spaces. These factors complicate interpretation, making it difficult to determine if the poor performance is due to bias or the task design. To address this, we use inverse socio-demographic prompting (ISDP), where we prompt LLMs to discriminate and predict the demographic proxy from actual and simulated user behavior from different users. We use the Goodreads-CSI dataset (Saha et al., 2025), which captures difficulty in understanding English book reviews for users from India, Mexico, and the USA, and test four LLMs: Aya-23, Gemma-2, GPT-4o, and LLaMA-3.1 with ISDP. Results show that models perform better with actual behaviors than simulated ones, contrary to what SDP suggests. However, performance with both behavior types diminishes and becomes nearly equal at the individual level, indicating limits to personalization.

Fine-tuning Large Language Models(LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from O\bigl(N³\bigl) to O\bigl(N²\bigl) and maintaining the same memory complexity for a transformer based model. Our approach achieves significant training speedup while preserving accuracy. Our implementation is available online(https://github.com/devrev/One-Pass-to-Reason).

Recent research has delved into Retrieval-based In-Context Learning (RetICL), leveraging the power of large language models (LLMs) for text classification. Despite its promise, a persistent challenge lies in effectively retrieving relevant demonstrations from a support set. Many existing approaches have overlooked the essential role of linguistic label information in guiding this retrieval process. To bridge this gap, we present Contrastive Linguistic Label Retrieval-based In-Context Learning (CLL-RetICL), a novel framework designed to identify the most relevant and impactful sentences without altering the model parameters. Our approach uniquely integrates sentence-query similarity with sentence-label similarity, enabling a more nuanced and comprehensive evaluation of relevance. We tested CLL-RetICL across diverse text classification tasks and evaluated its performance on various LLMs. Experimental results demonstrate that CLL-RetICL consistently outperforms previous retrieval methods that do not incorporate linguistic label information. These findings highlight the critical importance of linguistic label-aware selection in enhancing text classification accuracy.

Historical documents in the Sinosphere are known to share common formats and practices, particularly in veritable records compiled by court historians. This shared linguistic heritage has led researchers to use Classical Chinese resources for cross-lingual transfer when processing historical documents from Korea and Japan, which remain relatively low-resource. In this paper, we question the assumption of cross-lingual transferability from Classical Chinese to Hanja and Kanbun, the ancient written languages of Korea and Japan, respectively. Our experiments across machine translation, named entity recognition, and punctuation restoration tasks show minimal impact of Classical Chinese datasets on language model performance for ancient Korean documents written in Hanja, with performance differences within ±0.0068 F1-score for sequence labeling tasks and up to +0.84 BLEU score for translation. These limitations persist consistently across various model sizes, architectures, and domain-specific datasets. Our analysis reveals that the benefits of Classical Chinese resources diminish rapidly as local language data increases for Hanja, while showing substantial improvements only in extremely low-resource scenarios for both Korean and Japanese historical documents. These findings emphasize the need for careful empirical validation rather than assuming benefits from indiscriminate cross-lingual transfer.

Automatic speech recognition for low-resource languages remains fundamentally constrained by the scarcity of labeled data and computational resources required by state-of-the-art models. We present a systematic investigation into cross-lingual continuous pretraining for low-resource languages, using Perso-Arabic languages (Persian, Arabic, and Urdu) as our primary case study. Our approach demonstrates that strategic utilization of unlabeled speech data can effectively bridge the resource gap without sacrificing recognition accuracy. We construct a 3,000-hour multilingual corpus through a scalable unlabeled data collection pipeline and employ targeted continual pretraining combined with morphologically-aware tokenization to develop a 300M parameter model that achieves performance comparable to systems 5 times larger. Our model outperforms Whisper Large v3 (1.5B parameters) on Persian and achieves competitive results on Arabic and Urdu despite using significantly fewer parameters and substantially less labeled data. These findings challenge the prevailing assumption that ASR quality scales primarily with model size, revealing instead that data relevance and strategic pretraining are more critical factors for low-resource scenarios. This work provides a practical pathway toward inclusive speech technology, enabling effective ASR for underrepresented languages without dependence on massive computational infrastructure or proprietary datasets.

Gaps arise between a language model’s use of concepts and people’s expectations. This gap is critical when LLMs generate text to help people communicate via Augmentative and Alternative Communication (AAC) tools. In this work, we introduce the evaluation task of Representation Alignment for measuring this gap via human judgment. In our study, we expand keywords and emotion representations into full sentences. We select four emotion representations: Words, Valence-Arousal-Dominance (VAD) dimensions expressed in both Lexical and Numeric forms, and Emojis. In addition to Representation Alignment, we also measure people’s judgments of the accuracy and realism of the generated sentences. While representations like VAD break emotions into easy-to-compute components, our findings show that people agree more with how LLMs generate when conditioned on English words (e.g., “angry”) rather than VAD scales. This difference is especially visible when comparing Numeric VAD to words. Furthermore, we found that the perception of how much a generated sentence conveys an emotion is dependent on both the representation type and which emotion it is.

The proliferation of educational videos on the Internet has changed the educational landscape by enabling students to learn complex concepts at their own pace. Our work outlines the vision of an automated tutor – a multimodal question answering (QA) system to answer questions from students watching a video. This can make doubt resolution faster and further improve learning experience. In this work, we take first steps towards building such a QA system. We curate and release a dataset named EduVidQA, with 3,158 videos and 18,474 QA-pairs. However, building and evaluating an educational QA system is challenging because (1) existing evaluation metrics do not correlate with human judgments, and (2) a student question could be answered in many different ways, training on a single gold answer could confuse the model and make it worse. We conclude with important research questions to develop this research area further.

Automatic ICD coding, the task of assigning disease and procedure codes to electronic medical records, is crucial for clinical documentation and billing. While existing methods primarily enhance model understanding of code hierarchies and synonyms, they often overlook the pervasive use of medical acronyms in clinical notes, a key factor in ICD code inference. To address this gap, we propose a novel effective data augmentation technique that leverages large language models to expand medical acronyms, allowing models to be trained on their full form representations. Moreover, we incorporate consistency training to regularize predictions by enforcing agreement between the original and augmented documents. Extensive experiments on the MIMIC-III dataset demonstrate that our approach, ACE-ICD establishes new state-of-the-art performance across multiple settings, including common codes, rare codes, and full-code assignments. Our code is publicly available.

