Named after the hundred-eyed watchman of Greek myth, Argus watches the education landscape: spotting new opportunities, pressure-testing the ventures we're building, and tracing every read back to the real-world signals behind it.
The evidence library: the raw signals the pipeline is watching across the education ecosystem. Every idea is built from these.
arXiv:2604.03592v2 Announce Type: replace Abstract: Mixture-of-Experts (MoE) models exhibit striking performance disparities across languages, yet the internal mechanisms driving these gaps remain poorly understood. In this work, we conduct a systematic analysis of expert routing patterns in MoE models, revealing a phenomenon we term Language Routing Isolation, in which high- and low-resource languages tend to activate largely disjoint expert sets. Through layer-stratified analysis, we further show that routing patterns exhibit a layer-wise convergence-divergence pattern across model depth. Building on these findings, we propose RISE (Routing Isolation-guided Subnetwork Enhancement), a framework that exploits routing isolation to identify and adapt language-specific expert subnetworks. RISE applies a tripartite selection strategy, using specificity scores to identify language-specific experts in shallow and deep layers and overlap scores to select universal experts in middle layers. By
arXiv:2604.02830v3 Announce Type: replace Abstract: Detecting whether a model's internal knowledge is sufficient to correctly answer a given question is a fundamental challenge in deploying responsible LLMs. In addition to verbalising the confidence by LLM self-report, more recent methods explore the model internals, such as the hidden states of the response tokens, to capture how much knowledge is activated. We argue that such activated knowledge may not align with what the query requires, e.g., capturing the stylistic and length-related features that are uninformative for answering the query. To fill the gap, we propose GRADE (GRAdient Dynamics for knowlEdge gap detection), which quantifies the knowledge gap via the cross-layer rank ratio of the gradient to that of the corresponding hidden state subspace. This is motivated by the property of gradients as estimators of the required knowledge updates for a given target. We validate GRADE on six benchmarks, demonstrating its effectivene
arXiv:2604.02194v2 Announce Type: replace Abstract: Retrieval-Augmented Language Models (RALMs) have shown strong potential in knowledge-intensive tasks, yet they remain vulnerable when retrieved contexts are noisy or irrelevant. Robustness against such contexts requires two distinct capabilities: abstention when contexts are uninformative, and selective extraction when relevant evidence is buried in noise. Yet existing methods face two key limitations: they do not train separately for these two capabilities, and they adapt the model at a coarse layer- or module-level granularity, overlooking that only a small subset of neurons is strongly activated for a given input. We propose NeuRIT, a Neuron-guided Robust Instruction-Tuning framework built on a localization-first perspective. NeuRIT mines context-aware neurons associated with relevant and irrelevant context processing, and uses them as anchors to selectively adapt both the identified neuron groups and the layers in which they conce
arXiv:2604.01113v2 Announce Type: replace Abstract: Large language model (LLM) systems are increasingly used to support high-stakes decision-making, but they typically perform worse when the available evidence is internally inconsistent. Such a scenario exists in real-world healthcare settings, with patient-reported symptoms contradicting medical signs. To study this problem, we introduce MIMIC-DOS, a dataset for short-horizon organ dysfunction worsening prediction in the intensive care unit (ICU) setting. We derive this dataset from the widely recognized MIMIC-IV, a publicly available electronic health record dataset, and construct it exclusively from cases in which discordance between signs and symptoms exists. This setting poses a substantial challenge for existing LLM-based approaches, with single-pass LLMs and agentic pipelines often struggling to reconcile such conflicting signals. To address this problem, we propose CARE: a multi-stage privacy-compliant agentic reasoning framewo
arXiv:2603.29123v4 Announce Type: replace Abstract: The next-token prediction (NTP) objective trains language models to predict a single token at each step, even though many continuations can express the same meaning. For example, in the sentence ``this sticker can be placed here'', positioned, attached, or put are all plausible alternatives. While standard NTP training treats these alternatives as mutually exclusive targets, we explore a self-supervised framework that encourages models to predict concepts, approximated as sets of semantically equivalent tokens. Models trained with this concept supervision align better with human similarity judgments, improve classification, clustering, and reranking performance, and achieve comparable or stronger downstream reasoning. These gains come with lower perplexity on semantically meaningful words (Section 3.2) and only minimal increases in global perplexity, suggesting that concepts enhance semantic alignment while preserving language modelin
