Representation Fine-tuning (ReFT), a recently proposed parameter-efficient fine-tuning (PeFT) method, significantly improves parameter efficiency by modifying the representation space alone. However, directly applying ReFT, which alters a fixed number of representations at the beginning and end positions of each layer, results in suboptimal performance for two reasons. (i) The impact of these fixed-position representations on the output is uncertain; (ii) As the sequence length increases, fine-tuning a fixed number of representations may have diminishing effects on the final results. Based on our observations that punctuation plays a crucial role in integrating representations from preceding layers and modulating those of subsequent layers, we introduce Punctuation-steered Representation Fine-tuning (PuReFT), a straightforward yet powerful approach that additionally fine-tunes punctuation representations to achieve performance improvements. Extensive evaluations on common-sense, arithmetic, and code datasets demonstrate the effectiveness and versatility of PuReFT. Furthermore, our analysis of its training speed and memory overhead confirms its greater ease of use and efficiency.

While self-consistency is known to improve performance on symbolic reasoning, its effect on the recall of encyclopedic knowledge is unclear due to a lack of targeted evaluation grounds.To address this, we establish such a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work.We validate this split by showing that the performance patterns on the symbolic reasoning and knowledge recall subsets mirror those of GSM8K and MedMCQA, respectively.Using this solid ground, we find that self-consistency consistently improves performance across both symbolic reasoning and knowledge recall, even though its underlying CoT prompting is primarily effective for symbolic reasoning.As a result, we achieve an 89% accuracy on MMLU, the best performance to date with the use of GPT-4o.

Commonsense reasoning often requires both textual and visual knowledge, yet Large Language Models (LLMs) trained solely on text lack visual grounding (e.g., "what color is an emperor penguin’s belly?"). Visual Language Models (VLMs) perform better on visually grounded tasks but face two limitations: (i) often reduced performance on text-only commonsense reasoning compared to text-trained LLMs, and (ii) adapting newly released LLMs to vision input typically requires costly multimodal training. An alternative augments LLMs with test-time visual signals, improving visual commonsense without harming textual reasoning, but prior designs often rely on early fusion and a single image, which can be suboptimal. We propose a late multi-image fusion method: multiple images are generated from the text prompt with a lightweight parallel sampling, and their prediction probabilities are combined with those of a text-only LLM through a late-fusion layer that integrates projected visual features just before the final prediction. Across visual commonsense and NLP benchmarks, our method significantly outperforms augmented LLMs on visual reasoning, matches VLMs on vision-based tasks, and, when applied to strong LLMs such as LLaMA 3, also improves NLP performance while adding only modest test-time overhead.

While full-duplex speech agents enable natural, low-latency interaction by speaking and listening simultaneously, their consistency and task performance in multi-turn settings remain underexplored. We introduce Full-Duplex-Bench-v2 (FDB-v2), a streaming framework that integrates with an automated examiner that enforces staged goals under two pacing setups (Fast vs. Slow). FDB-v2 covers four task families—Daily, Correction, Entity Tracking, and Safety—and reports turn-taking fluency, multi-turn instruction following, and task-specific competence. The framework is extensible, supporting both commercial APIs and open-source models. When we test full-duplex systems with FDB-v2, they often get confused when people talk at the same time, struggle to handle corrections smoothly, and sometimes lose track of who or what is being talked about. Through an open-source, standardized streaming protocol and a task set, FDB-v2 makes it easy to extend to new task families, allowing the community to tailor and accelerate evaluation of multi-turn full-duplex systems.

Retrieval-augmented generation (RAG) over long documents typically involves splitting the text into smaller chunks, which serve as the basic units for retrieval. However, due to dependencies across the original document, contextual information is often essential for accurately interpreting each chunk. To address this, prior work has explored encoding longer context windows to produce embeddings for longer chunks, yet their gains in retrieval and downstream tasks remain limited. This is because (1) longer chunks strain the capacity of embedding models due to the increased amount of information they must encode, and (2) many real-world applications still require returning localized evidence due to constraints on model or human bandwidth. To this end, we propose an alternative approach to this challenge by representing short chunks in a way that is conditioned on a broader context window to enhance retrieval performance – i.e., situating a chunk’s meaning within its context. We further show that existing embedding models are not well-equipped to encode such situated context effectively, and thus introduce a new training paradigm and develop the first situated embedding model. To evaluate our method, we curate a book-plot retrieval dataset specifically designed to assess situated retrieval capabilities. On this benchmark, our 1B-parameter model substantially outperforms state-of-the-art embedding models, including several with up to 7B parameters.

The world is in the grip of ecological, meaning, and language crises that are converging into a metacrisis. Big AI is accelerating them all. LLM engineering sits at the core. Despite the public good motives of language engineers and the promise of LLMs, this work is being leveraged to create unprecedented wealth and power for a handful of individuals and corporations while causing existential harm to life on earth. As a profession, we urgently need to come together to explore alternatives and to design a life-affirming future for our field of natural language processing that is centered on human flourishing on a living planet.

Current research on Event Factuality Prediction (EFP) predominantly treats LLMs as passive classifiers, where high aggregate metrics often mask shortcut learning and unreliable reasoning. In this position paper, we argue for a focus shift from event factuality to meta-factivity. We introduce the Meta-Factivity Framework (MFF), a theoretical roadmap that moves evaluation beyond surface recognition to belief trajectory reasoning and epistemic regulation. By framing hallucination as a failure of meta-cognitive control, we advocate for a transition from measuring black-box accuracy to evaluating white-box cognition, laying the groundwork for a more rigorous benchmark for explainable self-governance.

Transformers often display an *attention sink*: probability mass concentrates on a fixed, content-agnostic position. Are sinks a byproduct of the optimization/training regime? Or are they sometimes functionally necessary in softmax Transformers? We prove that, in some settings, it is the latter: computing a simple trigger-conditional behavior *necessarily* induces a sink in softmax self-attention models. Our results formalize a familiar intuition: normalization over a probability simplex must force attention to collapse onto a stable anchor to realize a default state (e.g., when the model needs to ignore the input). We instantiate this with a concrete task: when a designated trigger token appears, the model must return the *average of all preceding token representations*, and otherwise output zero, a task which mirrors the functionality of attention heads in the wild (Barbero et al., 2025; Guo et al., 2024). We also prove that non-normalized ReLU attention can solve the same task without any sink, confirming that the normalization constraint is the fundamental driver of sink behavior. Experiments validate our predictions and demonstrate they extend beyond the theoretically analyzed setting: softmax models develop strong sinks while ReLU attention eliminates them in both single-head and multi-head variants.

Transformers commonly exhibit an attention sink: disproportionately high attention to the first position. We study this behavior in GPT-2–style models with learned query biases and absolute positional embeddings. Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection. Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks. This indicates that attention sinks may arise through distinct circuits across architectures. These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerges.

Classifier-based Quality Filtering has recently emerged as a fundamental technique in constructing pre-training corpora. The ability to deploy a single model that can replace or supplement a set of heuristics has proven effective across numerous Large Language Models. In this work, we expose a critical vulnerability in this approach by demonstrating how a straightforward Wikipedia-style reformatting operation can substantially alter a model’s quality assessment and enable low-quality content to surpass filtering thresholds. Our analysis reveals that the FineWeb-Edu CQF model would reverse its filtering decision for approximately 7% of evaluated documents, thereby admitting content into the pre-training corpus that would otherwise have been excluded.