Logical Table-to-Text (LT2T) generation requires models to both verbalize tabular data and reason over it - performing comparisons, aggregations, and causal inference. While many generation tasks struggle with similar analytical demands, LT2T provides a structured perspective on reasoning capabilities in natural language generation. This survey uses LT2T as a lens to focus on reasoning in data-to-text tasks. By focusing narrowly on LT2T, we present a deep taxonomy of methods that inject, structure, or verify reasoning steps, allowing a level of technical granularity missing in broader surveys. We review representative models and evaluation metrics, and highlight how LT2T techniques transfer to general generation challenges involving logic, numeracy, and faithfulness. Our goal is to distill lessons from LT2T that apply more widely, while also guiding future research in table-based reasoning.

As Large Language Models (LLMs) become integral to human-centered applications, understanding their personality-like behaviors is increasingly important for responsible development and deployment. This paper systematically evaluates six LLMs, applying the Big Five Inventory-2 (BFI-2) framework, to assess trait expressions under varying sampling temperatures. We find significant differences across four of the five personality dimensions, with Neuroticism and Extraversion susceptible to temperature adjustments. Further, hierarchical clustering reveals distinct model clusters, suggesting that architectural features may predispose certain models toward stable trait profiles. Taken together, these results offer new insights into the emergence of personality-like patterns in LLMs and provide a new perspective on model tuning, selection, and the ethical governance of AI systems. We share the data and code for this analysis here: https://osf.io/bsvzc/?view_only=6672219bede24b4e875097426dc3fac1

Knowledge distillation (KD) is a popular method of transferring knowledge from a large “teacher” model to a small “student” model. Previous work has explored various layer-selection strategies (e.g., forward matching and in-order random matching) for intermediate-layer matching in KD, where a student layer is forced to resemble a certain teacher layer. In this work, we revisit such layer-selection strategies and observe an intriguing phenomenon that layer-selection strategy does not matter (much) in intermediate-layer matching—even seemingly nonsensical matching strategies such as *reverse matching* still result in surprisingly good student performance. We provide an interpretation for this phenomenon by examining the angles between teacher layers viewed from the student’s perspective. Our work sheds light on KD practice, as layer-selection strategies may not be the main focus of KD system design and vanilla forward matching works well in most setups.

Recent instruction-tuned large language models (LLMs) have demonstrated remarkable performance on various downstream tasks, including named entity recognition (NER). However, previous approaches often generate incorrect predictions, particularly regarding entity boundaries and types. Many of these errors can be corrected to match the ground truth by revising the entity boundaries and/or types. In this paper, we propose a self-correcting instruction tuning approach that simultaneously learns to perform NER and correct errors through natural language instructions. Self-correcting instruction tuning requires only a standard annotated NER dataset. Supervision for self-correction can be automatically generated from error patterns observed in LLMs fine-tuned solely on NER tasks. We conducted extensive experiments on eight NER datasets with two LLMs to validate the effectiveness of the proposed approach. The results demonstrate that the proposed approach enhances NER performance by effectively correcting prediction errors and substantially reducing false positives. We further analyze the self-correction behavior to better understand how the models improve performance.

Large Language Models (LLMs) have shown remarkable reasoning capabilities in mathematical and scientific tasks. To enhance complex reasoning, multi-agent systems have been proposed to harness the collective intelligence of LLM agents. However, existing collaboration structures are either predefined or rely on majority voting or round-table debates, which can suppress correct but less dominant agent contributions. Recent approaches model multi-agent systems as graph networks but optimize purely for agent performance, neglecting the quality of interactions. We hypothesize that effective agent communication is crucial for multi-agent reasoning and that debating quality plays a significant role. To address this, we propose OptAgent, a multi-agent verbal reinforcement learning algorithm that dynamically constructs and refines multi-agent collaboration structures. Our method defines action spaces and a feedback mechanism that evaluates communication robustness and coherence throughout the debate. The final decision is achieved through a majority vote over all the agents. We assess OptAgent on various reasoning tasks, including mathematical reasoning, creative writing, scientific reasoning, and numerical sorting. Results demonstrate that our approach significantly outperforms single-agent prompting methods and state-of-the-art multi-agent frameworks on diverse tasks.

Low-Rank Adaptation (LoRA) has emerged as a widely adopted parameter-efficient fine-tuning (PEFT) approach for language models. By restricting weight updates to a low-rank subspace, LoRA achieves cost-effective finetuning of large, generalist models to more specialized target domains. While LoRA achieves impressive results for a variety of individual downstream tasks, it struggles to capture the diverse expertise needed when presented with a more heterogeneous finetuning corpus. To address this, we propose Expert Weighted Low-Rank Adaptation (EWoRA), a novel LoRA variant that partitions a rank-(r) adapter into (n) independent adapters of rank (r/n). A lightweight “routing” matrix (W_r R^r n) aggregates the outputs of these adapters by learning specialized weights for each context. Experiments show EWoRA improves performance over LoRA when finetuning on heterogeneous data while generally matching or exceeding LoRA performance on individual finetuning tasks under the same low-rank parameter budget.

In-context learning (ICL) has become a prominent paradigm to rapidly customize LLMs to new tasks without fine-tuning. However, despite the empirical evidence of its usefulness, we still do not truly understand how ICL works. In this paper, we compare the behavior of in-context learning with supervised classifiers trained on ICL demonstrations to investigate three research questions: (1) Do LLMs with ICL behave similarly to classifiers trained on the same examples? (2) If so, which classifiers are closer, those based on gradient descent (GD) or those based on k-nearest neighbors (kNN)? (3) When they do not behave similarly, what conditions are associated with differences in behavior? Using text classification as a use case, with six datasets and three LLMs, we observe that LLMs behave similarly to these classifiers when the relevance of demonstrations is high. On average, ICL is closer to kNN than logistic regression, giving empirical evidence that the attention mechanism behaves more similarly to kNN than GD. However, when demonstration relevance is low, LLMs perform better than these classifiers, likely because LLMs can back off to their parametric memory, a luxury these classifiers do not have.

