Accents play a pivotal role in shaping human communication, enhancing our ability to convey and comprehend messages with clarity and cultural nuance. While there has been significant progress in Automatic Speech Recognition (ASR), African-accented English ASR has been understudied due to a lack of training datasets, which are often expensive to create and demand colossal human labor. Combining several active learning paradigms and the core-set approach, we propose a new multi-round adaptation process that uses epistemic uncertainty to automate annotation, significantly reducing the associated costs and human labor. This novel method streamlines data annotation and strategically selects data samples that contribute most to model uncertainty, enhancing training efficiency. We define a new U-WER metric to track model adaptation to hard accents. We evaluate our approach across several domains, datasets, and high-performing speech models. Our results show that our approach leads to a 27% WER relative average improvement while requiring, on average, 45% less data than established baselines. Our approach also improves out-of-distribution generalization for very low-resource accents, demonstrating its viability for building generalizable ASR models in the context of accented African ASR. We open-source the code here.

Originally developed to reduce the manual burden of grading standardised language tests, Automated Essay Scoring (AES) research has long focused on holistic scoring methods which offer minimal formative feedback in the classroom. With the increasing demand for technological tools that support language acquisition, the field is turning to analytic AES (evaluating essays according to different linguistic traits). This approach holds promise for generating more detailed essay feedback, but relies on analytic scoring data that is both more cognitively demanding for humans to produce, and prone to bias. The dominant paradigm in AES is to aggregate disagreements between raters into a single gold-standard label, which fails to account for genuine examiner variability. In an attempt to make AES more representative and trustworthy, we propose to explore the sources of disagreements and lay out a novel AES system design that learns from individual raters instead of the gold standard labels.

With the advent of highly capable instruction-tuned neural language models, benchmarking in natural language processing (NLP) is increasingly shifting towards pairwise comparison leaderboards, such as LMSYS Arena, from traditional global pointwise scores (e.g., GLUE, BIG-bench, SWE-bench). This paper empirically investigates the strengths and weaknesses of both global scores and pairwise comparisons to aid decision-making in selecting appropriate model evaluation strategies. Through computational experiments on synthetic and real-world datasets using standard global metrics and the popular Bradley–Terry model for pairwise comparisons, we found that while global scores provide more reliable overall rankings, they can underestimate strong models with rare, significant errors or low confidence. Conversely, pairwise comparisons are particularly effective for identifying strong contenders among models with lower global scores, especially where quality metrics are hard to define (e.g., text generation), though they require more comparisons to converge if ties are frequent. Our code and data are available at https://github.com/HSPyroblast/srw-ranking under a permissive license.

Filtering and annotating textual data are routine tasks in many areas, like social media or news analytics. Automating these tasks allows to scale the analyses wrt. speed and breadth of content covered and decreases the manual effort required. Due to technical advancements in Natural Language Processing, specifically the success of large foundation models, a new tool for automating such annotation processes by using a text-to-text interface given written guidelines without providing training samples has become available. In this work, we assess these advancements in-the-wild by empirically testing them in an annotation task on German Twitter data about social and political European crises. We compare the prompt-based results with our human annotation and preceding classification approaches, including Naive Bayes and a BERT-based fine-tuning/domain adaptation pipeline. Our results show that the prompt-based approach – despite being limited by local computation resources during the model selection – is comparable with the fine-tuned BERT but without any annotated training data. Our findings emphasize the ongoing paradigm shift in the NLP landscape, i.e., the unification of downstream tasks and elimination of the need for pre-labeled training data.

Active learning (AL) aims to reduce annotation effort by iteratively selecting the most informative samples for labeling. The dominant strategy in AL involves fully finetuning the model on all acquired data after each round, which is computationally expensive in multilingual and low-resource settings. This paper investigates continual finetuning (CF), an alternative update strategy where the model is updated only on newly acquired samples at each round. We evaluate CF against full finetuning (FA) across 28 African languages using MasakhaNEWS and SIB-200. Our analysis reveals three key findings. First, CF matches or outperforms FA for languages included in the model’s pretraining, achieving up to 35% reductions in GPU memory, FLOPs, and training time. Second, CF performs comparably even for languages not seen during pretraining when they are typologically similar to those that were. Third, CF’s effectiveness depends critically on uncertainty-based acquisition; without it, performance deteriorates significantly. While FA remains preferable for some low-resource languages, the overall results establish CF as a robust, cost-efficient alternative for active learning in multilingual NLP. These findings motivate developing hybrid AL strategies that adapt fine-tuning behavior based on pretraining coverage, language typology, and acquisition dynamics.

Many pre-trained language models (PLMs) exhibit suboptimal performance on mid- and low-resource languages, largely due to limited exposure to these languages during pre-training. A common strategy to address this is to introduce new tokens specific to the target languages, initialize their embeddings, and apply continual pre-training on target-language data. Among such methods, OFA (Liu et al., 2024a) proposes a similarity-based subword embedding initialization heuristic that is both effective and efficient. However, OFA restricts target-language token embeddings to be convex combinations of a fixed number of source-language embeddings, which may limit expressiveness. To overcome this limitation, we propose HYPEROFA, a hypernetwork-based approach for more adaptive token embedding initialization. The hypernetwork is trained to map from an external multilingual word vector space to the PLM’s token embedding space using source-language tokens. Once trained, it can generate flexible embeddings for target-language tokens, serving as a good starting point for continual pretraining. Experiments demonstrate that HYPEROFA consistently outperforms random initialization baseline and matches or exceeds the performance of OFA in both continual pre-training convergence and downstream task performance. We make the code publicly available.

Deception is the intentional practice of twisting information. It is a nuanced societal practice deeply intertwined with human societal evolution, characterized by a multitude of facets. This research explores the problem of deception through the lens of psychology, employing a framework that categorizes deception into three forms: lies of omission, lies of commission, and lies of influence. The primary focus of this study is specifically on investigating only lies of omission. We propose a novel framework for deception detection leveraging NLP techniques. We curated an annotated dataset of 876,784 samples by amalgamating a popular large-scale fake news dataset and scraped news headlines from the Twitter handle of “Times of India”, a well-known Indian news media house. Each sample has been labeled with four layers, namely: (i) the type of omission (speculation, bias, distortion, sounds factual, and opinion), (ii) colors of lies (black, white, grey, and red), and (iii) the intention of such lies (to influence, gain social prestige, etc) (iv) topic of lies (political, educational, religious, racial, and ethnicity). We present a novel multi-task learning [MTL] pipeline that leverages the dataless merging of fine-tuned language models to address the deception detection task mentioned earlier. Our proposed model achieved an impressive F1 score of 0.87, demonstrating strong performance across all layers including the type, color, intent, and topic aspects of deceptive content. Finally, our research aims to explore the relationship between the lies of omission and propaganda techniques. To accomplish this, we conducted an in-depth analysis, uncovering compelling findings. For instance, our analysis revealed a significant correlation between loaded language and opinion, shedding light on their interconnectedness. To encourage further research in this field, we are releasing the SEPSIS dataset and code at https://huggingface.co/datasets/ankurani/deception.

Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample and high-quality unit and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1% pass@1 solve rate. With our best inference technique, which combines muti-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open source our code at https://github.com/kraritt/zolve.

While LLMs demonstrate remarkable reasoning capabilities and multi-agent applicability, their tendency to “overthink” and “groupthink” pose intriguing parallels to human cognitive limitations. Inspired by this observation, we conduct an exploratory simulation to investigate whether LLMs are wise enough to be thinkers of philosophical reflection. We design two frameworks, Philosopher and Symposium, which simulate self- and group-reflection for multi-persona in hybrid table reasoning tasks. Through experiments across four benchmarks, we discover that while introducing varied perspectives might help, LLMs tend to under-perform simpler end-to-end approaches. We reveal from close reading five emergent behaviors which strikingly resemble human cognitive closure-seeking behaviors, and identify a consistent pattern of “overthinking threshold” across all tasks, where collaborative reasoning often reaches a critical point of diminishing returns. This study sheds light on a fundamental challenge shared by both human and machine intelligence: the delicate balance between deliberation and decisiveness.

Large Language Models (LLMs) face deployment challenges due to high computational costs, and while Post-Training Quantization (PTQ) offers a solution, existing rotation-based methods struggle at very low bit-widths like 2-bit. We introduce a novel, training-free approach to construct an improved rotation matrix, addressing the limitations of current methods. The key contributions include leveraging the Walsh-Hadamard transform with sequency ordering, which clusters similar frequency components to reduce quantization error compared to standard Hadamard matrices, significantly improving performance. Furthermore, we propose a Grouped Sequency-arranged Rotation (GSR) using block-diagonal matrices with smaller Walsh blocks, effectively isolating outlier impacts and achieving performance comparable to optimization-based methods without requiring any training. Our method demonstrates robust performance on reasoning tasks and Perplexity (PPL) score on WikiText-2. Our method also enhances results even when applied over existing learned rotation techniques.