Group-based reinforcement learning methods, like Group Relative Policy Optimization (GRPO), are widely used nowadays to post-train large language models. Despite their empirical success, they exhibit structural mismatches between reward optimization and the underlying training objective. In this paper, we present a theoretical analysis of GRPO style methods by studying them within a unified surrogate formulation. This perspective reveals recurring properties that affect all the methods under analysis: (i) non-uniform group weighting induces systematic gradient biases on shared prefix tokens; (ii) interactions with the AdamW optimizer make training dynamics largely insensitive to reward scaling; and (iii) optimizer momentum can push policy updates beyond the intended clipping region under repeated optimization steps. We believe that these findings highlight fundamental limitations of current approaches and provide principled guidance for the design of future formulations.

Long-term conversational agents need memory systems that capture relationships between events, not merely isolated facts, to support temporal reasoning and multi-hop question answering. Current approaches face a fundamental trade-off: flat memory is efficient but fails to model relational structure, while graph-based memory enables structured reasoning at the cost of expensive and fragile construction. To address these issues, we propose StructMem, a structure-enriched hierarchical memory framework that preserves event-level bindings and induces cross-event connections. By temporally anchoring dual perspectives and performing periodic semantic consolidation, StructMem improves temporal reasoning and multi-hop performance on LoCoMo, while substantially reducing token usage, API calls, and runtime compared to prior memory systems.

3D visual grounding (3DVG) aims to localize objects in a 3D scene based on natural language queries. In this work, we explore zero-shot 3DVG from multi-view images alone, without requiring any geometric supervision or object priors. We introduce Z3D, a universal grounding pipeline that flexibly operates on multi-view images while optionally incorporating camera poses and depth maps. We identify key bottlenecks in prior zero-shot methods causing significant performance degradation and address them with (i) a state-of-the-art zero-shot 3D instance segmentation method to generate high-quality 3D bounding box proposals and (ii) advanced reasoning via prompt-based segmentation, which utilizes full capabilities of modern VLMs. Extensive experiments on the ScanRefer and Nr3D benchmarks demonstrate that our approach achieves state-of-the-art performance among zero-shot methods.

Retrieval-Augmented Generation (RAG) improves the factual accuracy of large language model (LLM) outputs by grounding generation in external knowledge. Recent agentic RAG systems extend this paradigm by introducing critic agents that evaluate model responses and iteratively refine outputs. However, most prior work implicitly assumes reliable critic feedback and focuses on planning strategies, while paying limited attention to the robustness of the error correction process itself, which is often hindered by misaligned error categories and ineffective or incorrect corrections. We hypothesize that RAG performance can be improved without explicit error categorization. To this end, we propose RePAIR, a response–action learning paradigm that directly maps flawed RAG outputs to error-mitigating action plans without relying on fine-grained error taxonomies or explicit critic supervision. Across multiple benchmarks, RePAIR consistently improves agentic RAG performance.

Agentic LLM inference with long contexts is increasingly limited by memory bandwidth rather than compute. In this setting, SwiGLU MLP blocks, whose large weights exceed cache capacity, become a major yet under-optimized bottleneck in the Transformer architecture. We propose DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, delivering up to 13.2% speedup on H100 and 9.7% on A100 over SGLang. Integrated with SGLang and paired with a kernel scheduler, DeepFusionKernel ensures consistent accelerations across generation lengths, while remaining adaptable to diverse models, inference configurations, and hardware platforms.

Anchoring bias causes large language models (LLMs) to shift quantitative judgments in response to irrelevant numerical primes. We analyze this bias as a function of model confidence and accuracy in base, instruction-tuned, and distilled variants of Llama and Qwen models. We find that anchoring susceptibility is negatively correlated with model confidence without regard to accuracy: confidently incorrect models resist anchoring as effectively as accurate ones, provided their internal priors are sufficiently strong. We further show that post-training impacts the strength of this relationship, and that models are more susceptible to high anchors than to low anchors. Our findings suggest anchoring resistance is a structural property of distributional concentration (certainty) rather than knowledge correctness (factual accuracy), with implications for deploying LLMs in numerical reasoning tasks.

Relation extraction represents a fundamental component in the process of creating knowledge graphs, among other applications. Large language models (LLMs) have been adopted as a promising tool for relation extraction, both in supervised and in-context learning settings. However, in this work we show that their performance still lags behind much smaller architectures when the linguistic graph underlying a text has great complexity. To demonstrate this, we evaluate four LLMs against a graph-based parser on six relation extraction datasets with sentence graphs of varying sizes and complexities. Our results show that the graph-based parser increasingly outperforms the LLMs, as the number of relations in the input documents increases. This makes the much lighter graph-based parser a superior choice in the presence of complex linguistic graphs.

One of biggest missing capabilities in state-of-the-art AI systems is the ability to learn continually after deployment. However, implementing an inference-time learning system has several challenges including the large memory requirement of gradient-based algorithms that are used to train state-of-the-art LLMs. Evolutionary Strategies (ES) have recently re-emerged as a gradient-free alternative to traditional learning algorithms and have shown encouraging performance on specific tasks in LLMs. In this paper, we perform a more comprehensive analysis of ES and specifically evaluate its forgetting curves when training for a larger number of update steps. We find that although ES is able to reach performance numbers closer to GRPO for math and reasoning tasks, it is accompanied by significant forgetting of prior abilities. We also show that the updates made using ES are much less sparse and have a larger l2 norm compared to corresponding GRPO updates, explaining the contrasting forgetting curves between the two algorithms. With this study, we aim to specifically highlight the issue of forgetting in gradient-free algorithms like ES and hope to inspire future work to mitigate these issues.

Reinforcement Learning (RL) enhances LLM reasoning, yet a paradox emerges as models scale: strong base models saturate standard benchmarks (e.g., MATH), yielding correct but homogeneous solutions. In such environments, the lack of failure cases causes the advantage signal in group-relative algorithms (e.g., GRPO) to vanish, driving policies into mode collapse. To address this, we propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration. Unlike standard sampling that follows model biases, CUTS flattens the local optimization landscape by sampling uniformly from constrained high-confidence candidates. We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance. Experiments on Qwen3 models demonstrate that our approach prevents policy degeneration and significantly boosts out-of-domain generalization. Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO, validating that maintaining diversity within the semantic manifold is critical for rigorous reasoning.

Recent large language models (LLMs) perform strongly on mathematical benchmarks yet often misapply lemmas, importing conclusions without validating assumptions. We formalize lemma-judging as a structured prediction task: given a statement and a candidate lemma, the model must output a precondition check and a conclusion-utility check, from which a usefulness decision is derived. We present RULES, which encodes this specification via a two-section output and trains with reinforcement learning plus section-aware loss masking to assign penalty to the section responsible for errors. Training and evaluation draw on diverse natural-language and formal proof corpora; robustness is assessed with a held-out perturbation suite; and end-to-end evaluation spans competition-style, perturbation-aligned, and theorem-based problems across various LLMs. Results show consistent in-domain gains over both a vanilla model and a single-label RL baseline, larger improvements on applicability-breaking perturbations, and parity or modest gains on end-to-end tasks; ablations indicate that the two-section outputs and section-aware reinforcement are both necessary for robustness.