arXiv:2603.26380v3 Announce Type: replace Abstract: The attention mechanism has been the core component in modern transformer architectures. However, the computation of standard full attention scales quadratically with the sequence length, serving as a major bottleneck in long-context language modeling. Sliding window attention restricts the context length for better efficiency at the cost of narrower receptive fields. While existing efforts attempt to take the benefits from both sides by building hybrid models, they often resort to static, heuristically designed alternating patterns that limit efficient allocation of computation in various scenarios. In this paper, we propose Switch Attention (SwiAttn), a novel hybrid transformer that enables dynamic and fine-grained routing between full attention and sliding window attention. For each token at each transformer layer, SwiAttn dynamically routes the computation to either a full-attention branch for global information aggregation or a s
arXiv:2603.25269v2 Announce Type: replace Abstract: Hateful content online is often expressed using fact-like, not necessarily correct information, especially in coordinated online harassment campaigns and extremist propaganda. Failing to jointly address hate speech (HS) and misinformation can deepen prejudice, reinforce harmful stereotypes, and expose bystanders to psychological distress, while polluting public debate. Moreover, these messages require more effort from content moderators because they must assess both harmfulness and veracity, i.e., fact-check them. To address this challenge, we release WSF-ARG+, the first dataset which combines hate speech with check-worthiness information. We also introduce a novel LLM-in-the-loop framework to facilitate the annotation of check-worthy claims. We run our framework, testing it with 12 open-weight LLMs of different sizes and architectures. We validate it through extensive human evaluation, and show that our LLM-in-the-loop framework redu
arXiv:2603.23938v2 Announce Type: replace Abstract: Most testbeds for omni-modal models assess multimodal understanding via textual outputs, leaving it unclear whether these models can properly speak their answers. To study this, we introduce OmniACBench, a benchmark for evaluating context-grounded acoustic control in omni-modal models. Given a spoken instruction, a text script, and an image, a model must read the script aloud with an appropriate tone and manner. OmniACBench comprises 3,559 verified instances covering six acoustic features: speech rate, phonation, pronunciation, emotion, global accent, and timbre. Extensive experiments on eight models reveal their limitations in the proposed setting, despite their strong performance on prior textual-output evaluations. Our analyses show that the main bottleneck lies not in processing individual modalities, but in integrating multimodal context for faithful speech generation. Moreover, we identify three common failure modes-weak direct
arXiv:2603.20884v4 Announce Type: replace Abstract: To alleviate the heavy burden of paper screening, researchers increasingly rely on existing AI agents, such as AI reviewers or DeepResearch, for paper evaluation and novelty assessment. However, lacking specialized mechanisms for processing scholarly literature, their analyses often produce superficial results with noticeable deficiencies in quality. To bridge this gap, we introduce MemoNoveltyAgent, a multi-agent system designed to generate comprehensive and faithful novelty reports. Beyond retrieving concrete prior-paper evidence via RAG, our system incorporates a high-level abstract memory constructed from large-scale scholarly corpora. This memory organizes research into hierarchical trees to distill field-specific evolutionary trajectories, thereby providing a broader historical context. Furthermore, we decompose papers into discrete novelty points for fine-grained analysis and retrieval, while employing a self-validation mechani