In this reproduction study, we revisit recent claims that self-attention implements kernel principal component analysis (KPCA) (Teo and Nguyen, 2024), positing that (i) value vectors V capture the eigenvectors of the Gram matrix of the keys, and (ii) that self-attention projects queries onto the principal component axes of the key matrix K in a feature space. Our analysis reveals three critical inconsistencies: (1) No alignment exists between learned self-attention value vectors and what is proposed in the KPCA perspective, with average similarity metrics (optimal cosine similarity ≤ 0.32, linear CKA (Centered Kernel Alignment) ≤ 0.11, kernel CKA ≤ 0.32) indicating negligible correspondence; (2) Reported decreases in reconstruction loss Jproj, arguably justifying the claim that the self-attentionminimizes the projection error of KPCA, are misinterpreted, as the quantities involved differ by orders of magnitude (∼ 10³); (3) Gram matrix eigenvalue statistics, introduced to justify that V captures the eigenvector of the gram matrix, are irreproducible without undocumented implementation-specific adjustments. Across 10 transformer architectures, we conclude that the KPCA interpretation of self-attention lacks empirical support.

This study proposes a novel, scalable, non-invasive and channel-independent approach for early dementia detection, particularly Alzheimer’s Disease (AD), by representing Electroencephalography (EEG) microstates as symbolic, language-like sequences. These representations are processed via text embedding and time-series deep learning models for classification. Developed on EEG data from 1001 participants across multiple countries, the proposed method achieves a high accuracy of 94.31% for AD detection. By eliminating the need for fixed EEG configurations and costly/invasive modalities, the introduced approach improves generalisability and enables cost-effective deployment without requiring separate AI models or specific devices. It facilitates scalable and accessible dementia screening, supporting timely interventions and enhancing AD detection in resource-limited communities.

Language tagging, a method whereby source and target inputs are prefixed with a unique language token, has become the de facto standard for conditioning Multilingual Neural Machine Translation (MNMT) models on specific language directions. This conditioning can manifest effective zero-shot translation abilities in MT models at scale for many languages. Expanding on previous work, we propose a novel method of language tagging for MNMT, injection, in which the embedded representation of a language token is concatenated to the input of every linear layer. We explore a variety of different tagging methods, with and without injection, showing that injection improves zero-shot translation performance with up to a 2+ BLEU score point gain for certain language directions in our dataset.

Music has long served as a vehicle for political expression, with protest songs playing a central role in articulating dissent and mobilizing collective action. Yet, despite their cultural significance, the linguistic and acoustic signatures that define protest music remain understudied. We present a multimodal computational analysis of protest and non-protest songs spanning multiple decades. Using NLP and audio analysis, we identify the linguistic and musical features that differentiate protest songs. Instead of focusing on classification performance, we treat classification as a diagnostic tool to investigate these features and reveal broader patterns. Protest songs are not just politically charged they are acoustically and linguistically distinct, and we quantify how.

Curriculum learning is a widely adopted training strategy in natural language processing (NLP), where models are exposed to examples organized by increasing difficulty to enhance learning efficiency and performance. However, most existing approaches rely on manually defined difficulty metrics – such as text length – which may not accurately reflect the model’s own perspective. To overcome this limitation, we present a self-adaptive curriculum learning paradigm that prioritizes fine-tuning examples based on difficulty scores predicted by pre-trained language models (PLMs) themselves. Building on these scores, we explore various training strategies that differ in the ordering of examples for the fine-tuning: from easy-to-hard, hard-to-easy, to mixed sampling. We evaluate our method on four natural language understanding (NLU) datasets covering both binary and multi-class classification tasks.Experimental results show that our approach leads to faster convergence and improved performance compared to standard random sampling.

This paper introduces CausalGraphBench, a benchmark designed to evaluate the ability of large language models (LLMs) to construct Causal Graphs (CGs), a critical component of reasoning models like Bayesian Networks. The benchmark comprises 35 CGs sourced from publicly available repositories and academic papers, each enriched with detailed metadata to facilitate systematic and consistent evaluation. We explore various LLM-driven methods for CG discovery, analyzing their performance across different graph sizes and complexity levels. Additionally, we examine the effects of data contamination on the quality of the generated CGs.Our findings reveal that methods relying on approaches with a limited number of queries to LLM, particularly those leveraging the full graph context, consistently outperform query-intensive and exhaustive approaches, which tend to overemphasize local relationships. Across all methods, performance declines as graph size increases.

As Large Language Models (LLMs) continue to advance in capability, prompt engineering has emerged as a crucial method for optimizing their performance on specialized tasks. While prompting strategies like Zero-shot, Few-shot, Chain-of-Thought, and Tree-of-Thought have demonstrated significant improvements in reasoning tasks, their application to machine translation has received comparatively less attention. This paper systematically evaluates these prompting techniques across diverse language pairs and domains, measuring their effect on translation quality. Our findings reveal substantial performance variations between prompting methods, with certain strategies offering consistent improvements for specific language directions and complexity levels. These results provide valuable insights for developing more effective LLM-based translation systems without requiring model fine-tuning and complement existing works in the field.

Prompt engineering has made significant contributions to the era of large language models, yet its effectiveness depends on the skills of a prompt author. This paper introduces iPrOp, a novel interactive prompt optimization approach, to bridge manual prompt engineering and automatic prompt optimization while offering users the flexibility to assess evolving prompts. We aim to provide users with task-specific guidance to enhance human engagement in the optimization process, which is structured through prompt variations, informative instances, predictions generated by large language models along with their corresponding explanations, and relevant performance metrics. This approach empowers users to choose and further refine the prompts based on their individual preferences and needs. It can not only assist non-technical domain experts in generating optimal prompts tailored to their specific tasks or domains, but also enable to study the intrinsic parameters that influence the performance of prompt optimization. The evaluation shows that our approach has the capability to generate improved prompts, leading to enhanced task performance.

We present a comparative analysis of the parseability of structured outputs generated by small language models for open attribute-value extraction from clinical notes. We evaluate three widely used serialization formats: JSON, YAML, and XML, and find that JSON consistently yields the highest parseability. Structural robustness improves with targeted prompting and larger models, but declines for longer documents and certain note types. Our error analysis identifies recurring format-specific failure patterns. These findings offer practical guidance for selecting serialization formats and designing prompts when deploying language models in privacy-sensitive clinical settings.

Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo & Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets—either collected from the Web or generated by another model—which may contain out-of-distribution (OOD) data beyond the model’s generalisation capabilities. This can result in hallucinated SAE features, which we term ”Fake Features”, that misrepresent the model’s internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model’s own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on webbased datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.

Large language models (LLMs) have advanced natural language processing by understanding, generating, and manipulating texts.Although recent studies have shown that prompt engineering can reduce computational effort and potentially improve translation quality, prompt designs specific to different domains remain challenging. Besides, movie subtitle translation is particularly challenging and understudied, as it involves handling colloquial language, preserving cultural nuances, and requires contextual information such as the movie’s theme and storyline to ensure accurate meaning. This study aims to fill this gap by focusing on the translation of movie subtitles through the use of prompting strategies that incorporate the movie’s meta-information, e.g., movie title, summary, and genre. We build a multilingual dataset which aligns the OpenSubtitles dataset with their corresponding Wikipedia articles and investigate different prompts and their effect on translation performance. Our experiments with GPT-3.5, GPT-4o, and LLaMA-3 models have shown that the presence of meta-information improves translation accuracy. These findings further emphasize the importance of designing appropriate prompts and highlight the potential of LLMs to enhance subtitle translation quality.

Puns, as a unique form of linguistic creativity, present significant challenges in cross-lingual translation, particularly between linguistically distant languages like Chinese and English, where it’s often considered a “mission impossible”. We introduce Pun2Pun, a novel benchmark for quantitatively evaluating pun translation between Chinese and English while preserving both linguistic mechanisms and humorous effects. We propose the adaptation of Constant-Variable Optimization (CVO) Model for translation strategy and concomitant Overlap (Ovl) metric for translation quality assessment. Our approach provides a robust quantitative evaluation framework to assess models’ complex linguistic and cultural reasoning capabilities in pun translation. Through extensive experiments on both textual and visual puns, we demonstrate that our translation strategy model significantly improves performance, particularly for better-performing models. Our findings reveal exciting potentials and current limitations of LLMs in preserving sophisticated humor across linguistic and cultural boundaries.