Punctuation plays a vital role in structuring meaning, yet current models often struggle to restore it accurately in transcripts of spontaneous speech, especially in the presence of disfluencies such as false starts and backtracking. These limitations hinder the performance of downstream tasks like translation, text-to-speech, summarization, etc. where sentence boundaries are critical for preserving quality. In this work, we introduce Cadence, a generalist punctuation restoration model adapted from a pretrained large language model. Cadence is designed to handle both clean written text and highly spontaneous spoken transcripts. It surpasses the previous state-of-the-art in performance while expanding support from 14 to all 22 Indian languages and English. We conduct a comprehensive analysis of model behavior across punctuation types and language families, identifying persistent challenges under domain shift and with rare punctuation marks. Our findings demonstrate the efficacy of utilizing pretrained language models for multilingual punctuation restoration and highlight Cadence’s practical value for low-resource NLP pipelines at scale.

Large language models (LLMs) power modern AI applications, but processing sensitive data on untrusted servers raises privacy concerns. Homomorphic encryption (HE) enables computation on encrypted data for secure inference. However, neural text generation requires decoding methods like argmax and sampling, which are non-polynomial and thus computationally expensive under encryption, creating a significant performance bottleneck. We introduce cutmax, an HE-friendly argmax algorithm that reduces ciphertext operations compared to prior methods, enabling practical greedy decoding under encryption. We also propose the first HE-compatible nucleus (top-p) sampling method, leveraging cutmax for efficient stochastic decoding with provable privacy guarantees. Both techniques are polynomial, supporting efficient inference in privacy-preserving settings. Moreover, their differentiability facilitates gradient-based sequence-level optimization as a polynomial alternative to straight-through estimators. We further provide strong theoretical guarantees for cutmax, proving its convergence via exponential amplification of the gap ratio between the maximum and runner-up elements. Evaluations on realistic LLM outputs show latency reductions of 24×–35× over baselines, advancing secure text generation.

Esports is a competitive sport in which highly skilled players face off in fast-paced video games. Matches consist of intense, moment-by-moment plays that require exceptional technique and strategy. These moments often involve complex interactions, including team fights, positioning, or strategic decisions, which are difficult to interpret without expert explanation. In this study, we set up the task of generating commentary for a specific game moment from multimodal game data consisting of a gameplay screenshot and structured JSON data. Specifically, we construct the first large-scale tri-modal dataset for League of Legends, one of the most popular multiplayer strategy esports titles, and then design evaluation criteria for the task. Using this dataset, we evaluate various large vision language models in generating commentary for a specific moment. We will release the scripts to reconstruct our dataset.

Federated large language models (FedLLMs) enable cross-silo collaborative training among institutions while preserving data locality, making them appealing for privacy-sensitive domains such as law, finance, and healthcare. However, the memorization behavior of LLMs can lead to privacy risks that may cause cross-client data leakage. In this work, we study the threat of *cross-client data extraction*, where a semi-honest participant attempts to recover personally identifiable information (PII) memorized from other clients’ data. We propose three simple yet effective extraction strategies that leverage contextual prefixes from the attacker’s local data, including frequency-based prefix sampling and local fine-tuning to amplify memorization. To evaluate these attacks, we construct a Chinese legal-domain dataset with fine-grained PII annotations consistent with CPIS, GDPR, and CCPA standards, and assess extraction performance using two metrics: *coverage* and *efficiency*. Experimental results show that our methods can recover up to 56.6% of victim-exclusive PII, where names, addresses, and birthdays are particularly vulnerable. These findings highlight concrete privacy risks in FedLLMs and establish a benchmark and evaluation framework for future research on privacy-preserving federated learning. Code and data are available at https://github.com/SMILELab-FL/FedPII.

Generating accurate and concise textual summaries from multimodal documents is challenging, especially when dealing with visually complex content like scientific posters. We introduce PosterSum, a novel benchmark to advance the development of vision-language models that can understand and summarize scientific posters into research paper abstracts. Our dataset contains 16,305 conference posters paired with their corresponding abstracts as summaries. Each poster is provided in image format and presents diverse visual understanding challenges, such as complex layouts, dense text regions, tables, and figures. We benchmark state-of-the-art Multimodal Large Language Models (MLLMs) on PosterSum and demonstrate that they struggle to accurately interpret and summarize scientific posters. We propose Segment & Summarize, a hierarchical method that outperforms current MLLMs on automated metrics, achieving a 3.14% gain in ROUGE-L. This will serve as a starting point for future research on poster summarization.

Self-attention mechanisms have become foundational across modern deep learning architectures. Recent efforts focus on improving their efficiency, particularly for signal processing tasks. The existing approaches employ complex-valued representations for inputs and weights and achieve higher accuracy at the cost of increased model size and inference latency. Dual-numbered algebra offers a promising alternative that allows efficient multiplication and faster inference with the same representational capacity. Inspired by previous studies in the field of hypercomplex neural networks, we introduce a generalized hypercomplex attention block and integrate it into Transformer-based models for EEG classification. Our experiments include adaptation of the hypercomplex models, so that the number of parameters is equal to that of their real-valued counterparts. Across all scenarios, the dual- and complex-numbered models consistently outperform the real ones, demonstrating superior accuracy. This work presents hypercomplex attention as a competitive and computationally efficient strategy with potential value to solve multiple NLP tasks.

Translating natural language requirements into Signal Temporal Logic (STL) is essential for safety-critical systems but requires mathematical expertise. We propose a translational grammar mapping Universal Dependencies (UD) structures to STL Operators through 17 theoretically-motivated patterns, evaluated on the NL2TL benchmarking dataset of 7,002 expert-annotated sentence-STL pairs, and an additional cross-domain analysis. We built a parser guided by this grammar to explore the formal deterministic relationship between UDR Compositions and STL Operators, achieving ~99% sentence coverage, ~54% exact matches (and ~97% similarity). Sentence-level regression analyses predict STL statements and STL Operator classes, considering the co-occurance of UDR substructures (UDR components) with an accuracy of more than ~74% and ~81%, respectively. They uncover a new logical grammatical link between temporal NL and formal logic, that is conditioned by the sentence-level context, and provide insights into how linguistic theory unfolds in practice through temporal linguistic expressions.