Analog in-memory computing (AIMC) offers substantial efficiency gains for transformer inference but introduces hardware-induced noise that can distort attention behavior. Prior studies primarily focus on AIMC evaluations for vision tasks and CNN-based models. They largely overlook how hardware-induced noise perturbs internal attention dynamics in NLP models. In this work, we present the first fine-grained analysis of analog vulnerability in pretrained transformers, examining projection submodules, attention heads, and layer-wise dynamics across multiple NLP tasks. Results show that query (Q), key (K), and value (V) projections are the most sensitive components, while feed-forward layers remain comparatively robust. Also, analog noise yields depth-dependent degradation in higher layers, leading to scattered attention and disrupted token routing. This pre-deployment analysis mitigates potential resource misuse before physical deployment and offers practical guidance for designing noise-resilient analog NLP transformers.

Evaluation of modern large language model (LLM) systems increasingly relies on multi-stage pipelines such as retrieval-augmented generation, tool-using agents, and prompt chains. Reproducing reported evaluation results for these systems often requires evaluation-specific artifacts beyond model weights and datasets, including prompts, judge configurations, retrieval snapshots, and intermediate traces, yet their availability has not been systematically examined.We introduce ReproEvalCard, a lightweight reporting standard that specifies the minimal artifacts required to reproduce and validate evaluations of LLM pipelines. To motivate this standard, we audit 55 pipeline-based LLM papers published between 2022 and 2025 and quantify the availability of reproducibility-critical evaluation artifacts. We find that randomness controls are missing in 75% of papers and intermediate execution traces in 61%, substantially limiting evaluation reproducibility. We further demonstrate ReproEvalCard through a worked example and provide a concise checklist for authors and reviewers, aiming to improve reproducibility and comparability in LLM evaluation.

Large language models (LLMs) now give reasoning before answering, excelling in tasks like multiple-choice question answering (MCQA). Yet, a concern is that LLMs do not solve MCQs as intended, as work finds LLMs sans reasoning succeed in MCQA without using the question, i.e., choices-only. Such partial-input success is often deemed problematic, but reasoning traces could reveal if these strategies are truly shallow in choices-only settings. To study these strategies, reasoning LLMs solve MCQs in full and choices-only inputs; test-time reasoning often boosts accuracy on full and in choices-only half the time. While possibly due to shallow shortcuts, choices-only success is barely affected by the length of reasoning traces, and after finding traces pass faithfulness tests, we show they use less problematic strategies like inferring missing questions. In all, we challenge claims that partial-input success is always a flaw, so we discuss how reasoning traces could separate problematic data from less problematic reasoning.

Automated 3D radiology report generation often suffers from clinical hallucinations and a lack of the iterative verification found in human practice. While recent Vision-Language Models (VLMs) have advanced the field, they typically operate as monolithic "black-box" systems without the collaborative oversight characteristic of clinical workflows. To address these challenges, we propose MARCH (Multi-Agent Radiology Clinical Hierarchy), a multi-agent framework that emulates the professional hierarchy of radiology departments and assigns specialized roles to distinct agents. MARCH utilizes a Resident Agent for initial drafting with multi-scale CT feature extraction, multiple Fellow Agents for retrieval-augmented revision, and an Attending Agent that orchestrates an iterative, stance-based consensus discourse to resolve diagnostic discrepancies. On the RadGenome-ChestCT dataset, MARCH significantly outperforms state-of-the-art baselines in both clinical fidelity and linguistic accuracy. Our work demonstrates that modeling human-like organizational structures enhances the reliability of AI in high-stakes medical domains.

Prefix parsing asks whether an input prefix can be extended to a complete string generated by a given grammar. In the weighted setting, it also provides prefix probabilities, which are central to context-free language modeling, psycholinguistic analysis, and syntactically constrained generation from large language models. We introduce the prefix grammar transformation, an efficient reduction of prefix parsing to ordinary parsing. Given a grammar, our method constructs another grammar that generates exactly the prefixes of its original strings. Prefix parsing is then solved by applying any ordinary parsing algorithm on the transformed grammar without modification. The reduction is both elegant and practical: the transformed grammar is only a small factor larger than the input, and any optimized implementation can be used directly, eliminating the need for bespoke prefix-parsing algorithms. We also present a strategy—based on algorithmic differentiation—for computing the next-token weight vector, i.e., the prefix weights of all one-token extensions, enabling efficient prediction of the next token. Together, these contributions yield a simple, general, and efficient framework for prefix parsing.

Speech representations that capture prosodic information can be useful for both understanding and generation. However, speaker characteristics are reflected in acoustic-prosodic features (e.g., pitch). To address privacy concerns from the leakage of identity information, we propose a new self-supervised approach to learning prosody representations that incorporates speaker disentanglement strategies. We evaluate our encoder on three tasks to probe representation capabilities, including pitch reconstruction and detection of different prosodic events. Our encoder outperforms raw prosody and HuBERT-base baselines, achieving strong speaker disentanglement without adverse impact on prosody-related downstream tasks.

Language models (LMs) are pre-trained on raw text datasets to generate text sequences token-by-token. While this approach facilitates the learning of world knowledge and reasoning, it does not explicitly optimize for linguistic competence. To bridge this gap, we propose L2T, a pre-training framework integrating Language Learning Tasks alongside standard next-token prediction. Inspired by human language acquisition, L2T transforms raw text into structured input-output pairs to provide explicit linguistic stimulation. Pre-training LMs on a mixture of raw text and L2T data not only improves overall performance on linguistic competence benchmarks but accelerates their acquisition, while maintaining competitive performance on general reasoning tasks.

Data mixing strategy is essential for large language model (LLM) training. Empirical evidence shows that inappropriate strategies can significantly reduce generalization. Although recent methods have improved empirical performance, several fundamental questions remain open: what constitutes a domain, whether human and model perceptions of domains are aligned, and how domain weighting influences generalization. We address these questions by establishing formal connections between gradient dynamics and domain distributions, offering a theoretical framework that clarifies the role of domains in training dynamics. Building on this analysis, we introduce DoGraph, a reweighting framework that formulates data scheduling as a graph-constrained optimization problem. Extensive experiments on GPT-2 models of varying scales demonstrate that DoGraph consistently achieves competitive performance.

The landscape of extremely low-resource machine translation (MT) is characterized by perplexing variability in reported performance, often making results across different language pairs difficult to contextualize. For researchers focused on specific language groups—such as ancient languages—it is nearly impossible to determine if breakthroughs reported in other contexts (e.g., African or American languages) result from superior methodologies or are merely artifacts of benchmark collection. To address this, we introduce the FRED Difficulty Metrics—Fertility Ratio (F), Retrieval Proxy (R) Pre-training Exposure (E) and Corpus Diversity (D) —that serve as dataset-intrinsic metrics to contextualize reported scores. Our findings reveal that a significant portion of result variability is explained by train-test overlap and pre-training exposure rather than model capability. Additionally, we identify that underperforming XLR languages—particularly extinct and non-Latin indigenous languages—suffer from poor tokenization coverage (high token fertility), highlighting structural limitations of transfer learning for languages outside pre-trained models’ representation space. By providing these indices alongside performance scores, we enable more transparent evaluation of cross-lingual transfer and provide a more reliable foundation for the XLR MT community.