Low-resource languages (LRLs) face significant challenges in natural language processing (NLP) due to limited data. While current state-of-the-art large language models (LLMs) still struggle with LRLs, smaller multilingual models (mLMs) such as mBERT and XLM-R offer greater promise due to a better fit of their capacity to low training data sizes. This study systematically investigates parameter-efficient adapter-based methods for adapting mLMs to LRLs, evaluating three architectures: Sequential Bottleneck, Invertible Bottleneck, and Low-Rank Adaptation. Using unstructured text from GlotCC and structured knowledge from ConceptNet, we show that small adaptation datasets (e.g., up to 1 GB of free-text or a few MB of knowledge graph data) yield gains in intrinsic (masked language modeling) and extrinsic tasks (topic classification, sentiment analysis, and named entity recognition). We find that Sequential Bottleneck adapters excel in language modeling, while Invertible Bottleneck adapters slightly outperform other methods on downstream tasks due to better embedding alignment and larger parameter counts. Adapter-based methods match or outperform full fine-tuning while using far fewer parameters, and smaller mLMs prove more effective for LRLs than massive LLMs like LLaMA-3, GPT-4, and DeepSeek-R1-based distilled models. While adaptation improves performance, pre-training data size remains the dominant factor, especially for languages with extensive pre-training coverage.The code for our experiments is available at https://github.com/d-gurgurov/Knowledge-Driven-Adaptation-LLMs.

We explore how different types of nominal compound complexity in scientific writing, in particular different types of compound structure, affect the reading times of experts and novices. We consider both in-domain and out-of-domain reading and use PoTeC (Jakobi et al. 2024), a corpus containing eye-tracking data of German native speakers reading passages from scientific textbooks. Our results suggest that some compound types are associated with longer reading times and that experts may not only have an advantage while reading in-domain texts, but also while reading out-of-domain.

This study evaluates Direct Preference Optimization (DPO) and its variants for aligning Large Language Models (LLMs) with human preferences, testing three configurations: (1) with Supervised Fine-Tuning (SFT), (2) without SFT, and (3) without SFT but using an instruction-tuned model. We further investigate how training set size influences model performance. Our evaluation spans 13 benchmarks—covering dialogue, reasoning, mathematical problem-solving, question answering, truthfulness, MT-Bench, Big Bench, and the Open LLM Leaderboard. We find that: (1) alignment methods often achieve near-optimal performance even with smaller subsets of training data; (2) although they offer limited improvements on complex reasoning tasks, they enhance mathematical problem-solving; and (3) using an instruction-tuned model improves truthfulness. These insights highlight the conditions under which alignment methods excel, as well as their limitations.

Retrieval-Augmented Generation (RAG) has emerged as a crucial framework in natural language processing (NLP), improving factual consistency and reducing hallucinations by integrating external document retrieval with large language models (LLMs). However, the effectiveness of RAG is often hindered by coreferential complexity in retrieved documents, which can introduce ambiguity and interfere with in-context learning. In this study, we systematically investigate how entity coreference affects both document retrieval and generative performance in RAG-based systems, focusing on retrieval relevance, contextual understanding, and overall response quality. We demonstrate that coreference resolution enhances retrieval effectiveness and improves question-answering (QA) performance. Through comparative analysis of different pooling strategies in retrieval tasks, we find that mean pooling demonstrates superior context capturing ability after applying coreference resolution. In QA tasks, we discover that smaller models show greater improvement from the disambiguation process, likely due to their limited inherent capacity for handling referential ambiguity. With these findings, this study aims to provide a deeper understanding of the challenges posed by coreferential complexity in RAG, offering guidance for improving retrieval and generation in knowledge-intensive AI applications.

Several recent works have examined the generations produced by large language models (LLMs) on subjective topics such as political opinions and attitudinal questionnaires. There is growing interest in controlling these outputs to align with specific users or perspectives using model steering techniques. However, several studies have highlighted unintended and unexpected steering effects, where minor changes in the prompt or irrelevant contextual cues influence model-generated opinions.This work empirically tests how irrelevant information can systematically bias model opinions in specific directions. Using the Political Compass Test questionnaire, we conduct a detailed statistical analysis to quantify these shifts using the opinions generated by LLMs in an open-generation setting. The results demonstrate that even seemingly unrelated contexts consistently alter model responses in predictable ways, further highlighting challenges in ensuring the robustness and reliability of LLMs when generating opinions on subjective topics.

Although LLMs have made significant progress in various languages, there are still concerns about their effectiveness with low-resource agglutinative languages compared to languages such as English. In this study, we focused on Korean, a language known for its complex sentence endings, and evaluated LLMs on this challenging aspect. We introduce the Korean Sentence Endings (KoSEnd) dataset, which includes 3,000 sentences, each annotated for the naturalness of 15 sentence ending forms. These were collected from diverse sources to cover a range of contexts. We evaluated 11 LLMs to assess their understanding of Korean sentence endings, analyzing them based on parameter count and prediction consistency. Notably, we found that informing models about the possibility of missing sentence endings improved performance, highlighting the impact of explicitly considering certain linguistic features.

In the field of Natural Language Processing (NLP), a common approach for resolving human disagreement involves establishing a consensus among multiple annotators. However, previous research shows that overlooking individual opinions can result in the marginalization of minority perspectives, particularly in subjective tasks, where annotators may systematically disagree due to their personal preferences. Emerging Multi-Perspective approaches challenge traditional methodologies that treat disagreement as mere noise, instead recognizing it as a valuable source of knowledge shaped by annotators’ diverse backgrounds, life experiences, and values.This thesis proposal aims to (1) identify the challenges of designing disaggregated datasets i.e., preserving individual labels in human-annotated datasets for subjective tasks (2) propose solutions for developing Perspective-Aware by design systems and (3) explore the correlation between human disagreement and model uncertainty leveraging eXplainable AI techniques (XAI).Our long-term goal is to create a framework adaptable to various subjective NLP tasks to promote the development of more responsible and inclusive models.

Large Language Models (LLMs) offer transformative potential for Software Requirements Engineering (SRE), yet critical challenges, including domain ignorance, hallucinations, and high computational costs, hinder their adoption. This paper proposes a conceptual framework that integrates Small Language Models (SLMs) and Knowledge-Augmented LMs (KALMs) with LangChain to address these limitations systematically. Our approach combines: (1) SLMs for efficient, locally deployable requirements processing, (2) KALMs enhanced with Retrieval-Augmented Generation (RAG) to mitigate domain-specific gaps, and (3) LangChain for structured, secure workflow orchestration. We identify and categorize six technical challenges and two research gaps through a systematic review of LLM applications in SRE. To guide practitioners, we distill evidence-based prompt engineering guidelines (Context, Language, Examples, Keywords) and propose prompting strategies (e.g., Chain-of-Verification) to improve output reliability. The paper establishes a theoretical foundation for scalable, trustworthy AI-assisted SRE and outlines future directions, including domain-specific prompt templates and hybrid validation pipelines.

Grounding large language models (LLMs) in verifiable external sources is a well-established strategy for generating reliable answers. Retrieval-augmented generation (RAG) is one such approach, particularly effective for tasks like question answering: it retrieves passages that are semantically related to the question and then conditions the model on this evidence. However, multi-hop questions, such as “Which company among NVIDIA, Apple, and Google made the biggest profit in 2023?,” challenge RAG because relevant facts are often distributed across multiple documents rather than co-occurring in one source, making it difficult for standard RAG to retrieve sufficient information. To address this, we propose a RAG pipeline that incorporates question decomposition: (i) an LLM decomposes the original query into sub-questions, (ii) passages are retrieved for each sub-question, and (iii) the merged candidate pool is reranked to improve the coverage and precision of the retrieved evidence. We show that question decomposition effectively assembles complementary documents, while reranking reduces noise and promotes the most relevant passages before answer generation. We evaluate our approach on the MultiHop-RAG and HotpotQA, showing gains in retrieval (MRR@10: +36.7%) and answer accuracy (F1: +11.6%) over standard RAG baselines. The pipeline is a practical, drop-in enhancement requiring no task-specific training or specialized indexing.

In Recent years, advances in Neural Machine Translation (NMT) heavily rely on large-scale parallel corpora. Within the context of China’s Belt and Road Initiative, there is increasing demand for improving translation quality from agglutinative languages (e.g., Mongolian, Arabic) to Chinese. However, the translation scenarios for agglutinative languages (which form words by concatenating morphemes with clear boundaries) face significant challenges including data sparsity, quality imbalance, and inactive sample proliferation due to their morphological complexity and syntactic flexibility. This study presents a systematic analysis of data distribution characteristics in agglutinative languages and proposes a dual-module framework combining fine-grained inactive sample identification with target-side rejuvenation. Our framework first establishes a multi-dimensional evaluation system to accurately identify samples exhibiting low-frequency morphological interference or long-range word order mismatches. Subsequently, the target-side rejuvenation mechanism generates diversified noise-resistant translations through iterative optimization of sample contribution weights. Experimental results on four low-resource agglutinative language tasks demonstrate significant performance improvements (BLEU +2.1–3.4) across mainstream NMT architectures. Architecture-agnostic validation further confirms the framework’s generalizability.

Structured code comments in docstring format are essential for code comprehension and maintenance, but existing machine learning models for their generation perform poorly for Russian compared to English. To bridge this gap, we present StRuCom — the first large-scale dataset (153K examples) specifically designed for Russian code documentation. Unlike machine-translated English datasets that distort terminology (e.g., technical loanwords vs. literal translations) and docstring structures, StRuCom combines human-written comments from Russian GitHub repositories with synthetically generated ones, ensuring compliance with Python, Java, JavaScript, C#, and Go standards through automated validation.