The vast majority of the world’s languages, particularly low-resource and indigenous ones like Bhili, remain critically underserved by modern language technologies. The primary bottleneck is the lack of large-scale corpora required for standard pre-training. To address this gap, we introduce cross-lingual contrastive distillation, a novel and data-efficient, compute-efficient paradigm for creating powerful language models without a massive monolingual corpus. Our method adapts a pre-existing multilingual model (MuRIL) by using a fixed, expert teacher model (HindBERT) to distill semantic knowledge from a related high-resource language (Hindi) via a contrastive objective on a modest parallel corpus. Through comprehensive experiments, we show that our resulting model, GARuD-Bhili, significantly outperforms strong zero-shot and MLM-only baselines on a suite of evaluations, including intrinsic language modeling, downstream sentiment analysis, and cross-lingual benchmarks (Tatoeba, XNLI). Our work presents a generalizable and scalable blueprint for linguistic empowerment, offering a practical pathway to develop robust language technologies for other underserved languages globally.

Transformer-based models have revolutionized the field of natural language processing. To understand why they perform so well and to assess their reliability, several studies have focused on questions such as: Which linguistic properties are encoded by these models, and to what extent? How robust are these models in encoding linguistic properties when faced with perturbations in the input text? However, these studies have mainly focused on BERT and the English language. In this paper, we investigate similar questions regarding encoding capability and robustness for 8 linguistic properties across 13 different perturbations in 6 Indic languages, using 9 multilingual Transformer models (7 universal and 2 Indic-specific). To conduct this study, we introduce a novel multilingual benchmark dataset, IndicSentEval, containing approximately ~47K sentences. Our probing analysis of surface, syntactic, and semantic properties reveals that, while almost all multilingual models demonstrate consistent encoding performance for English, surprisingly, they show mixed results for Indic languages. As expected, Indic-specific multilingual models capture linguistic properties in Indic languages better than universal models. Intriguingly, universal models broadly exhibit better robustness compared to Indic-specific models, particularly under perturbations such as dropping both nouns and verbs, dropping only verbs, or keeping only nouns. Overall, this study provides valuable insights into probing and perturbation-specific strengths and weaknesses of popular multilingual Transformer-based models for different Indic languages.

The Nepali language has distinct linguistic features, especially its complex script (Devanagari script), morphology, and various dialects, which pose a unique challenge for Natural Language Understanding (NLU) tasks. While the Nepali Language Understanding Evaluation (Nep-gLUE) benchmark provides a foundation for evaluating models, it remains limited in scope, covering four tasks. This restricts their utility for comprehensive assessments of Natural Language Processing (NLP) models. To address this limitation, we introduce twelve new datasets, creating a new benchmark, the Nepali Language Understanding Evaluation (NLUE) benchmark for evaluating the performance of models across a diverse set of Natural Language Understanding (NLU) tasks. The added tasks include Single-Sentence Classification, Similarity and Paraphrase Tasks, Natural Language Inference (NLI), and General Masked Evaluation Task (GMET). Through extensive experiments, we demonstrate that existing top models struggle with the added complexity of these tasks. We also find that the best multilingual model outperforms the best monolingual models across most tasks, highlighting the need for more robust solutions tailored to the Nepali language. This expanded benchmark sets a new standard for evaluating, comparing, and advancing models, contributing significantly to the broader goal of advancing NLP research for low-resource languages.

Developing robust Natural Language Understanding (NLU) for morphologically rich Dravidian languages like Kannada, Malayalam, Tamil, and Telugu presents significant challenges due to their agglutinative nature and syntactic complexity. In this work, we present the Dravidian NLP Suite tackling five core tasks: Morphological Analysis (MA), POS Tagging (POS), Named Entity Recognition (NER), Dependency Parsing (DEP), and Coreference Resolution (CR), trained for monolingual models and multilingual models. To facilitate this, we present the Dravida dataset, meticulously annotated multilingual corpus for these tasks across all four languages. Our experiments demonstrate that a multilingual model, which utilizes shared linguistic features and cross-lingual patterns inherent to the Dravidian family, consistently outperforms its monolingual counterparts across all tasks. These findings suggest that multilingual learning is an effective approach for enhancing Natural Language Understanding (NLU) capabilities, particularly for languages belonging to the same family. To the best of our knowledge, this is the first work to jointly address all these core tasks on the Dravidian languages.

Visual Question Answering (VQA) often requires complex multi-hop reasoning encompassing both vision and language. Despite the remarkable performance of Large Multimodal Models (LMMs) in vision-language tasks, they encounter difficulties when faced with challenging scenarios that require complex reasoning and may be susceptible to object hallucination. This paper introduces a novel framework named Spatial-aware Visual Program Reasoning (SVPR). The primary goal of SVPR is to enhance the alignment between vision and language within LMMs, fostering their multi-hop reasoning abilities and ultimately strengthening their capacity to address complex visual reasoning tasks. We first utilize the strong visual understanding abilities of LMMs to generate scene graphs, facilitating coordination between vision and language at semantic levels. Then, we leverage the in-context learning ability of LMMs to generate visual programs, which guide the question decomposition process. Finally, we employ a program solver to execute the programs and derive the final answer. This process makes our approach both explanatory and robust, providing clear explanations of its reasoning process while ensuring the faithfulness of the answer to the visual input. We evaluate our framework on two challenging multi-hop multimodal VQA datasets and show its effectiveness under zero-shot settings.

Retrieval-Augmented Generation (RAG) has emerged as a powerful framework to overcome the knowledge limitations of Large Language Models (LLMs) by integrating external retrieval with language generation. While early RAG systems based on static pipelines have shown effectiveness in well-structured tasks, they struggle in real-world scenarios requiring complex reasoning, dynamic retrieval, and multi-modal integration. To address these challenges, the field has shifted toward Reasoning Agentic RAG, a paradigm that embeds decision-making and adaptive tool use directly into the retrieval process. In this paper, we present a comprehensive review of Reasoning Agentic RAG methods, categorizing them into two primary systems: predefined reasoning, which follow fixed modular pipelines to boost reasoning, and agentic reasoning, where the model autonomously orchestrates tool interaction during inference. We analyze representative techniques under both paradigms, covering architectural design, reasoning strategies, and tool coordination. Finally, we discuss key research challenges and propose future directions to advance the flexibility, robustness, and applicability of reasoning agentic RAG systems.