Membership inference attacks (MIAs) are a canonical way to assess a machine learning model’s privacy properties. Although several attempts have been made to evaluate MIAs on language models, the extant literature has suffered numerous difficulties in constructing clean evaluations to test new techniques. In particular, subtle distribution shifts between member and non-member sets can undermine the statistical validity of MIAs; recent work has underscored this by showing that “blind” methods with no access to the underlying model can perform far better than published methods on the same benchmarks. This paper constructs a benchmark for principled evaluation of MIAs against LLMs, by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution. Therefore, all open-source models with intermediate checkpoints and public training data can be converted into MIA testbeds. We apply our framework to a half-dozen published attacks on the Pythia and OLMo family of models, from 70M to 7B parameters. To facilitate further privacy research, we open-source a modular library for designing and implementing attacks in this setting: https://github.com/safr-ai-lab/pandora_llm.

As the use of digital devices and social media grows among younger users, cybergrooming has emerged as a critical social concern for protecting vulnerable minors online. However, research on automated cybergrooming detection remains limited due to data scarcity. Building on previous studies that conceptualize cybergrooming as a form of luring communication, this paper investigates the potential transferability of corpora from luring or manipulative contexts for cybergrooming detection.

Recent work has shown that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon, where models generate malicious responses even to prompts unrelated to the original insecure code-writing task. Such cross-domain generalization of harmful behavior underscores the need for a deeper understanding of the algorithms, tasks, and datasets that induce emergent misalignment. In this work, we extend this study by demonstrating that emergent misalignment can also arise from narrow refusal unlearning in specific domains. We perform refusal unlearning on Cybersecurity and Safety concept, and evaluate EMA by monitoring refusal scores across seven responsible AI (RAI) domains, Cybersecurity, Safety, Toxicity, Bias, Sensitive Content, Medical/Legal, and Privacy. Our work shows that narrow domain unlearning can yield compliance responses for the targeted concept, however, it may also propagate EMA to unrelated domains. Among the two intervened concepts, Cybersecurity and Safety, we find that the safety concept can have larger EMA impact, i.e, causing lower refusal scores, across other unrelated domains such as bias. We observe this effect consistently across two model families, Mistral-7b-0.3v, and Qwen-7b-2.5. Further, we show that refusal unlearning augmented with cross-entropy loss function on a small set of retain data from the affected domains can largely, if not fully, restore alignment across the impacted domains while having lower refusal rate on the concept we perform unlearning on. To investigate the underlying causes of EMA, we analyze concept entanglements at the representation level via concept vectors. Our analysis reveals that concepts with higher representation similarity in earlier layers are more susceptible to EMA after intervention when the refusal stream is altered through targeted refusal unlearning.

Lowering the numerical precision of model parameters and computations is widely adopted to improve the efficiency of retrieval systems. However, when computing relevance scores between the query and documents in low-precision, we observe spurious ties due to the reduced granularity. This introduces high variability in the results based on tie resolution, making the evaluation less reliable. To address this, we propose a more robust retrieval evaluation protocol designed to reduce score variation. It consists of: (1) High-Precision Scoring (HPS), which upcasts the final scoring step to higher precision to resolve tied candidates with minimal computational cost; and (2) Tie-aware Retrieval Metrics (TRM), which report expected scores, range, and bias to quantify order uncertainty of tied candidates. Our experiments test multiple models with three scoring functions on twelve retrieval datasets to demonstrate that HPS dramatically reduces tie-induced instability, and TRM accurately recovers expected metric values. This combination enables a more consistent and reliable evaluation system for lower-precision retrieval.

Vision-language models (VLMs) are increasingly used to make visual content accessible via text-based descriptions. In current systems, however, description specificity is often conflated with their length. We argue that these two concepts must be disentangled: descriptions can be concise yet dense with information, or lengthy yet vacuous. We define specificity relative to a contrast set, where a description is more specific to the extent that it picks out the target image better than other possible images. We construct a dataset that controls for length while varying information content, and validate that people reliably prefer more specific descriptions regardless of length. We find that controlling for length alone cannot account for differences in specificity; it matters how the length budget is applied. These results support evaluation approaches that directly prioritize specificity over verbosity.

Large language models (LLMs) trained on massive text corpora may inadvertently memorize sensitive or copyrighted content, motivating the need for more targeted unlearning. Selective LLM unlearning focuses on identifying token-level or span-level unlearning targets within a text, rather than treating entire sequences as unlearning targets. However, many existing selective approaches depend on external supervision to identify unlearning targets, which may misalign unlearning objectives with the model’s internal behavior. In this paper, we propose a selective span-level unlearning method that is grounded entirely in model-intrinsic information. Our method first estimates token-level importance scores by contrasting gradient information induced by forget and retain datasets, identifying tokens that disproportionately contribute to information targeted for unlearning. These token-level importance scores are then used as anchors to identify coherent span-level unlearning targets via a self-consistency–based generation process, allowing the model to determine stable spans based on its own predictions. Experiments on two LLM unlearning benchmarks show that our approach achieves comparable unlearning performance while substantially better preserving retained knowledge.

Safe clinical deployment of Large Language Models (LLMs) requires not only high accuracy but also robust uncertainty calibration to ensure models defer to clinicians when appropriate. Our paper investigates how social descriptors of a patient (specifically sexual orientation and religious affiliation) distort these uncertainty signals and model accuracy. Evaluating nine general-purpose and biomedical LLMs on 2,364 medical questions and their counterfactual variants, we demonstrate that identity markers cause a "calibration crisis". *Homosexual* markers consistently trigger performance drops, and intersectional identities produce idiosyncratic, non-additive harms to calibration. Moreover, a clinician-validated case study in an open-ended generation setting confirms that these failures are not an artifact of the multiple-choice format. Our results demonstrate that the presence of social identity cues does not merely shift predictions; it affects the reliability of confidence signals, posing a significant risk to equitable care and safe deployment in confidence-based clinical workflows.

Multimodal abductive reasoning — the generation and selection of explanatory hypotheses from partial observations — is a cornerstone of intelligence. Current evaluations of such ability in vision–language models (VLMs) are largely confined to static, single-agent tasks. Inspired by Dixit, we introduce DixitWorld, a comprehensive evaluation suite designed to deconstruct this challenge. DixitWorld features two core components: DixitArena, a dynamic, multi-agent environment that evaluates both hypothesis generation (a "storyteller" crafting cryptic clues) and hypothesis selection ("listeners" choosing the target image from decoys) under imperfect information; and DixitBench, a static QA benchmark that isolates the listener’s task for efficient, controlled evaluation. Results from DixitArena reveal distinct, role-dependent behaviors: smaller open-source models often excel as creative storytellers, producing imaginative yet less discriminative clues, whereas larger proprietary models demonstrate superior overall performance, particularly as listeners. Performance on DixitBench strongly correlates with listener results in DixitArena, validating it as a reliable proxy for hypothesis selection. Our findings reveal a key trade-off between generative creativity and discriminative understanding in multimodal abductive reasoning, a central challenge for developing more balanced and capable vision-language agents.