Back-translation has been proven effective in enhancing the performance of Neural Machine Translation (NMT), with its core mechanism relying on synthesizing parallel corpora to strengthen model training. However, while traditional back-translation methods alleviate the data scarcity in low-resource machine translation, their dependence on random sampling strategies ignores the semantic quality of monolingual data. This results in the contamination of model training through the inclusion of substantial low-quality samples in the generated corpora. To mitigate noise interference, additional training iterations or model scaling are required, significantly increasing computational costs. To address this challenge, this study proposes a Semantic Uncertainty Sampling strategy, which prioritizes sentences with higher semantic uncertainty as training samples by computationally evaluating the complexity of unannotated monolingual data. Experiments were conducted on three typical low-resource agglutinative language pairs: Mongolian-Chinese, Uyghur-Chinese, and Korean-Chinese. Results demonstrate an average BLEU score improvement of +1.7 on test sets across all three translation tasks, confirming the method’s effectiveness in enhancing translation accuracy and fluency. This approach provides a novel pathway for the efficient utilization of unannotated data in low-resource language scenarios.

The task of automatic dialect classification is typically tackled using traditional machine-learning models with bag-of-words unigram features. We explore two alternative methods for distinguishing dialects across 20 Spanish-speaking countries:(i) Support vector machine and decision tree models were trained on dialectal features tailored to the Spanish dialects, combined with standard unigrams. (ii) A pre-trained BERT model was fine-tuned on the task.Results show that the tailored features generally did not have a positive impact on traditional model performance, but provide a salient way of representing dialects in a content-agnostic manner. The BERT model wins over traditional models but with only a tiny margin, while sacrificing explainability and interpretability.

As the use of Large Language Models (LLMs) expands, so do concerns about their vulnerability to jailbreak attacks. We introduce SequentialBreak, a novel single-query jailbreak technique that arranges multiple benign prompts in sequence with a hidden malicious instruction among them to bypass safety mechanisms. Sequential prompt chains in a single query can lead LLMs to focus on certain prompts while ignoring others. By embedding a malicious prompt within a prompt chain, we show that LLMs tend to ignore the harmful context and respond to all prompts including the harmful one. We demonstrate the effectiveness of our attack across diverse scenarios—including Q&A systems, dialogue completion tasks, and levelwise gaming scenario—highlighting its adaptability to varied prompt structures. The variability of prompt structures shows that SequentialBreak is adaptable to formats beyond those discussed here. Experiments show that SequentialBreak only uses a single query to significantly outperform existing baselines on both open-source and closed-source models. These findings underline the urgent need for more robust defenses against prompt-based attacks. The Results and website are available on https://anonymous.4open.science/r/JailBreakAttack-4F3B/.

As large language models (LLMs) are increasingly deployed across diverse linguistic and cultural contexts, understanding their behavior in both factual and disputable scenarios is essential—especially when their outputs may shape public opinion or reinforce dominant narratives. In this paper, we define two types of bias in LLMs: model bias (bias stemming from model training) and inference bias (bias induced by the language of the query), through a two-phase evaluation.Phase 1 evaluates LLMs on factual questions where a single verifiable answer exists, assessing whether models maintain consistency across different query languages. Phase 2 expands the scope by probing geopolitically sensitive disputes, where responses may reflect culturally embedded or ideologically aligned perspectives. We construct a manually curated dataset spanning both factual and disputable QA, across four languages and question types. The results show that Phase 1 exhibits query language-induced alignment, while Phase 2 reflects an interplay between the model’s training context and query language. This paper offers a structured framework for evaluating LLM behavior across neutral and sensitive topics, providing insights for future LLM deployment and culturally-aware evaluation practices in multilingual contexts.WARNING: this paper covers East Asian issues which may be politically sensitive.

Recognizing biomedical concepts in the text is vital for ontology refinement, knowledge graph construction, and concept relationship discovery. However, traditional concept recognition methods, relying on explicit mention identification, often fail to capture complex concepts not explicitly stated in the text. To overcome this limitation, we introduce MA-COIR, a framework that reformulates concept recognition as an indexing-recognition task. By assigning semantic search indexes (ssIDs) to concepts, MA-COIR resolves ambiguities in ontology entries and enhances recognition efficiency. Using a pretrained BART-based model fine-tuned on small datasets, our approach reduces computational requirements to facilitate adoption by domain experts. Furthermore, we incorporate large language model (LLM)-generated queries and synthetic data to improve recognition in low-resource settings. Experimental results on three scenarios (CDR, HPO, and HOIP) highlight the effectiveness of MA-COIR in recognizing both explicit and implicit concepts without the need for mention-level annotations during inference, advancing ontology-driven concept recognition in biomedical domain applications. Our code and constructed data are available at https://github.com/sl-633/macoir-master.

Open-source AI libraries are foundational to modern AI systems, yet they present significant, underexamined risks spanning security, licensing, maintenance, supply chain integrity, and regulatory compliance. We introduce LibVulnWatch, a system that leverages recent advances in large language models and agentic workflows to perform deep, evidence-based evaluations of these libraries. Built on a graph-based orchestration of specialized agents, the framework extracts, verifies, and quantifies risk using information from repositories, documentation, and vulnerability databases. LibVulnWatch produces reproducible, governance-aligned scores across five critical domains, publishing results to a public leaderboard for ongoing ecosystem monitoring. Applied to 20 widely used libraries—including ML frameworks, LLM inference engines, and agent orchestration tools—our approach covers up to 88% of OpenSSF Scorecard checks while surfacing up to 19 additional risks per library, such as critical RCE vulnerabilities, missing SBOMs, and regulatory gaps. By integrating advanced language technologies with the practical demands of software risk assessment, this work demonstrates a scalable, transparent mechanism for continuous supply chain evaluation and informed library selection.

The cross-modal grounding of LLMs has recently garnered significant attention, while grounding them in textual interactions has been less explored. As the first of its kind, the GLAM framework utilises LLMs as agents in interactive text-based games to investigate their grounding capabilities. However, it faces the challenge of low computational efficiency, which hinders further experiments. This paper proposes the use of Lookahead models for action selection, demonstrating through empirical results that the approach can substantially improve training speed, achieving performance gains relative to the size of the action space.

Large language models (LLMs) are widelyused in search engines to provide direct an-swers, while AI chatbots retrieve updated infor-mation from the web. As these systems influ-ence how billions access information, evaluat-ing the credibility of news outlets has becomecrucial. We audit nine LLMs from OpenAI,Google, and Meta to assess their ability to eval-uate the credibility and political bias of the top20 most popular news outlets in Bangladesh.While most LLMs rate the tested outlets, largermodels often refuse to rate sources due to in-sufficient information, while smaller modelsare more prone to hallucinations. We create adataset of credibility ratings and political iden-tities based on journalism experts’ opinions andcompare these with LLM responses. We findstrong internal consistency in LLM credibil-ity ratings, with an average correlation coeffi-cient (ρ) of 0.72, but moderate alignment withexpert evaluations, with an average ρ of 0.45.Most LLMs (GPT-4, GPT-4o-mini, Llama 3.3,Llama-3.1-70B, Llama 3.1 8B, and Gemini 1.5Pro) in their default configurations favor theleft-leaning Bangladesh Awami League, givinghigher credibility ratings, and show misalign-ment with human experts. These findings high-light the significant role of LLMs in shapingnews and political information

Generation Alpha (born 2010-2024) is the first generation fully raised within the digital ecosystem. They exhibit unique linguistic behaviours influenced by rampant online communication and platform-specific cultures. This study examines the rapid evolution of Gen Alpha slang through a comparative analysis of Millennial and Gen Z vernacular. We identify three core linguistic patterns: extreme lexical compression, digital culture-driven semantic shifts and part-of-speech conversion. We construct a comprehensive slang corpus sourced from online platforms and evaluate the performance of four AI translation systems (viz. Google Translate, ChatGPT 4, Gemini 1.0, DeepSeek v3) on over 100 slang terms. Our results reveal significant translation challenges rooted in culturally-bound terms from gaming, meme culture, and mental health discourse. Most errors are the result of inadequate cultural contextualization, with literal translations dominating the error patterns. Our findings highlight the critical limitations in current language models and emphasize the need for adaptive, culturally sensitive and context-aware frameworks that can handle the dynamic lexicon of evolving youth vernacular.

We propose a simple and light-weight, yet effective hallucination detection method for conditional text generation. Hallucinated outputs include information that is either absent from and/or difficult to infer from the input context. Leveraging this feature, we add contrastive learning to the hallucination detection classifier to pull faithful outputs and input contexts together while pushing hallucinated outputs apart. Experimental results confirm that our method on top of RoBERTa improves binary hallucination detection performance, outperforming much larger GPT-4o prompting. Remarkably, our method shows higher performance for outputs where hallucinated spans are sparse.