Open-domain Information Extraction (IE) plays an essential role in constructing large-scale knowledge bases and supports downstream applications such as Question Answering, Text Summarization, etc. While most prior research in IE has centered around extracting categorical relational tuples (e.g., president of, located in), the extraction of numerical relations (e.g., literacy rate, area, molecular weight), that link quantitative mentions to corresponding entities, remains relatively underexplored. This work addresses this gap by targeting the extraction of open-domain numeric assertions, which require identifying both the relevant entity and the appropriate measuring attribute associated with a quantity in natural language text. We begin by refining an existing OpenIE system through a rule-based approach where retrieving implicit measuring attributes for a quantity mention becomes the main challenge. To overcome this, we propose a neural framework that jointly identifies the relevant entity for a numeric mention and infers the measuring attribute to relate them, using contextual cues in the sentence. Experimental evaluation shows that our proposed model outperforms the baseline and a general-purpose large language model with a significantly large margin.

Multimodal models play a key role in empathy detection, but their performance can suffer when modalities provide conflicting cues. To understand these failures, we examine cases where unimodal and multimodal predictions diverge. Using fine-tuned models for text, audio, and video, along with a gated fusion model, we find that such disagreements often reflect underlying ambiguity, as evidenced by annotator uncertainty. Our analysis shows that dominant signals in one modality can mislead fusion when unsupported by others. We also observe that humans, like models, do not consistently benefit from multimodal input. These insights position disagreement as a useful diagnostic signal for identifying challenging examples and improving empathy system robustness.

We introduce PHLoRA (Pronounced “flora”) (Post-hoc LoRA), a simple yet powerful method to extract low-rank adaptation adapters from full-rank fine-tuned models without requiring access to training data or gradients. By computing the low-rank decomposition of weight differences between a base model and its fine-tuned counterpart, our method reconstructs adapter modules that can be merged or dynamically routed at inference time via S-LoRA, AdapterFusion, or served in scalable, industry settings using platforms like NVIDIA NIM.This approach amortizes latency overhead across requests and yields substantial cost savings. Unlike prior work that trains each adapter explicitly, our approach decouples fine-tuning from adapter generation, allowing adapter extraction from existing full-rank models or third-party checkpoints. Experiments on text, image, and video benchmarks using the Amazon Nova model family demonstrate that extracted adapters preserve high energy from the full weight delta, can be pruned safely, and yield negligible degradation in downstream task performance when re-merged. Overall, PHLoRA provides a practical path for making all existing full-rank checkpoints adapter-ready, democratizing scalable inference for all models.

Understanding how large language models (LLMs) reason across semantically distinct domains remains an open challenge. In this work, we investigate whether LLMs can connect personality traits to musical preferences, specifically chord progressions. Drawing on psychological theory and symbolic music structure, we introduce a novel benchmark that evaluates two interdependent tasks: (1) inferring personality traits from a textual context and (2) selecting a musically appropriate chord progression aligned with the inferred trait. We release a synthetic, expert-guided dataset grounded in Cattell’s 16 Personality Factors (PF16), genre-conditioned chord structures, and diverse situational contexts. We explore multiple learning strategies, including fine-tuning task-specific corpora, model merging with LoRA adapters, and advanced prompt-based reasoning techniques such as verbalization. Additionally, we propose a teacher-student framework to evaluate the quality of model-generated explanations using a five-dimensional rubric. Our findings show that verbalization outperforms standard reasoning methods, achieving up to 11% improvement over zero-shot baselines.

Multiple Choice Question (MCQ) answering is a widely used method for evaluating the performance of Large Language Models (LLMs). However, LLMs often exhibit selection bias in MCQ tasks, where their choices are influenced by factors like answer position or option symbols rather than the content. This bias undermines the reliability of MCQ as an evaluation framework. Most existing selection bias metrics require answer labels and measure divergences between prediction and answer distributions, but do not fully capture the consistency of a model’s predictions across different orderings of answer choices. Existing selection bias mitigation strategies have notable limitations: majority voting, though effective, is computationally prohibitive; calibration-based methods require validation sets and often fail to generalise across datasets. To address these gaps, we propose three key contributions: (1) a new unsupervised label-free Permutation Bias Metric (PBM) that directly quantifies inconsistencies in model predictions across answer permutations, providing a more precise measure of selection bias, (2) an efficient majority voting approach called Batch Question-Context KV caching (BaQCKV), to significantly reduce computational costs while preserving bias mitigation effectiveness, and (3) an unsupervised Low-Rank Adaptation (LoRA)-1 fine-tuning strategy based on our proposed metric and the BaQCKV that mitigates selection bias, providing a computationally efficient alternative that maintains model generalizability. Experiments across multiple MCQ benchmarks demonstrate that our approaches reduce bias, increasing consistency in accuracy while minimizing computational costs.

Social norms are implicit, culturally grounded expectations that guide interpersonal communication. Unlike factual commonsense, norm reasoning is subjective, context-dependent, and varies across cultures—posing challenges for computational models. Prior works provide valuable normative annotations but mostly target isolated utterances or synthetic dialogues, limiting their ability to capture the fluid, multi-turn nature of real-world conversations. In this work, we present Norm-RAG, a retrieval-augmented, agentic framework for nuanced social norm inference in multi-turn dialogues. Norm-RAG models utterance-level attributes including communicative intent, speaker roles, interpersonal framing, and linguistic cues and grounds them in structured normative documentation retrieved via a novel Semantic Chunking approach. This enables interpretable and context-aware reasoning about norm adherence and violation across multilingual dialogues. We further introduce MINDS (Multilingual Interactions with Norm-Driven Speech), a bilingual dataset comprising 31 multi-turn Mandarin-English and Spanish-English conversations. Each turn is annotated for norm category and adherence status using multi-annotator consensus, reflecting cross-cultural and realistic norm expression. Our experiments show that Norm-RAG improves norm detection and generalization, demonstrates improved performance for culturally adaptive and socially intelligent dialogue systems.