Events exhibit rich semantic relations that are essential for understanding the unfolding of real-world processes. Although large language models (LLMs) have achieved strong performance on event relation extraction, how event relations are internally represented and utilized remains unclear. In this paper, we present UERLens, an interpretability framework for understanding event relations in LLMs. Specifically, we first construct UERBench, a counterfactual dataset for event relation analysis that covers causal, temporal, and sub-event relations. Based on counterfactual pairs, we identify relation-sensitive internal features by comparing model activations. We then examine the functional role of these features through model manipulation, including model intervention and model training. Experimental results show that event relations are encoded through structured and layer-specific internal features. Disabling relation-sensitive features leads to performance drops of over 22%, while enhancing them yields improvements of up to 7%. Furthermore, leveraging these interpretable features to train a lightweight classifier significantly improves event relation extraction, achieving F1 gains of up to 24% for causal relations.

We present a multilingual coreference dataset of 827k tokens of fiction in 7 languages: Bahasa Indonesia, Chinese, Dutch, English, Italian, Korean, and Spanish. The dataset includes full stories of diverse lengths, ranging from 500 to 17k words. We discuss our annotation scheme focusing on characters and language-specific challenges we encountered. Finally we present evaluation results of a neural coreference system trained on our dataset. We show that jointly training a system across all languages provides a strong improvement over monolingually trained models. The dataset is available under a creative commons license in CoNLL-2012 and CorefUD format at https://github.com/GOLEM-lab/GOLEMcoref/

Large language models (LLMs) sometimes exhibit language confusion when generating non-English text. Existing approaches typically rely on fine-tuning to mitigate this issue. In contrast, we propose a tuning-free paradigm for reducing language confusion. Within this paradigm, we introduce two methods: Language-Aware Token Boosting (LATB), which applies targeted perturbations to tokens associated with the desired language, and Adaptive Language-Aware Token Boosting (Adaptive-LATB), which dynamically adjusts these perturbations based on the model’s confidence in the intended language. Experiments demonstrate that our methods effectively improve multilingual alignment by reducing language confusion, while maintain the summarization quality without requiring any additional fine-tuning. Our code is publicly available.[<https://github.com/scbdatax/genai-datax-language-aware-token-boosting>].

Test-time alignment methods offer a promising alternative to fine-tuning by steering the outputs of large language models (LLMs) at inference time with lightweight interventions on their internal representations. Recently, a prominent and effective approach, RE-Control (Kong et al., 2024), has proposed leveraging an external value function trained over the LLM’s hidden states to guide generation via gradient-based editing. While effective, this method overlooks a key characteristic of alignment tasks, i.e. that they are typically formulated as learning from human preferences between candidate responses. To address this, in this paper we propose a novel preference-based training framework, **Pref-CTRL**, that uses a multi-objective value function to better reflect the structure of preference data. Our approach has outperformed RE-Control on two benchmark datasets and showed greater generalization on out-of-domain datasets. Our source code is available at https://github.com/UTS-nlPUG/pref-ctrl.

Textual humor is enormously diverse and computational studies need to account for this range, including intentionally bad humor. In this paper, we curate and analyze a novel corpus of sentences from the Bulwer-Lytton Fiction Contest to better understand "bad" humor in English. Standard humor detection models perform poorly on our corpus, and an analysis of literary devices finds that these sentences combine features common in existing humor datasets (e.g., puns, irony) with metaphor, metafiction and simile. LLMs prompted to synthesize contest-style sentences imitate the form but exaggerate the effect by over-using certain literary devices, and including far more novel adjective-noun bigrams than human writers.

When an AI assistant remembers that Sarah is a single mother working two jobs, does it interpret her stress differently than if she were a wealthy executive? As personalized AI systems increasingly incorporate long-term user memory, understanding how this memory shapes emotional reasoning is critical. We investigate how user memory affects emotional intelligence in large language models (LLMs) by evaluating 15 models on human validated emotional intelligence tests. We find that identical scenarios paired with different user profiles produce systematically divergent emotional interpretations. Across validated user-independent emotional scenarios and diverse user profiles, systematic biases emerged in several high-performing LLMs where advantaged profiles received more accurate emotional interpretations. Moreover, LLMs demonstrate significant disparities across demographic factors in emotion understanding and supportive recommendations tasks, indicating that personalization mechanisms can embed social hierarchies into models’ emotional reasoning. These results highlight a key challenge for memory-enhanced AI: systems designed for personalization may inadvertently reinforce social inequalities.

As human-agent interaction (HAI) evolves toward long-term social companionship, users expect *Original Character (OC)* agents to maintain a consistent persona, manage shared memories, and adapt to ever-changing preferences. However, LLM-based agents optimized by prompting or SFT exhibit a generalization gap: they behave as myopic instruction followers, leading to cascading errors in multi-turn interactions. For the agents to learn trajectory-level value functions that enable farsighted decision-making, we propose the NSARL framework, which formalizes OC companion agents’ interactions as a POMDP and decomposes the agent into three sub-policies (Router, Memory, and Persona), optimized via closed-loop RL from AI feedback (RLAIF) with verifiable rewards in a graph-constrained action space. Our preliminary experiments indicate a trade-off: SFT yields stronger persona generation, while NSARL improves structural logic, through conservative strategies (e.g., over-routing) that increase workflow completeness, advocating for a hybrid deployment strategy.

Group Relative Policy Optimization (GRPO) has emerged as a promising approach for improving the reasoning capabilities of large language models. However, it struggles to effectively balance the trade-off between exploration and exploitation during training, often resulting in suboptimal performance. Motivated by the theoretical insight that changes in entropy are governed by the covariance between token probabilities and their corresponding advantages, we propose a hyperparameter-free, covariance-weighted optimization method that dynamically down-weights extreme token-level updates via a Gaussian kernel. This approach automatically reduces the instability caused by the exploration-exploitation trade-off while preserving informative learning signals. Extensive empirical evaluations show that our approach improves downstream performance across reasoning benchmarks compared with GRPO, and effectively stabilizes entropy as training progresses.

Recent works proposed test-time alignment methods that rely on a small aligned model as a proxy that guides the generation of a larger base (unaligned) model. The implicit reward approach skews the large model distribution, whereas the nudging approach defers the generation of the next token to the small aligned model when the large base one is unconfident about its outcome. In this work, we first show that both approaches can be reduced to sampling from similar graphical models, where they differ only in the definition of a rejection criterion (or distribution). Moreover, we argue that the confidence criterion is ill-motivated due to linguistic phenomena like ambiguous phrasing. We propose a novel rejection criterion based on a conservative confidence bet. Experimentally, our novel approach outperforms previous work on several datasets.

GraphRAG augments large language models with structured knowledge graphs, enabling graph-based context selection and a more integrated view of the knowledge space. However, recent work shows that GraphRAG exposes a new attack surface: corpus-level knowledge poisoning can inject spurious entities and relationships during graph construction, corrupting query-specific subgraphs and steering the generator toward incorrect answers. We propose Hop-wise Guard for GraphRAG (HoG-GRAG), a defense layer between retriever and generator that decomposes multi-hop questions into ordered subqueries, monitors hop-wise execution for poisoning-induced inconsistencies, and locally repairs the retrieved subgraph by pruning compromised entities and relationships and adding only minimal missing evidence. Experiments on multi-hop datasets and multiple GraphRAG configurations show that HoG-GRAG recovers a large fraction of the lost performance. The code is available at https://github.com/CyberScienceLab/HoG-GRAG.