Digitized literary corpora of the 19th century favor canonical and novelistic forms, sidelining a broader and more diverse literary production. Serialized fiction – widely read but embedded in newspapers – remains especially underexplored, particularly in low-resource languages like Danish. This paper addresses this gap by developing methods to identify fiction in digitized Danish newspapers (1818–1848).We (1) introduce a manually annotated dataset of 1,394 articles and (2) evaluate classification pipelines using both selected linguistic features and embeddings, achieving F1-scores of up to 0.91. Finally, we (3) analyze feuilleton fiction via interpretable features to test its drift in discourse from neighboring nonfiction.Our results support the construction of alternative literary corpora and contribute to ongoing work on modeling the fiction–nonfiction boundary by operationalizing discourse-level distinctions at scale.

Poetry has always distinguished itself from other literary genres in many ways, including grammatically and syntactically. These differences are evident not only in modern literature but also in earlier stages. Linguistic analysis tools struggle to address these differences. This paper focuses on the dichotomy between Old English poetry and prose, specifically in the context of the POS tagging task.Two annotated corpora representing each genre were analyzed to show that there are several types of structural differences between Old English poetry and prose. For POS tagging, we conduct experiments on both a detailed tag set with over 200 tags and a mapping to the UPOS tag set with 17 tags. We establish a baseline and conduct two cross-genre experiments to investigate the effect of different proportions of prose and poetry data. Across both tag sets, our results indicate that if the divergence between two genres is substantial, simply increasing the quantity of training data from the support genre does not necessarily improve prediction accuracy. However, incorporating even a small amount of target data can lead to better performance compared to excluding it entirely. This study not only highlights the linguistic differences between Old English poetry and prose but also emphasizes the importance of developing effective NLP tools for underrepresented historical languages across all genres.

This Ph.D. proposal introduces a plan to develop a computational framework to identify Self-aspects in text. The Self is a multifaceted construct and it is reflected in language. While it is described across disciplines like cognitive science and phenomenology, it remains underexplored in natural language processing (NLP). Many of the aspects of the Self align with psychological and other well-researched phenomena (e.g., those related to mental health), highlighting the need for systematic NLP-based analysis. In line with this, we plan to introduce an ontology of Self-aspects and a gold-standard annotated dataset. Using this foundation, we will develop and evaluate conventional discriminative models, generative large language models, and embedding-based retrieval approaches against four main criteria: interpretability, ground-truth adherence, accuracy, and computational efficiency. Top-performing models will be applied in case studies in mental health and empirical phenomenology.

This paper presents a workflow that compels an audio-enabled large language model to recite Latin poetry with metrically accurate stress. One hundred hexameters from the Aeneid and the opening elegiac epistula of Ovid’s Heroides constitute the test bed, drawn from the Pedecerto XML corpus, where ictic syllables are marked. A preprocessing pipeline syllabifies each line, converts alien graphemes into approximate English-Italian counterparts, merges obligatory elisions, adds commas on caesurae, upper-cases every ictic syllable, and places a grave accent on its vowel. Verses are then supplied, one at a time, to an LLM-based Text-to-Speech model under a compact system prompt that instructs slow, articulated delivery. From ten stochastic realisations per verse, a team of Latin experts retained the best; at least one fully correct file was found for 91% of the 216 lines. Upper-casing plus accent marking proved the strongest cue, while hyphenating syllables offered no benefit. Remaining errors cluster around cognates where the model inherits a Romance or English stress template. The corpus of validated audio and all scripts are openly released on Zenodo, opening avenues for pedagogy, accessibility, and prosodic research.

Agents controlled by Large Language Models (LLMs) can assist with natural language tasks across domains and applications when given access to confidential data.When such digital assistants interact with their potentially adversarial environment, confidentiality of the data is at stake.We investigated whether an LLM-controlled agent can, in a manner similar to humans, consider confidentiality when responding to natural language requests involving internal data.For evaluation, we created a synthetic dataset consisting of confidentiality-aware planning and deduction tasks in organizational access control.The dataset was developed from human input, LLM-generated content, and existing datasets.It includes various everyday scenarios in which access to confidential or private information is requested.We utilized our dataset to evaluate the ability to infer confidentiality-aware behavior in such scenarios by differentiating between legitimate and illegitimate access requests.We compared a prompting-based and a fine-tuning-based approach, to evaluate the performance of Llama 3 and GPT-4o-mini in this domain.In addition, we conducted a user study to establish a baseline for human evaluation performance in these tasks. We found humans reached an accuracy of up to 79%.Prompting techniques, such as chain-of-thought and few-shot prompting, yielded promising results, but still fell short of real-world applicability and do not surpass human baseline performance. However, we found that fine-tuning significantly improves the agent’s access decisions, reaching up to 98% accuracy, making it promising for future confidentiality-aware applications when data is available.

We present a new dataset for chart question answering (CQA) constructed from visualization notebooks. The dataset features real-world, multi-view charts paired with natural language questions grounded in analytical narratives. Unlike prior benchmarks, our data reflects ecologically valid reasoning workflows. Benchmarking state-of-the-art multimodal large language models reveals a significant performance gap, with GPT-4.1 achieving an accuracy of 69.3%, underscoring the challenges posed by this more authentic CQA setting.

As Large Language Models (LLMs) become increasingly widespread, understanding how specific training data shapes their outputs is crucial for transparency, accountability, privacy, and fairness. To explore how LLMs leverage and replicate their training data, we introduce a systematic approach centered on analyzing low-perplexity sequences—high-probability text spans generated by the model. Our pipeline reliably extracts such long sequences across diverse topics while avoiding degeneration, then traces them back to their sources in the training data. Surprisingly, we find that a substantial portion of these low-perplexity spans cannot be mapped to the corpus. For those that do match, we quantify the distribution of occurrences across source documents, highlighting the scope and nature of verbatim recall and paving a way toward better understanding of how LLMs training data impacts their behavior.

Tables are extensively utilized to represent and store data, however, they often lack explicit semantics necessary for machine interpretation of their contents. Semantic table interpretation is essential for integrating structured data with knowledge graphs, yet existing methods face challenges with Russian-language tables due to limited labeled data and linguistic peculiarities. This paper introduces a contrastive learning approach to minimize reliance on manual labeling and enhance the accuracy of column annotation for rare semantic types. The proposed method adapts contrastive learning for tabular data through augmentations and employs a distilled multilingual BERT model trained on the unlabeled RWT corpus (comprising 7.4 million columns). The resulting table representations are incorporated into the RuTaBERT pipeline, reducing computational overhead. Experimental results demonstrate a micro-F1 score of 97% and a macro-F1 score of 92%, surpassing several baseline approaches. These findings emphasize the efficiency of the proposed method in addressing data sparsity and handling unique features of the Russian language. The results further confirm that contrastive learning effectively captures semantic similarities among columns without explicit supervision, which is particularly vital for rare data types.

Large Language Models (LLMs) demonstrate strong performance on different language tasks, but tend to hallucinate – generate plausible but factually incorrect outputs. Recently, several approaches to integrate Knowledge Graphs (KGs) into LLM inference were published to reduce hallucinations. This paper presents a systematic literature review (SLR) of such approaches. Following established SLR methodology, we identified relevant work by systematically search in different academic online libraries and applying a selection process. Nine publications were chosen for in-depth analysis. Our synthesis reveals differences and similarities of how the KG is accessed, traversed, and how the context is finally assembled. KG integration can significantly improve LLM performance on benchmark datasets and additionally to mitigate hallucination enhance reasoning capabilities, explainability, and access to domain-specific knowledge. We also point out current limitations and outline directions for future work.

Existing approaches to fine-grained emotion classification (FEC) often operate in Euclidean space, where the flat geometry limits the ability to distinguish semantically similar emotion labels (e.g., *annoyed* vs. *angry*). While prior research has explored hyperbolic geometry to capture fine-grained label distinctions, it typically relies on predefined hierarchies and ignores semantically similar negative labels that can mislead the model into making incorrect predictions. In this work, we propose HyCoEM (Hyperbolic Contrastive Learning for Emotion Classification), a semantic alignment framework that leverages the Lorentz model of hyperbolic space. Our approach embeds text and label representations into hyperbolic space via the exponential map, and employs a contrastive loss to bring text embeddings closer to their true labels while pushing them away from adaptively selected, semantically similar negatives. This enables the model to learn label embeddings without relying on a predefined hierarchy and better captures subtle distinctions by incorporating information from both positive and challenging negative labels. Experimental results on two benchmark FEC datasets demonstrate the effectiveness of our approach over baseline methods.

We introduce the first dependency treebank containing Universal Dependencies (UD) annotations for Spanish learner writing from the UC Davis COWSL2H corpus. Our annotations include lemmatization, POS tagging, and syntactic dependencies. We adapt the existing UD framework for Spanish L1 to account forlearner-specific features such as code-switching and non-canonical syntax. A suite of parsing evaluation experiments shows that parsers trained on learner data together with moderate sizes of Spanish L1 data can yield reasonable performance. Our annotations are openly accessible to motivate future development of learner-oriented language technologies.