Large language models (LLMs), despite their remarkable text generation capabilities, often hallucinate and generate text that is factually incorrect and not grounded in real-world knowledge. This poses serious risks in domains like healthcare, finance, and customer support. A typical way to use LLMs is via the APIs provided by LLM vendors where there is no access to model weights or options to fine-tune the model. Existing methods to detect hallucinations in such settings where the model access is restricted or constrained by resources typically require making multiple LLM API calls, increasing latency and API cost. We introduce CONFACTCHECK, an efficient hallucination detection approach that does not leverage any external knowledge base and works on the simple intuition that responses to factual probes within the generated text should be consistent within a single LLM and across different LLMs. Rigorous empirical evaluation on multiple datasets that cover both the generation of factual texts and the open generation shows that CONFACTCHECK can detect hallucinated facts efficiently using fewer resources and achieves higher accuracy scores compared to existing baselines that operate under similar conditions. Our code is available here.

Natural Language to SQL (NL2SQL) has seen significant advancements with large language models (LLMs). However, these models often depend on closed-source methods and high computational resources, posing challenges in data privacy and deployment. In contrast, small language models (SLMs) struggle with NL2SQL tasks, exhibiting poor performance and incompatibility with existing frameworks. To address these issues, we introduce Feather-SQL, a new lightweight framework tailored for SLMs. Feather-SQL improves SQL executability and accuracy through: (i) schema pruning and linking, (ii) multi-path and multi-candidate generation. Additionally, we introduce 1+1 Model Collaboration Paradigm, which pairs a strong general-purpose chat model with a fine-tuned SQL model, combining strong analytical reasoning with high-precision SQL generation. Experimental results on BIRD demonstrate that Feather-SQL improves NL2SQL performance on SLMs, with around 10% boost for models without fine-tuning. The proposed paradigm raises the accuracy ceiling of SLMs to 54.76%, highlighting its effectiveness.

Vision-Language Models (VLMs) have demonstrated impressive capabilities across a range of tasks, yet concerns about their potential biases persist. This work investigates the cultural biases in state-of-the-art VLMs by evaluating their performance on an image-based country identification task at the country level. Utilizing the geographically diverse Country211 (CITATION) dataset, we probe VLMs via open-ended questions, multiple-choice questions (MCQs), and include challenging multilingual and adversarial task settings. Our analysis aims to uncover disparities in model accuracy across different countries and question formats, providing insights into how training data distribution and evaluation methodologies may influence cultural biases in VLMs. The findings highlight significant variations in performance, suggesting that while VLMs possess considerable visual understanding, they inherit biases from their pre-training data and scale, which impact their ability to generalize uniformly across diverse global contexts.

Synthetic reviews mislead users and erode trust in online marketplaces, and the advent of Large Language Models (LLMs) makes detecting such AI-generated content increasingly challenging due to their human-like fluency and coherence. In the literature, LLM-generated review detection datasets are limited to one or a few domains, with reviews generated by only a few LLMs. Consequently, datasets are limited in diversity in terms of both domain coverage and review generation styles. Models trained on such datasets generalize poorly, lacking cross-model adaptation and struggling to detect diverse LLM-generated reviews in real-world, open-domain scenarios. To address this, we introduce DetectAIRev, a benchmark dataset for AI-generated review detection that includes human-written reviews from diverse domains and AI-generated reviews generated by various categories of LLMs. We evaluate the quality and reliability of the proposed dataset through several ablation studies and human evaluations. Furthermore, we propose an AI-generated text detection method ProtoFewRoBERTa, a few-shot framework that combines prototypical networks with RoBERTa embeddings, which learn discriminative features across multiple LLMs and human-written text using only a few labeled examples per class to discriminate between LLMs as the author for text author detection. We conduct our experiments on the DetectAIRev and a publicly available benchmark dataset. Our experimental results suggest that our proposed methods outperform the state-of-the-art baseline models in detecting AI-generated reviews and text detection.

To explore a more scalable path for adding multimodal capabilities to existing LLMs, this paper addresses a fundamental question: Can a unimodal LLM, relying solely on text, reason about its own informational needs and provide effective feedback to optimize a multimodal model? To answer this, we propose a method that enables a language agent to give feedback to a vision-language model (VLM) to adapt text generation to the agent’s preferences. Our results from different experiments affirm this hypothesis, showing that LLM preference feedback significantly enhances VLM descriptions. Using our proposed method, we find that the VLM can supply multimodal scene descriptions to help the LLM better understand multimodal context, leading to improvements of maximum 13% in absolute accuracy compared to the baseline multimodal approach. Furthermore, a human study validated our AI-driven feedback, showing a 64.6% preference alignment rate between the LLM’s choices and human judgments. Extensive experiments provide insights on how and why the method works and its limitations.

Despite significant progress in large language models (LLMs), their knowledge and evaluation continue to be centered around high-resource languages, leaving critical gaps in low-resource settings. This raises questions about how effectively LLMs handle subjects that require locally relevant knowledge. To address this challenge, we need a robust dataset that reflects the knowledge of underrepresented regions such as Bangladesh. In this paper, we present ***SOMAJGYAAN***, a Bangla multiple-choice dataset consisting of 4,234 questions, annotated across five levels of difficulty. The questions are drawn from Bangladesh’s National Curriculum and Global Studies textbooks, covering a wide range of domains including History, Geography, Economics, Social Studies, Politics and Law, and Miscellaneous topics. Difficulty levels were assigned by four expert annotators to minimize annotation bias. The experiments reveal that closed-source LLMs perform better than open-source LLMs. While fine-tuning open-source models on improves their performance, they still fall short of matching closed-source LLMs. Our findings highlight the importance of culturally grounded evaluation datasets and task-specific adaptation to improve LLM performance in low-resource language settings.