LLM-based agents for text-to-SQL often struggle with latency-performance trade-off, where performance improvements come at the cost of latency or vice versa. We reformulate text-to-SQL generation within the lens of software test coverage where the original query is prepared with a suite of test cases with simpler, atomic SQLs that are executed in parallel and together ensure semantic coverage of the original query. After iterating on test case coverage, the final SQL is generated only when enough information is gathered, leveraging the explored test case SQLs to ground the final generation. We validated our framework on a state-of-the-art benchmark for text-to-SQL, Spider 2.0, achieving a new state-of-the-art with 70.2% execution accuracy.

Large reasoning models such as DeepSeek-R1 and their distilled variants achieve strong performance on complex reasoning tasks. Yet, distilling these models often demands large-scale data for supervised fine-tuning (SFT), motivating the pursuit of data-efficient training methods. To address this, we propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model’s weaker skills, and (2) Skill-aware fine-tuning, which encourages explicit skill decomposition during problem solving. With only 1,000 training examples selected from a 100K teacher-generated corpus, our method surpasses random SFT baselines by +1.6% on Qwen3-4B and +1.4% on Qwen3-8B across five mathematical reasoning benchmarks. Further analysis confirms that these gains concentrate on skills emphasized during training, highlighting the effectiveness of skill-centric training for efficient reasoning distillation.

Scientific feasibility assessment asks whether a claim is consistent with established knowledge and whether experimental evidence could support or refute it. We frame feasibility assessment as a diagnostic reasoning task in which, given a hypothesis, a model predicts feasible or infeasible and justifies its decision. We evaluate large language models (LLMs) under controlled knowledge conditions (hypothesis-only, with experiments, with outcomes, or both) and probe robustness by progressively removing portions of the experimental and/or outcome context. Across multiple LLMs and two datasets, providing outcome evidence is generally more reliable than providing experiment descriptions. Outcomes tend to improve accuracy beyond what internal knowledge alone provides, whereas experimental text can be brittle and may degrade performance when the context is incomplete. These findings clarify when experimental evidence benefits LLM-based feasibility assessment and when it introduces fragility.

Analyses of parasocial cues in live-stream chats require accurate, efficient, and scalable annotation. However, manual annotation is tedious, and large language models (LLMs) often make mistakes when applying subjective, discourse-dependent labels. This study proposes Context-aware Batching for Stream Annotation with LLMs (CaBSALLM), an efficient pipeline that incorporates lightweight conversational context and a novel dynamic batching method to improve throughput and scalability. Compared with state-of-the-art pipelines, this generalizable approach is significantly more time- and cost-efficient while achieving comparable or better predictive performance and agreement.

In this paper, we find that the generated preceding tokens, which are not directly related to the answer, may still significantly push the large language model (LLM) towards the target answer. More crucially, the biased connotations of target answer in the preceding tokens can also transfer to other prompts. This finding suggests that the LLM may intentionally use the semantically unrelated tokens to help the generation of the target answer. Our finding offers a new perspective on understanding the long-range dependency phenomena in LLMs.

While having a significant potential for parallel processing in theory, diffusion-based non-autoregressive text generation remains inefficient due to the need for multiple denoising steps. Performance degrades sharply if a low number of steps is used, such as in flow matching. To enable accurate one-step generation, we propose a novel shortcut flow-matching model that learns to directly predict multi-step denoising outcomes in a single step. Experiments conducted on three datasets demonstrate consistent improvements over classic flow-matching, with BLEU scores more than doubling on two datasets. We also tested five different ways of extending shortcut models with commonly used techniques.

Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has emerged as the standard post-training paradigm for large language models (LLMs). However, the conventional SFT process, driven by Cross-Entropy (CE) loss, often induces mode collapse, where models over-concentrate on specific response patterns. This lack of distributional diversity severely restricts the exploration efficiency required for subsequent RL. While recent studies have attempted to improve SFT by replacing CE loss, aiming to preserve diversity or refine the update policy, they fail to adequately balance diversity and accuracy, thereby achieving sub-optimal performance after RL. To address the mode collapse problem, we propose SED-SFT, which adaptively encourages diversity based on the token exploration space. This framework introduces a selective entropy regularization term with a selective masking mechanism into the optimization objective. Extensive experiments across eight mathematical benchmarks demonstrate that SED-SFT significantly enhances generation diversity with a negligible computational overhead increase compared with CE loss, yielding average improvements of 2.06 and 1.20 points in subsequent RL performance over standard CE-based baselines on Llama-3.2-3B-Instruct and Qwen2.5-Math-7B-Instruct, respectively.

Large language models (LLMs) are fluent but often brittle when interpretation depends on external information (e.g., events or participant roles), as next-token prediction does not explicitly encode situation-level semantic constraints. FrameNet provides a structured account of semantics through its inventory of frames, roles, and relations. We present a scalable framework that injects frame-semantic knowledge into LLMs via LoRA, moving from fact-oriented prompting to principle-oriented supervision over the full FrameNet inventory. The supervision encodes semantic constraints through semantic types, sense-aware definitions, frame relations, and role-annotated examples. To test whether this knowledge generalizes beyond surface cues, we use Natural Language Inference (NLI) as a diagnostic task for event-level reasoning. Experiments on CONFER and SNLI show consistent gains over Meta-Llama-3.1-8B-Instruct in zero-shot and few-shot settings, especially for entailment and contradiction. Complementary semantic role labeling analyses further indicate improved sensitivity to frame, role, and span structure.

Federated learning (FL) enables collaborative training without raw data sharing, but still risks training data memorization. Existing FL memorization detection techniques focus on one sample at a time, underestimating more subtle risks of cross-sample memorization. In contrast, recent work on centralized learning (CL) has introduced fine-grained methods to assess memorization across all samples in training data, but these assume centralized access to data and cannot be applied directly to FL. We bridge this gap by proposing a framework that quantifies both intra- and inter-client memorization in FL using fine-grained cross-sample memorization measurement across all clients. Based on this framework, we conduct two studies: (1) measuring subtle memorization across clients and (2) examining key factors that influence memorization, including decoding strategies, prefix length, and FL algorithms. Our findings reveal that FL models do memorize client data, particularly intra-client data, more than inter-client data, with memorization influenced by training and inferencing factors.

Figurative language recognition poses significant challenges in NLP, particularly when distinguishing between fine-grained rhetorical categories such as metaphor, metonymy, and simile. This paper formulates the problem as a four-way sentence-level classification task and proposes FL-MSCL, a unified framework integrating prompt-based knowledge injection with supervised contrastive learning. Experiments across both unified and single-class benchmarks demonstrate that FL-MSCL achieves competitive performance compared to State-of-the-Art (SOTA) methods, indicating consistent advantages in cross-category generalization and category-specific detection.