Large language models (LLMs), which are primarily trained on high-resource programming languages (HRPLs), tend to perform sub-optimally for low-resource programming languages (LRPLs). This study investigates the impact of tokenizer adaptation methods on improving code generation for LRPLs. StarCoder 2 and DeepSeek-Coder models adapted to Elixir and Racket using methods such as Fast Vocabulary Transfer (FVT), FOCUS, and Zero-shot Tokenizer Transfer (ZeTT) are evaluated and compared with the original and fine-tuned models. Our experiments reveal that ZeTT outperforms other methods, achieving significant improvements in handling syntax, program logic, and data types for LRPLs. However, we also highlight performance declines in non-target languages like Python after tokenizer adaptation. The study approves the positive impact of tokenizer adaptation in enhancing LRPL code generation and suggests directions for future research, including token embeddings improvement.

Controlling the output of Large Language Models (LLMs) through context-sensitive constraints has emerged as a promising approach to overcome the limitations of Context-Free Grammars (CFGs) in guaranteeing generation validity. However, such constraints typically require manual specification—a significant barrier demanding specialized expertise. We introduce a framework that automatically learns context-sensitive constraints from LLM interactions through a two-phase process: syntactic exploration to gather diverse outputs for constraint learning, followed by constraint exploitation to enforce these learned rules during generation. Experiments demonstrate that our method enables even small LLMs (1B parameters) to learn and generate with perfect constraint adherence, outperforming larger counterparts and state-of-the-art reasoning models. This work represents the first integration of context-sensitive grammar learning with LLM generation, eliminating manual specification while maintaining generation validity.

Large Language Models (LLMs) are typically trained to predict the next token in a sequence. However, their internal representations often encode signals that go beyond immediate next-token prediction. In this work, we investigate whether these hidden states also carry information about the remaining length of the generated output—an implicit form of foresight (CITATION). We formulate this as a regression problem where, at generation step t, the target is the number of remaining tokens y_t = T - t, with T as the total output length.We propose two approaches: (1) an aggregation-based model that combines hidden states from multiple transformer layers ℓ ∈ {8, \dots, 15} using element-wise operations such as mean or sum, and (2) a Layerwise Graph Regressor that treats layerwise hidden states as nodes in a fully connected graph and applies a Graph Neural Network (GNN) to predict y_t. Both models operate on frozen LLM embeddings without requiring end-to-end fine-tuning.Accurately estimating remaining output length has both theoretical and practical implications. From an interpretability standpoint, it suggests that LLMs internally track their generation progress. From a systems perspective, it enables optimizations such as output-length-aware scheduling (CITATION). Our graph-based model achieves state-of-the-art performance on the Alpaca dataset using LLaMA-3-8B-Instruct, reducing normalized mean absolute error (NMAE) by over 50% in short-output scenarios.

Most state-of-the-art large language models (LLMs) are trained mainly on English data, limiting their effectiveness on non-English, especially low-resource, languages. This study investigates whether language adapters can facilitate cross-lingual transfer in English-centric LLMs. We train language adapters for 13 languages using Llama 2 (7B) and Llama 3.1 (8B) as base models, and evaluate their effectiveness on two downstream tasks (MLQA and SIB-200) using either task adapters or in-context learning. Our results reveal that language adapters improve performance for languages not seen during pretraining, but provide negligible benefit for seen languages. These findings highlight the limitations of language adapters as a general solution for multilingual adaptation in English-centric LLMs.

Recent approaches to Hierarchical Text Classification (HTC) rely on capturing the global label hierarchy, which contains static and often redundant relationships. Instead, the hierarchical relationships within the instance-specific set of positive labels are more important, as they focus on the relevant parts of the hierarchy. These localized relationships can be modeled as a semantic alignment between the text and its positive labels within the embedding space. However, without explicitly encoding the global hierarchy, achieving this alignment directly in Euclidean space is challenging, as its flat geometry does not naturally support hierarchicalrelationships. To address this, we propose Hyperbolic Instance-Specific Local Relationships (HyILR), which models instance-specific relationships using the Lorentz model of hyperbolic space. Text and label features are projected into hyperbolic space, where a contrastive loss aligns text with its labels. This loss is guided by a hierarchy-aware negative sampling strategy, ensuring the selection of structurally and semantically relevant negatives. By leveraging hyperbolic geometry for this alignment, our approach inherently captures hierarchical relationships and eliminates the need for global hierarchy encoding. Experimental results on four benchmark datasets validate the superior performance of HyILR over baseline methods.

While large language models (LLMs) have demonstrated impressive capabilities across diverse tasks, their ability to generate valid graph structures remains underexplored. We evaluate fifteen state-of-the-art LLMs on five specialized graph generation tasks spanning delivery networks, social networks, quantum circuits, gene-disease networks, and transportation systems. We also test the LLMs using 3 different prompt types: direct, iterative feedback, and program-augmented. Models supported with explicit reasoning modules (o3-mini-high, o1, Claude 3.7 Sonnet, DeepSeek-R1) solve more than twice as many tasks as their general-purpose peers, independent of parameter count. Error forensics reveals two recurring failure modes: smaller parameter size Llama models often violate basic structural constraints, whereas Claude models respect topology but mismanage higher-order logical rules. Allowing models to refine their answers iteratively yields uneven gains, underscoring fundamental differences in error-correction capacity. This work demonstrates that graph competence stems from specialized training methodologies rather than scale, establishing a framework for developing truly graph-savvy language models. Results and verification scripts available at https://github.com/egedemirci/Are-LLMs-Truly-Graph-Savvy-A-Comprehensive-Evaluation-of-Graph-Generation.

Drawing on pragmatic theories of implicature by Grice (1975) and Levinson (1983), according to which speakers often convey more than it is explicitly said, the paper argues that interpreting texts with implicit meaning correctly is essential for precise natural language understanding. To illustrate the challenges in computational interpretation of implicatures, the study introduces a series of illustrative micro-experiments with the use of four transformer models fine-tuned for sentiment analysis. In these micro-experiments, the models classified sentences specifically designed to expose difficulties in handling implicit meaning. The study demonstrates that contrasting qualitative pragmatic analysis with the models’ tendency to focus on formal linguistic markers can reveal the limitations of supervised machine learning methods in detecting implicit sentiments.

In education, peer instruction (PI) is widely recognized as an effective active learning strategy. However, real-world evaluations of PI are often limited by logistical constraints and variability in classroom settings. This paper introduces PEERS (Peer Enhanced Educational Realistic Simulation), a simulation framework that integrates Agent-Based Modeling (ABM), Large Language Models (LLMs), and Bayesian Knowledge Tracing (BKT) to emulate student learning dynamics. As an initial step, this study focuses on evaluating whether LLM-powered agents can effectively assume the roles of teachers and students within the simulation. Human evaluations and topic-based metrics show that LLMs can generate role-consistent and contextually appropriate classroom dialogues. These results serve as a foundational milestone toward building realistic, AI-driven educational simulations. Future work will include simulating the complete PEERS framework and validating its accuracy through actual classroom-based PI sessions. This research aims to contribute a scalable, cost-effective methodology for studying instructional strategies in controlled yet realistic environments.

Integrating knowledge graphs (KGs) into the reasoning processes of large language models (LLMs) has emerged as a promising approach to mitigate hallucination. However, existing work in this area often relies on proprietary or extremely large models, limiting accessibility and scalability. In this study, we investigate the capabilities of existing integration methods for small language models (SLMs) in KG-based question answering and observe that their performance is often constrained by their limited ability to traverse and reason over knowledge graphs. To address this limitation, we propose leveraging simple and efficient exploration modules to handle knowledge graph traversal in place of the language model itself. Experiment results demonstrate that these lightweight modules effectively improve the performance of small language models on knowledge graph question answering tasks. Source code: https://github.com/yijie-cheng/SLM-ToG/.

This paper quantifies the “embodiment gap” between disembodied language models and embodied agricultural knowledge communication through mixed-methods analysis with 78 farmers. Our key contributions include: (1) the Embodied Knowledge Representation Framework (EKRF), a novel computational architecture with specialized lexical mapping that incorporates embodied linguistic patterns from five identified domains of agricultural expertise; (2) the Embodied Prompt Engineering Protocol (EPEP), which reduced the embodiment gap by 47.3% through systematic linguistic scaffolding techniques; and (3) the Embodied Knowledge Representation Index (EKRI), a new metric for evaluating embodied knowledge representation in language models. Implementation results show substantial improvements across agricultural domains, with particularly strong gains in tool usage discourse (58.7%) and soil assessment terminology (67% reduction in embodiment gap). This research advances both theoretical understanding of embodied cognition in AI and practical methodologies to enhance LLM performance in domains requiring embodied expertise.