Social networks extensively feature memes, particularly cartoon images, as a prevalent form of communication often conveying complex sentiments or harmful content. Detecting such content, particularly when it involves Bengali and English text, remains a multimodal challenge. This paper introduces ***CMBan***, a novel and culturally relevant dataset of 2,641 annotated cartoon memes. It addresses meme classification based on their sentiment across five key categories: Humor, Sarcasm, Offensiveness, Motivational Content, and Overall Sentiment, incorporating both image and text features. Our curated dataset specifically aids in detecting nuanced offensive content and navigating complexities of pure Bengali, English, or code-mixed Bengali-English languages. Through rigorous experimentation involving over 12 multimodal models, including monolingual, multilingual, and proprietary architectures, and utilizing prompting methods like Chain-Of-Thought (CoT), findings suggest this cartoon-based, code-mixed meme content poses substantial understanding challenges. Experimental results demonstrate that closed models excel over open models. While the LoRA fine-tuning strategy equalizes performance across model architectures and improves classification of challenging aspects in multilingual meme contexts, this work advances meme classification by providing effective solution for detecting harmful content in multilingual meme contexts.

The spread of fake news during the COVID-19 pandemic era triggered widespread chaos and confusion globally, causing public panic and misdirected health behavior. Automated fact checking in non-English languages is challenging due to the low availability of trusted resources. There are several prior work that attempted automated fact checking in multilingual settings. However, most of them fine-tune pre-trained language models (PLMs) and only produce veracity prediction without providing explanations. The absence of explanatory reasoning in these models reduces the credibility of their predictions. This paper proposes a multi-agent explainable cross-lingual fake news detection method that leverages credible English evidence and Large Language Models (LLMs) to verify and generate explanations for non-English claims, overcoming the scarcity of non-English evidence. The experimental results show that the proposed method performs well across three non-English written multilingual COVID-19 datasets in terms of veracity predictions and explanations. Our source code is available online. (https://github.com/bassamtiano/crosslingual_efnd)

The rapid progress of generative AI (Gen-AI) and large language models (LLMs) offers significant potential for geospatial applications, but simultaneously introduces critical privacy, security, and ethical risks. Existing general-purpose AI safety frameworks inadequately cover GeoAI-specific risks such as geolocation privacy violations and re-identification, with False Safe Rates exceeding 40% in some models. To address this, we present GeoSAFE (Geospatial Safety Assurance Framework and Evaluation), introducing the first GeoAI-specific safety taxonomy with six hazard categories and a multimodal GeoSAFE-Dataset. It includes 11694 textual prompts with explanations, augmented by real-world queries and images to reduce synthetic bias and reflect operational use. We benchmark model performance on detecting unsafe geospatial queries. Additionally, we present GeoSAFEGuard, an instruction-tuned LLM achieving 4.6% False Safe Rate, 0.4% False Unsafe Rate, and 97% F1-score on text-to-text evaluation of GeoSAFE-Dataset. An anonymous user-survey confirms human-GeoSAFE alignment emphasizing the urgent need for domain-specific safety evaluations as general-purpose LLMs fail to detect unsafe location-powered queries.

Simultaneous speech translation (SiST) requires balancing translation quality and latency. While most SiST systems follow machine translation assumptions that prioritize full semantic accuracy to the source, human interpreters often omit less critical content to catch up with the speaker. This study investigates whether omission can be used to reduce latency while preserving meaning in SiST.We construct a dataset that includes omission using large language models (LLMs) and propose a Target-Duration Latency (TDL), target-based latency metric that measures the output length considering the start and end timing of translation. Our analysis shows that LLMs can omit less important words while retaining the core meaning. Furthermore, experimental results show that although standard metrics overlook the benefit of the model trained with proposed omission-involving dataset, alternative evaluation methods capture it, as omission leads to shorter outputs with acceptable quality.

Reasoning over procedural sequences, where the order of steps directly impacts outcomes, is a critical capability for large language models (LLMs). In this work, we study the task of reconstructing globally ordered sequences from shuffled procedural steps, using a curated dataset of food recipes, a domain where correct sequencing is essential for task success. We evaluate several LLMs under zero-shot and few-shot settings and present a comprehensive evaluation framework that adapts established metrics from ranking and sequence alignment. These include Kendall’s Tau, Normalized Longest Common Subsequence (NLCS), and Normalized Edit Distance (NED), which capture complementary aspects of ordering quality. Our analysis shows that model performance declines with increasing sequence length, reflecting the added complexity of longer procedures. We also find that greater step displacement in the input, corresponding to more severe shuffling, leads to further degradation. These findings highlight the limitations of current LLMs in procedural reasoning, especially with longer and more disordered inputs.

Contemporary vision-language models (VLMs) perform well on existing multimodal reasoning benchmarks (78-85% accuracy on MMMU, MathVista). Yet, these results fail to sufficiently distinguish true scientific reasoning articulation capabilities from pattern-matching. To address this gap, we introduce mmJEE-Eval, a multimodal bilingual (English and Hindi) benchmark comprising 1,460 questions from India’s JEE Advanced examination (2019-2025) spanning pre-college Physics, Chemistry, and Mathematics domains. Our evaluation of 17 state-of-the-art models reveals that while frontier VLMs (GPT-5, Gemini 2.5 Pro/Flash) achieve 77-84% accuracy on held-out 2025 questions, open-source models plateau at 37-45% despite scaling to 400B parameters, a significant difference not observed on existing benchmarks. While closed frontiers from Google and OpenAI show high problem-solving accuracies (up to 100% pass@3 scores), they fully collapse when the reasoning load is increased meta-cognitively (GPT-5 fixes just 5.2% errors). Systematic ablations show mmJEE-Eval’s difficulty stems from complexity and reasoning depth rather than memorization. Effectively, our benchmark segregates superior training and reasoning methodologies where alternatives fail. We publicly release our code and data: https://mmjee-eval.github.io

Large Language Models (LLMs) are increasingly used as daily recommendation systems for tasks like education planning, yet their recommendations risk perpetuating societal biases. This paper empirically examines geographic, demographic, and economic biases in university and program suggestions from three open-source LLMs: LLaMA-3.1-8B, Gemma-7B, and Mistral-7B. Using 360 simulated user profiles varying by gender, nationality, and economic status, we analyze over 25,000 recommendations. Results show strong biases: institutions in the Global North are disproportionately favored, recommendations often reinforce gender stereotypes, and institutional repetition is prevalent. While LLaMA-3.1 achieves the highest diversity, recommending 481 unique universities across 58 countries, systemic disparities persist. To quantify these issues, we propose a novel, multi-dimensional evaluation framework that goes beyond accuracy by measuring demographic and geographic representation. Our findings highlight the urgent need for bias consideration in educational LMs to ensure equitable global access to higher education.