Evaluating Question Answering (QA) systems in low-resource Indic languages remains challenging due to the scarcity of annotated data, high linguistic diversity, and the absence of reliable evaluation metrics. Many Indian languages are severely underrepresented, making it difficult to accurately assess the performance of Large Language Models (LLMs) on QA tasks. Commonly used metrics like BLEU, ROUGE-L, and BERTScore, while successful in machine translation and resource-rich scenarios, tend to perform poorly in low-resource QA settings. These metrics often exhibit issues such as compressed scoring ranges, excessive zero scores, and weak alignment with human judgments. To overcome these limitations, this work introduces the LRM²QAS (Language Robust Multi-aspect Metrics for Question Answering Systems). This composite evaluation framework integrates semantic similarity, factual completeness, numerical accuracy, and contextual relevance. The proposed metric is evaluated across eight Indic-language QA tasks using multiple LLMs, as well as on open-domain benchmarks NaturalQuestions (NQ) and TriviaQA (TQ). Across all settings, LRM²QAS demonstrates stronger agreement with human evaluation, as measured by Pearson, Spearman, and Kendall correlation coefficients. Experimental findings highlight that LRM²QAS provides more precise distinctions between model outputs and aligns more closely with human judgment, offering a reliable framework for evaluating multilingual QA in low-resource Indic languages.

Aligning text-to-speech (TTS) system outputs with human feedback through preference optimization has been shown to effectively improve the robustness and naturalness of LLM-based TTS models. Current approaches primarily require paired desirable and undesirable samples at the utterance level. However, such pairs are often limited in TTS output data, and utterance-level formulation prevents fine-grained token-level optimization needed for accurate pronunciation alignment. In this study, we propose TKTO that eliminates the need for paired data, enabling a more data-efficient training paradigm, and directly targets token-level units, automatically providing fine-grained alignment signals without token-level annotations. TKTO improves the challenging Japanese TTS accuracy by 39% and reduces CER by 54%, leveraging 6× more training data and assigning 12.8× stronger reward to targeted tokens.

We study how visual artifacts introduced by diffusion-based inpainting affect language generation in vision-language models. We use a two-stage diagnostic setup in which masked image regions are reconstructed and then provided to captioning models, enabling controlled comparisons between captions generated from original and reconstructed inputs. Across multiple datasets, we analyze the relationship between reconstruction fidelity and downstream caption quality. We observe consistent associations between pixel-level and perceptual reconstruction metrics and both lexical and semantic captioning performance. Additional analysis of intermediate visual representations and attention patterns shows that inpainting artifacts lead to systematic, layer-dependent changes in model behavior. Together, these results provide a practical diagnostic framework for examining how visual reconstruction quality influences language generation in multimodal systems.

Commonsense reasoning often involves evaluating multiple plausible interpretations rather than selecting a single atomic answer, yet most benchmarks rely on single-label evaluation, obscuring whether statements are jointly plausible, mutually exclusive, or jointly implausible. We introduce LOGICAL-COMMONSENSEQA, a benchmark that reframes commonsense reasoning as logical composition over pairs of atomic statements using plausibility-level operators (AND, OR and NEITHER/NOR). Evaluating instruction-tuned, reasoning-specialized, and fine-tuned models under zero-shot, few-shot, and chain-of-thought prompting, we find that while models perform reasonably on conjunctive and moderately on disjunctive reasoning, performance degrades sharply on negation-based questions. LOGICAL-COMMONSENSEQA exposes fundamental reasoning limitations and provides a controlled framework for advancing compositional commonsense reasoning.

Effective biomedical information retrieval requires modeling domain semantics and hierarchical relationships among biomedical texts. Existing biomedical generative retrievers built on coarse binary relevance signals, limiting their ability to capture semantic overlap. We propose BioHiCL - Biomedical Retrieval with Hierarchical Multi-Label Contrastive Learning, which leverages hierarchical MeSH annotations to provide structured supervision for multi-label contrastive learning. Our models, BioHiCL-Base (0.1B) and BioHiCL-Large (0.3B), achieve promising performance on biomedical retrieval, sentence similarity, and question answering tasks, while remaining computationally efficient for deployment.

Large Language Model agents can seeminglyplan and act, yet their language use is oftentreated primarily as an interface for instructingactions and reporting results. We argue that thisframing is one important cause of recurrent coordination failures in human-facing and multiagent settings, including ungrounded assumptions, silent goal misalignment, brittle protocoladherence, and failures to maintain or updateshared dialogue state over time, a limitation previously linked to the absence of explicit common ground tracking in collaborative systems(Geib et al., 2022). Drawing from classical dialogue system research on joint action, commonground, grounding, repair, and incremental processing, we re-frame dialogue as part of theplanning loop itself (rather than its output). Wedistill this re-framing into concrete implicationsfor agentic architecture and evaluation, including explicit representations of shared commitments, clarification as a first class action available to the policy, and process metrics that approximate grounding behavior, repair, and commitment formation rather than task completionalone. We lastly discuss how dialogue-centeredrequirements can inform standards and governance for safe deployment of agentic systems.

This paper introduces Late Code Chunking (LC²), a chunking strategy designed to improve the semantic understanding of code segments for Large Language Models (LLMs). Repository-level code completion requires predicting the completion of unfinished code by leveraging cross-file context spread across a repository. However, when retrieved fragments have missing semantics—the loss of structural or behavioral information during chunking—LLMs struggle to interpret the target code. To address this, LC² refines retrieved chunks by constructing a dual context: a "Code Retrieval Context" optimized for similarity-based search, and a "Code Comprehension Context" that serves as a late enrichment step through context expansion and augmentation. This dual-context design reduces information loss due to chunking and enhances the ability of LLMs to utilize retrieved code. Additionally, we introduce an Asymmetric Query-Chunk Sizing strategy to further optimize retrieval quality by minimizing query noise. Our experiments demonstrate that LC² provides robust performance gains, achieving a statistically significant 19.7% improvement in Exact Match accuracy on the CrossCodeEval benchmark compared to the best existing chunking method.

Search-engine date filters are widely used to enforce pre-cutoff retrieval in retrospective evaluations of search-augmented forecasters. We show this approach is unreliable across two major search engines: auditing Google Search’s before: filter and DuckDuckGo’s date-range filter, we find that at least one retrieved page contains major post-cutoff leakage for 71% of questions on Google and 81% on DuckDuckGo, and the answer is directly revealed for 41% and 55%, respectively. Using gpt-oss-120b to forecast with these leaky documents, we demonstrate inflated prediction accuracy (Brier score 0.10 vs. 0.24 with leak-free documents). We characterize recurring leakage mechanisms, including updated articles, related-content modules, unreliable metadata, and absence-based signals, and argue that date-restricted search on these engines is insufficient for credible retrospective evaluation. We recommend stronger retrieval safeguards or evaluation on frozen, time-stamped web snapshots.

Large language models (LLMs) encode problem difficulty as an internal signal that can be linearly decoded from their residuals. Given their multilingual capabilities, we investigate whether this meta-cognitive signal is language-agnostic and how it is organized across the model’s layers by training linear probes on the AMC subset of the Easy2Hard benchmark, translated into 21 languages. We found that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layers) internal representations, that exhibit functionally different behaviors. Probes trained on deep representations achieve high accuracy when evaluated on the same language but exhibit weaker cross-lingual transfer. In contrast, probes trained on shallow representations generalize better across languages, despite achieving lower within-language performance. This closely aligns with existing findings in LLM interpretability, showing that models tend to operate in an abstract conceptual space before producing language-specific outputs. Our results suggest that this two-stage organizational principle extends beyond simple semantic processing to meta-cognitive properties such as problem difficulty, highlighting an internal control signal that is not tied to surface meaning.