To evaluate the creativity of large language models (LLMs) in Japanese, we construct three benchmarks: Japanese Creativity Questions (JCQ), Divergent Association Task (DAT), and Story Alteration Task (SAT). JCQ comprehensively evaluates creativity using LLMs. Meanwhile, DAT and SAT measure specific aspects of creative ability using embeddings. We also analyze correlations between JCQ and DAT, JCQ and SAT, and DAT and SAT. While JCQ provides comprehensive evaluation, it is relatively time and resource intensive. In contrast, DAT and SAT offer lower comprehensiveness but enable quick, low-cost assessment. Additionally, we investigate whether training with DAT contributes to enhancing LLM creativity.

Sentiment analysis in policy-related studies typically involves annotating a subset of data to fine-tune a pre-trained model, which is subsequently used to classify sentiments in the remaining unlabeled texts, enabling policy researchers to analyze sentiments in novel policy contexts under resource constraints. We argue that existing methods fail to adequately capture the temporal volatility inherent in policy-related sentiments, which are subject to external shocks and evolving discourse of opinions. We propose methods accounting for the temporal dynamics of policy-related texts. Specifically, we propose leveraging continuous time-series clustering to select data points for annotation based on temporal trends and subsequently apply model merging techniques - each fine-tuned separately on data from distinct time intervals. Our results indicate that continuous time-series clustering followed by fine-tuning a single unified model achieves superior performance, outperforming existing methods by an average F1-score of 2.71%. This suggests that language models can generalize to temporally sensitive texts when provided with temporally representative samples. Nevertheless, merging multiple time-specific models - particularly via greedy soup and TIES - achieves competitive performance, suggesting practical applications in dynamically evolving policy scenarios.

Discourse relations are sometimes explicitly conveyed by specific connectives.However, some connectives can signal multiple discourse relations; in such cases, disambiguation is necessary to determine which relation is intended.This task is known as *discourse connective disambiguation* (Pitler and Nenkova, 2009), and particular attention is often given to connectives that can convey both *concession* and other relations (e.g., *synchronous*).In this study, we conducted experiments to analyze which linguistic features play an important role in the disambiguation of polysemous connectives in Japanese.A neural language model (BERT) was fine-tuned using inputs from which specific linguistic features (e.g., word order, specific lexicon, etc.) had been removed.We analyzed which linguistic features affect disambiguation by comparing the model’s performance.Our results show that even after performing drastic removal, such as deleting one of the two arguments that constitute the discourse relation, the model’s performance remained relatively robust.However, the removal of certain lexical items or words belonging to specific lexical categories significantly degraded disambiguation performance, highlighting their importance in identifying the intended discourse relation.

Recent studies on semantic frame induction have demonstrated that the emergence of pre-trained language models (PLMs) has led to more accurate results.However, most existing studies evaluate the performance using frame resources such as FrameNet, which may not accurately reflect real-world language usage.In this study, we conduct semantic frame induction using the Colossal Clean Crawled Corpus (C4) and assess the applicability of existing frame induction methods to real-world data.Our experimental results demonstrate that existing frame induction methods are effective on real-world data and that frames corresponding to novel concepts can be induced.

Morphological defectivity is an intriguing and understudied phenomenon in linguistics. Addressing defectivity, where expected inflectional forms are absent, is essential for improving the accuracy of NLP tools in morphologically rich languages. However, traditional linguistic resources often lack coverage of morphological gaps as such knowledge requires significant human expertise and effort to document and verify. For scarce linguistic phenomena in under-explored languages, Wikipedia and Wiktionary often serve as among the few accessible resources. Despite their extensive reach, their reliability has been a subject of controversy. This study customizes a novel neural morphological analyzer to annotate Latin and Italian corpora. Using the massive annotated data, crowd-sourced lists of defective verbs compiled from Wiktionary are validated computationally. Our results indicate that while Wiktionary provides a highly reliable account of Italian morphological gaps, 7% of Latin lemmata listed as defective show strong corpus evidence of being non-defective. This discrepancy highlights potential limitations of crowd-sourced wikis as definitive sources of linguistic knowledge, particularly for less-studied phenomena and languages, despite their value as resources for rare linguistic features. By providing scalable tools and methods for quality assurance of crowd-sourced data, this work advances computational morphology and expands linguistic knowledge of defectivity in non-English, morphologically rich languages.

Various evaluation metrics have been proposed for Grammatical Error Correction (GEC), but many, particularly reference-free metrics, lack explainability. This lack of explainability hinders researchers from analyzing the strengths and weaknesses of GEC models and limits the ability to provide detailed feedback for users. To address this issue, we propose attributing sentence-level scores to individual edits, providing insight into how specific corrections contribute to the overall performance. For the attribution method, we use Shapley values, from cooperative game theory, to compute the contribution of each edit. Experiments with existing sentence-level metrics demonstrate high consistency across different edit granularities and show approximately 70% alignment with human evaluations. In addition, we analyze biases in the metrics based on the attribution results, revealing trends such as the tendency to ignore orthographic edits.

Variation in human annotation and human perspectives has drawn increasing attention in natural language processing research. Disagreement observed in data annotation challenges the conventional assumption of a single “ground truth” and uniform models trained on aggregated annotations, which tend to overlook minority viewpoints and individual perspectives. This proposal investigates three directions of perspective-oriented research: First, annotation formats that better capture the granularity and uncertainty of individual judgments; Second, annotation modeling that leverages socio-demographic features to better represent and predict underrepresented or minority perspectives; Third, personalized text generation that tailors outputs to individual users’ preferences and communicative styles. The proposed tasks aim to advance natural language processing research towards more faithfully reflecting the diversity of human interpretation, enhancing both inclusiveness and fairness in language technologies.

Large language models often suffer from language confusion, a phenomenon in which responses are partially or entirely generated in unintended languages. This critically degrades the user experience, especially in low-resource settings. We hypothesize that this issue stems from limitations in conventional fine-tuning objectives, such as supervised learning, which optimize the likelihood of correct tokens without explicitly penalizing undesired outputs such as cross-lingual mixing. Analysis of loss trajectories during pretraining further reveals that models fail to distinguish between monolingual and language-mixed texts, highlighting the absence of inherent pressure to avoid such confusion. In this work, we apply ORPO, which adds penalties for unwanted output styles to standard SFT, effectively suppressing language-confused generations. ORPO maintains strong language consistency, even under high decoding temperatures, while preserving general QA performance. Our findings suggest that incorporating appropriate penalty terms can effectively mitigate language confusion in multilingual models, particularly in low-resource scenarios.

We introduce a modified sequence tagging architecture, proposed in (Omelianchuk et al., 2020), for the Grammatical Error Correction of the Russian language. We propose language-specific operation set and preprocessing algorithm as well as a classification scheme which makes distinct predictions for insertions and other operations. The best versions of our models outperform previous approaches and set new SOTA on the two Russian GEC benchmarks – RU-Lang8 and GERA, while achieve competitive performance on RULEC-GEC.

Recently, large language models (LLMs) have demonstrated impressive capabilities in dealing with new tasks with the help of in-context learning (ICL). In the study of Large Vision-Language Models (LVLMs), when implementing ICL, researchers usually adopt the naive strategies like fixed demonstrations across different samples, or selecting demonstrations directly via a visual-language embedding model. These methods do not guarantee the configured demonstrations fit the need of the LVLMs. To address this issue, we propose a novel framework, demonstration retriever for large multi-modal model (DRUM), which fine-tunes the CLIP embedding model to better meet the LVLM’s needs. First, we discuss the retrieval strategies for a visual-language task, assuming an embedding model is given. And we propose to concate the image and text embeddings to enhance the retrieval performance. Second, we propose to re-rank the the embedding model’s retrieved demonstrations via the LVLM’s feedbacks, and calculate a list-wise ranking loss for training the embedding model. Third, we propose an iterative demonstration mining strategy to improve the training of the embedding model. Through extensive experiments on 3 types of visual-language tasks, 7 benchmark datasets, our DRUM framework is proven to be effective in boosting the LVLM’s in-context learning performance via retrieving more proper demonstrations.

Due to strict privacy regulations, text corpora in non-English clinical contexts are scarce. Consequently, synthetic data generation using Large Language Models (LLMs) emerges as a promising strategy to address this data gap. To evaluate the ability of LLMs in generating synthetic data, we applied them to our novel German Medical Interview Questions Corpus (GerMedIQ), which consists of 4,524 unique, simulated question-response pairs in German. We augmented our corpus by prompting 18 different LLMs to generate responses to the same questions. Structural and semantic evaluations of the generated responses revealed that large-sized language models produced responses comparable to those provided by humans. Additionally, an LLM-as-a-judge study, combined with a human baseline experiment assessing response acceptability, demonstrated that human raters preferred the responses generated by Mistral (124B) over those produced by humans. Nonetheless, our findings indicate that using LLMs for data augmentation in non-English clinical contexts requires caution.