Small, resource-efficient language models are pivotal for extending high-quality text generation to low-resource and regional languages—the true frontier of linguistic equity in AI. Yet research largely prioritises massive English-centric systems, leaving regional-centric (low-resource) language modelling underexplored, particularly how tokenizer design, dataset diversity, and linguistic structure shape the inference of Small Language Models (SLMs) under realistic computational and data constraints. We present Regional-TinyStories, a lightweight framework that treats SLMs as cost-effective stand-ins for LLMs, enabling rapid, variable-wise inference-based analysis. Extending TinyStories to Hindi, Marathi, and Bangla, we release datasets of 2M synthetic and translated stories per language and train over 20 SLMs spanning 5–157M parameters. Using this framework, we (i) uncover contrasts between form-oriented (grammar, fluency) and content-oriented (context, completeness, creativity) metrics; (ii) chart language-specific learning dynamics; (iii) rank tokenizers, showing Indic-specific Sarvam-1 outperforming SUTRA and generic Tiktoken (GPT-2) across all metrics; and (iv) demonstrate that dataset semantic quality (translation vs. synthetic) strongly governs downstream generation. Validation through an LLM-as-Judge ensemble (GPT-4o, LLaMA-3.3-70B) and a 100+ participant human study confirms these trends while exposing systematic score inflation in automated evaluations. Regional-TinyStories offers a reproducible path to benchmark tokenizers, datasets, and SLM designs for scalable, context-faithful generation in low-resource settings.

Can today’s Text-to-SQL benchmarks still stretch modern LLMs? We argue no. Spider1.0 and BIRD, painstakingly hand-built, remain small, costly, and skewed toward middle complex SQL. Meanwhile, LLM-generated corpora are inexpensive but often superficial and fragile suffering from shallow nesting, semantic drift, template fatigue, and insufficient quality check.We address this gap with a Chain-of-Verifications framework that turns a handful of expert-labelled seeds into a large, reliably checked dataset at a fraction of the usual cost. The resulting corpus, AIGT2S, delivers: (1)18k Question–SQL pairs across 113 databases, 41–77% larger than current English sets; (2)55% queries in the Ultra band of our four-level difficulty taxonomy; (3)87.5% inter-annotator agreement; (4)≥80% labour and ≥98% monetary savings versus earlier efforts.Baselines including GPT-4o, Llama3, RESDSQL, and MAC-SQL, achieve at most 56% execution accuracy, indicating substantial room for improvement.

Soft prompts have emerged as a powerful alternative to adapters in parameter-efficient fine-tuning (PEFT), enabling large language models (LLMs) to adapt to downstream tasks without architectural changes or parameter updates. While prior work has focused on stabilizing training via parameter interaction in small neural prompt encoders, their broader potential for transfer across languages remains unexplored. In this paper, we demonstrate that a prompt encoder can play a central role in improving performance on low-performing languages—those that achieve poor accuracy even under full-model fine-tuning. We investigate a lightweight encoder paired with multi-source training on typologically diverse languages. We call this architecture-training combination the Cross-Prompt Encoder (XPE), and show that it advances the capture of abstract, transferable patterns across languages. To complement XPE, we propose a Dual Soft Prompt mechanism that combines an encoder-based prompt with a directly trained standard soft prompt. This hybrid design proves especially effective for target languages that benefit from both broadly shared structure and language-specific alignment. Text classification experiments with a transformer encoder (XLM-R) on the SIB-200 benchmark reveal a consistent trade-off: XPE is most effective for low-performing languages, while hybrid variants offer broader adaptability across multilingual settings.

Large language models (LLMs) have transformed machine translation, yet they have a high subword fertility issue for low-resource languages, which leads to slow inference speed and increased costs. While vocabulary expansion via continual pre-training is a common solution, it often degrades translation quality and requires large target-language corpora, which are unavailable for truly low-resource languages. To address this, we investigate tokenization efficiency through an information-theoretic lens, building on the established hypothesis that word length correlates with information content. From this perspective, we characterize tokenization inefficiency as having high fertility for low-information (highly predictable) words. Guided by this principle, we introduce a novel fine-tuning strategy that systematically identifies informationally redundant words—those with high fertility but low information content—for targeted vocabulary expansion and model fine-tuning. Experiments fine-tuning BLOOM and LLaMA-3 in English-Manipuri and other two language pairs show that our proposed method significantly reduces fertility by 50% and accelerates inference by more than 2 times, without compromising and often exceeding the translation quality of standard LLM baselines, providing a theoretically grounded solution for efficient LLM-based MT.

We present a novel agent-based approach for the automated claim matching task with instruction-following LLMs. We propose a two-step pipeline that first generates prompts with LLMs, to then perform claim matching as a binary classification task with LLMs. We demonstrate that LLM-generated prompts can outperform SOTA with human-generated prompts, and that smaller LLMs can do as well as larger ones in the generation process, allowing to save computational resources. We also demonstrate the effectiveness of using different LLMs for each step of the pipeline, i.e. using an LLM for prompt generation, and another for claim matching. Our investigation into the prompt generation process in turn reveals insights into the LLMs’ understanding and handling of the claim matching task.

We introduce Tooka-SBERT, a family of Persian sentence embedding models designed to enhance semantic understanding for Persian. The models are released in two sizes—Small (123M parameters) and Large (353M parameters)—both built upon the TookaBERT backbone. Tooka-SBERT is pretrained on the Targoman News corpus and fine-tuned using high-quality synthetic Persian sentence pair datasets to improve semantic alignment. We evaluate Tooka-SBERT on PTEB, a Persian adaptation of the MTEB benchmark, where the Large model achieves an average score of 70.54% and the Small model 69.49%, outperforming some strong multilingual baselines. Tooka-SBERT provides a compact and high-performing open-source solution for Persian sentence representation, with efficient inference suitable for both GPU and CPU environments. Our models are publicly available on Hugging Face, and the corresponding benchmark results can be viewed on the PTEB Leaderboard.