We present T⋆, a simple TraceRL-based curriculum for progressive block-size scaling in masked diffusion language models (MDMs).Starting from an AR-initialized small-block MDM, T⋆ gradually increases the block size while re-optimizing the denoising policy at each stage, enabling higher-parallelism decoding with limited degradation on math reasoning benchmarks. Across two SDAR scales and three benchmarks, T⋆ consistently outperforms direct large-block TraceRL and is substantially more stable during training. Our schedule analysis suggests that the learned policy does not simply revert to a strictly left-to-right order; instead, it retains block-size-specific non-monotone updates while improving accuracy.

Generative listwise reranking leverages global context for superior retrieval but is plagued by intrinsic position bias, where models exhibit structural sensitivity to input order independent of relevance. Existing mitigations present a dilemma: inference-time aggregation incurs prohibitive latency, while training-based methods often fail to eradicate ingrained priors, particularly in compact models. To resolve this dilemma, we propose CapCal (Content-Agnostic Probability Calibration), a training-free framework that mechanically decouples positional bias from ranking decisions. By estimating the bias distribution via content-free placeholders, CapCal rectifies output logits through an entropy-adaptive contrastive mechanism. Evaluations across 10 benchmarks confirm that CapCal achieves superior performance among training-free methods while preserving single-pass efficiency. Notably, it unlocks the latent potential of lightweight models (e.g., 0.6B), delivering absolute NDCG gains exceeding 10 points and outperforming computationally expensive data augmentation strategies.

Fine-tuning large language models (LLMs) for downstream tasks remains expensive, even with parameter-efficient methods like Low-Rank Adaptation (LoRA). In this regard, meta-learning approaches such as Model-Agnostic Meta-Learning for LLMs (MAML-en-LLM) and Amortized Bayesian Meta-Learning for LoRA (ABMLL) have emerged as promising solutions for rapid downstream LLM adaptation. However, these methods fundamentally couple two distinct objectives: learning generalizable initializations and enabling efficient task adaptation. We argue that this coupling limits both the quality of learned representations and adaptation efficiency. In this paper, we introduce **DeGAML-LLM** (**De**coupled **G**eneralization and **A**daptation in **M**eta-**L**earning for **LLM**s), a novel framework that explicitly separates these two objectives through dedicated parameter spaces. Specifically, we maintain a generalization module that learns task-agnostic representations across the task distribution, and an adaptation module that specializes in rapid task-specific adjustment. Extensive experiments on common-sense reasoning, mathematics, logic, social, medical and coding benchmarks across model scales demonstrate that DeGAML-LLM outperforms existing meta-learning and standard multi-task baselines.

Replicating AI research is a crucial yet challenging task for large language model (LLM) agents. Existing approaches often struggle to generate executable code, primarily due to insufficient background knowledge and the limitations of retrieval-augmented generation (RAG) methods, which fail to capture latent technical details hidden in referenced papers. Furthermore, previous approaches tend to overlook valuable implementation-level code signals and lack structured knowledge representations that support multi-granular retrieval and reuse. To overcome these challenges, we propose Executable Knowledge Graphs (xKG), a pluggable, paper-centric knowledge base that automatically integrates code snippets and technical insights extracted from scientific literature. When integrated into three agent frameworks with two different LLMs, xKG shows substantial performance gains (10.9% with o3-mini) on PaperBench, demonstrating its effectiveness as a general and extensible solution for automated AI research replication.

Multimodal Reasoning Models (MRMs) leveraging Chain-of-Though (CoT) based thinking have revolutionized mathematical and logical problem-solving. However, we show that this paradigm struggles with generalized spatial intelligence. We perform a comprehensive evaluation of sixteen models across thirteen spatial benchmarks and identify a critical gap: CoT prompting consistently degrades performance in visual spatial reasoning. Furthermore, through a novel No-Image++ ablation, we demonstrate that MRMs and CoT prompted MLMs suffer from severe shortcut learning, and hallucinate visual details from textual priors even when the image is absent. These findings challenge the efficacy of text-only CoT for spatial tasks and underscore the need for vision-centric reasoning paradigms.

Diffusion models have introduced a generative paradigm for Named Entity Recognition (NER), formulating the task as refining entityspans from noise. While promising, our analysis on the ACE2004 dataset reveals a limitation when training with Exponential MovingAverage (EMA): the model performance often peaks at a single inference step (γ = 1) and plateaus or degrades with additional steps. Thissuggests that under standard stable training configurations, the model may function primarily as a one-step generator rather thanleveraging the iterative refinement capability characteristic of diffusion models. To address this, we propose an Initializer-Restorerapproach. Instead of initializing the reverse process from random Gaussian noise, we utilize a preliminary set of candidate spansgenerated by a standard NER model (e.g., BERT or GLiNER). This allows the diffusion model to start from an informed, diverse prior,enabling effective iterative restoration. We investigate different training strategies for the restorer and find that a hybrid strategy mixingground truth and noisy predictions is essential. Experiments on ACE2004, GENIA, and CleanCoNLL show that our approach improvesperformance over the baseline, particularly when multiple restoration steps are employed. For instance, on CleanCoNLL, our methodachieves an F1 score of 94.70%, compared to 93.79% for the baseline. Our code is available at https://github.com/longtrieu-ai/Initializer-Restorer-NER.

Unsupervised representation learning using masked language modeling on the language of life has transformed protein research, enabling the analysis of a protein universe that is expanding at an exponential pace. However, most current models rely solely on sequence data, overlooking decades of expert-curated biological knowledge stored in natural language. While recent multimodal and knowledge-graph-based approaches attempt to bridge this gap, they often rely on shallow functional labels that lack the contextual depth of full textual narratives. We present Protein-STORY, a general pipeline that synthesizes protein embeddings from diverse, multi-source text descriptions. At the core of our approach is a novel network architecture designed for the semantic compression of multi document embeddings, which integrates high-fidelity functional and structural insights into a unified representation. Our experiments demonstrate that Protein-STORY produces biologically meaningful embeddings (r ≈ 0.75) that outperform existing models on diverse downstream tasks (+2 pts F1 in function prediction). Furthermore, by projecting the story of a protein into a natural language semantic space, our model enables effective zero-shot text-prompted protein search.

Non-autoregressive (NAR) models have been mainly developed to improve decoding efficiency. Lately, they have also shown great potential in controlled text generation tasks. In this work, we investigate the decoding space of NAR models through lexically constrained machine translation tasks, and develop a search-based decoding algorithm named LexMAP, which is comparable to the autoregressive Grid Beam Search (GBS) method. Our analysis reveals several interesting properties of NAR decoding: 1) the NAR-based method does not suffer from the MAP degradation issue as the autoregressive method does; 2) AR beam search exhibits strong positional bias, in which the candidates only diverge at the end of the sequence; 3) NAR search explores a larger portion of the probability space, suggesting that the search algorithm better exploits the model’s potential.