This paper examines differences between stream-of-consciousness (SoC) narratives written by humans and those generated by large language models (LLMs) to assess narrative coherence and personality expression. We generated texts by prompting LLMs (Llama-3.1-8B & DeepSeek-R1-Distill-Llama-8B) with the first half of SoC-essays while either providing the models with the personality characteristics (Big Five) or omitting them. Our analysis revealed consistently low similarity between LLM-generated continuations and original human texts, as measured by cosine similarity, perplexity, and BLEU scores. Including explicit personality traits significantly enhanced Llama-3.1-8B’s performance, particularly in BLEU scores.Further analysis of personality expression showed varying alignment patterns between LLMs and human texts. Specifically, Llama-3.1-8B exhibited higher extraversion but low agreeableness, while DeepSeek-R1-Distill-Llama-8B displayed dramatic personality shifts during its reasoning process, especially when prompted with personality traits, with all models consistently showing very low Openness.

This study investigates the use of Large Language Models (LLMs) for political stance detection in informal online discourse, where language is often sarcastic, ambiguous, and context-dependent. We explore whether providing contextual information, specifically user profile summaries derived from historical posts, can improve classification accuracy. Using a real-world political forum dataset, we generate structured profiles that summarize users’ ideological leaning, recurring topics, and linguistic patterns. We evaluate seven state-of-the-art LLMs across baseline and context-enriched setups through a comprehensive cross-model evaluation. Our findings show that contextual prompts significantly boost accuracy, with improvements ranging from +17.5% to +38.5%, achieving up to 74% accuracy that surpasses previous approaches. We also analyze how profile size and post selection strategies affect performance, showing that strategically chosen political content yields better results than larger, randomly selected contexts. These findings underscore the value of incorporating user-level context to enhance LLM performance in nuanced political classification tasks.

Health literacy plays a critical role in ensuring people can access, understand, and act on medical information. However, much of the health content available today is too complex for many people, and simplifying these texts manually is time-consuming and difficult to do at scale.To overcome this, we developed a new framework for automatically generating health answers at multiple, precisely controlled complexity levels.We began with a thorough analysis of 166 linguistic features, which we then refined into 13 key metrics that reliably differentiate between simple and complex medical texts. From these metrics, we derived a robust complexity scoring formula, combining them with weights learned from a logistic regression model. This formula allowed us to create a large, multi-level dataset of health question-answer pairs covering 21 distinct complexity levels, ranging from elementary patient-friendly explanations to highly technical summaries.Finally, we fine-tuned a Llama-3.1-8B-Instruct model using “control codes” on this dataset, giving users precise control over the complexity of the generated text and empowering them to select the level of detail and technicality they need.

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP), but their success remains largely confined to high-resource, general-purpose domains. In contrast, applying LLMs to low-resource domains poses significant challenges due to limited training data, domain drift, and strict terminology constraints. This survey provides an overview of the current landscape in domain-specific, low-resource QA with LLMs. We begin by analyzing the coverage and representativeness of specialized-domain QA datasets against large-scale reference datasets what we refer to as ParentQA. Building on this analysis, we survey data-centric strategies to enhance input diversity, including data augmentation techniques. We further discuss evaluation metrics for specialized tasks and consider ethical concerns. By mapping current methodologies and outlining open research questions, this survey aims to guide future efforts in adapting LLMs for robust and responsible use in resource-constrained, domain-specific environments. To facilitate reproducibility, we make our code available at https://github.com/kentrachmat/survey-da.

Time series modeling holds significant importance in many industrial applications and has been extensively studied. A series of recent studies have demonstrated that large language models (LLMs) possess robust pattern recognition and semantic understanding capabilities over time series data. However, the current literature have yet striked a high-quality balance between (a) effectively aligning the time series and natural language modalities and (b) keeping the inference efficiency for industrial deployment. To address the above issues, we now propose the Time-LlaMA framework. Time-LlaMA first converts the time series input into token embeddings through a linear tokenization mechanism. Second, the time series token embeddings are aligned with the text prompts. Third, to further adapt the large languag model (LLM) backbone for time series modeling, we have developed a dynamic low-rank adaptation technique (DynaLoRA). DynaLoRA dynamically chooses the most suitable LoRA modules at each layer of the Transformer backbone for each time series input, enhancing the model’s predictive capabilities. Our experimental results on an extensive collection of challenging open and proprietary time series tasks confirm that our proposed method achieves the state-of-the-art (SOTA) performance and have potentials for wide industrial usages.

This paper represents an implementation of an approach rather similar to that of (Zhu et al., 2024), adapted for the Russian-language data. We introduce the RusConText Benchmark for evaluating short-context understanding in Russian, comprising four distinct yet interrelated tasks: coreference resolution, discourse understanding, idiom interpretation and ellipsis resolution. Each task targets a specific aspect of linguistic processing, challenging a large language model to recover omitted information, resolve referential dependencies, interpret idioms and discourse. The RusConText Benchmark is an additional resource beyond standard benchmarks, designed to assess model performance from a specific perspective. In addition, we present the results of scoring 4 models on our benchmark.

With Large Language Model (LLM)-based applications becoming more common due to strong performance across many tasks, prompt optimization has emerged as a way to extract better solutions from frozen, often commercial LLMs that are not specifically adapted to a task. LLM-assisted prompt optimization methods provide a promising alternative to manual/human prompt engineering, where LLM “reasoning” can be used to make them optimizing agents. However, the cost of using LLMs for prompt optimization via commercial APIs remains high, especially for heuristic methods like evolutionary algorithms (EAs), which need many iterations to converge, and thus, tokens, API calls, and rate-limited network overhead. We propose GenDLN, an open-source, efficient genetic algorithm-based prompt pair optimization framework that leverages commercial API free tiers. Our approach allows teams with limited resources (NGOs, non-profits, academics, ...) to efficiently use commercial LLMs for EA-based prompt optimization. We conduct experiments on CLAUDETTE for legal terms of service classification and MRPC for paraphrase detection, performing in line with selected prompt optimization baselines, at no cost.

Sign language segmentation focuses on identifying temporal boundaries within sign language videos. As compared to previous segmentation techniques that have depended on frame-level and phrase-level segmentation, our study emphasizes on subtitle-level segmentation, using synchronized subtitle data to facilitate temporal boundary recognition. Based on Beginning-Inside-Outside (BIO) tagging for subtitle unit delineation, we train a sequence-to-sequence (Seq2Seq) model with and without attention for subtitle boundary identification. Training on optical flow data and aligned subtitles from BOBSL and YouTube-ASL, we show that the Seq2Seq model with attention outperforms baseline models, achieving improved percentage of segments, F1 and IoU score. An additional contribution is the development of an method for subtitle temporal resolution, aiming to facilitate manual annotation.

Large Language Models (LLMs) can enhance the performance of Named Entity Recognition (NER) tasks by leveraging external knowledge through in-context learning. When it comes to entity-type-related external knowledge, existing methods mainly provide LLMs with semantic information such as the definition and annotation guidelines of an entity type, leaving the effect of orthographic or morphological information on LLM-based NER unexplored. Besides, it is non-trivial to obtain literal patterns written in natural language to serve LLMs. In this work, we propose LiP-NER, an LLM-based NER framework that utilizes Literal Patterns, the entity-type-related knowledge that directly describes the orthographic and morphological features of entities. We also propose an LLM-based method to automatically acquire literal patterns, which requires only several sample entities rather than any annotation example, thus further reducing human labor. Our extensive experiments suggest that literal patterns can enhance the performance of LLMs in NER tasks. In further analysis, we found that entity types with relatively standardized naming conventions but limited world knowledge in LLMs, as well as entity types with broad and ambiguous names or definitions yet low internal variation among entities, benefit most from our approach. We found that the most effective written literal patterns are (1) detailed in classification, (2) focused on majority cases rather than minorities, and (3) explicit about obvious literal features.

It is widely assumed that Large Language Models (LLMs) are shaping language, with multiple studies noting the growing presence of LLM-generated content and suggesting homogenizing effects. However, it remains unclear if these effects are already evident in recent writing. This study addresses that gap by comparing two datasets of English online news articles – one from 2018, prior to LLM popularization, and one from 2024, after widespread LLM adoption. We define lexical homogenization as a decrease in lexical diversity, measured by the MATTR, Maas, and MTLD metrics, and introduce the LLM-Style-Word Ratio (SWR) to measure LLM influence. We found higher MTLD and SWR scores, yet negligible changes in Maas and MATTR scores in 2024 corpus. We conclude that while there is an apparent influence of LLMs on written online English, homogenization effects do not show in the measurements. We therefore propose to apply different metrics to measure lexical homogenization in future studies on the influence of LLM usage on language change.

Static and outdated datasets hinder the accuracy of Financial Sentiment Analysis (FSA) in capturing rapidly evolving market sentiment. We tackle this by proposing a novel data augmentation technique using Retrieval Augmented Generation (RAG). Our method leverages a generative LLM to infuse established benchmarks with up-to-date contextual information from contemporary financial news. This RAG-based augmentation significantly modernizes the data’s alignment with current financial language. Furthermore, a robust BERT-BiGRU judge model verifies that the sentiment of the original annotations is faithfully preserved, ensuring the generation of high-quality, temporally relevant, and sentiment-consistent data suitable for advancing FSA model development.