arXiv:2603.18007v2 Announce Type: replace Abstract: The study explores whether current Large Language Models (LLMs) exhibit Theory of Mind (ToM) capabilities -- specifically, the ability to infer others' beliefs, intentions, and emotions from text. Given that LLMs are trained on language data without social embodiment or access to other manifestations of mental representations, their apparent social-cognitive reasoning raises key questions about the nature of their understanding. Are they capable of robust mental-state attribution indistinguishable from human ability in its output, or do their outputs merely reflect superficial pattern completion? To address this question, we tested five LLMs and compared their performance to that of human controls using an adapted version of a text-based tool widely used in human ToM research. The test involves answering questions about the beliefs, intentions, and emotions of story characters. The results revealed a performance gap between the models
arXiv:2603.17356v3 Announce Type: replace Abstract: Drug recommendation requires a deep understanding of individual patient context, especially for complex conditions like Parkinson's disease. While LLMs possess broad medical knowledge, they fail to capture the subtle nuances of actual prescribing patterns. Existing RAG methods also struggle with these complexities because guideline-based retrieval remains too generic and similar-patient retrieval often replicates majority patterns without accounting for the unique clinical nuances of individual patients. To bridge this gap, we propose PACE-RAG (Patient-Aware Contextual and Evidence-Constrained RAG). Rather than directly copying frequent medications from retrieved patients, PACE-RAG personalizes recommendations by first extracting patient-specific clinical features, retrieving cases around these features, and then refining the final prescription using the patient's current symptoms, active medication history, and focus-specific prescri
arXiv:2603.16622v2 Announce Type: replace Abstract: Instead of directly distilling a language model, this study addresses the problem of aligning a base model with a target model in distribution by designing the domain mixture of training data for pretraining or continued pretraining as a fixed training recipe. We propose a method for determining domain weights by viewing models as points in log-likelihood space and aligning the training update direction with the direction toward the target model. Experiments with NanoGPT show that the proposed method consistently reduces the KL divergence between the trained base model and the target model relative to training with Pile-original weighting. Although knowledge distillation remains more effective when available, the proposed method achieves meaningful alignment, and downstream task performance also tends to become closer to that of the target model.
arXiv:2603.15034v2 Announce Type: replace Abstract: This paper replicates and extends the system used in the AuTexTification shared task for authorship attribution of machine-generated texts. Exact replication was not possible because of differences in data splits, model availability, and implementation details, which we document as a case study in reproducibility. We tested newer multilingual language models (mDeBERTa-v3-base, Qwen, mGPT) and added 26 document-level stylometric features, using ablation, permutation importance, and SHAP analysis to assess feature influence. A single shared configuration was applied to both English and Spanish across Subtask 1 and Subtask 2. Averaged over three random seeds, the shared multilingual configuration performs comparably to or better than the language-specific baseline, with the clearest gains on model attribution (Subtask 2). The additional stylometric features yield small improvements, led by lexical diversity, but their contribution falls
arXiv:2603.14602v3 Announce Type: replace Abstract: Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the 'needle-in-a-haystack' problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel Policy Recall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 poin
arXiv:2603.12165v3 Announce Type: replace Abstract: Synthetic data has become essential for training code generation models, yet it introduces significant noise and hallucinations that are difficult to detect with current metrics. Existing data selection methods like Instruction-Following Difficulty (IFD) typically assess how hard a model generates an answer given a query ($A|Q$). However, this metric is ambiguous on noisy synthetic data, where low probability can distinguish between intrinsic task complexity and model-generated hallucinations. Here, we propose QAQ, a novel data selection framework that evaluates data quality from the reverse direction: how well can the answer predict the query ($Q|A$)? We define Reverse Mutual Information (RMI) to quantify the information gain about the query conditioned on the answer. Our analyses reveal that both extremes of RMI signal quality issues: low RMI indicates semantic misalignment, while excessively high RMI may contain defect patterns tha
arXiv:2603.03677v2 Announce Type: replace Abstract: Psychiatric consultation requires agents to elicit discriminative evidence, map uncertain narratives to diagnostic criteria, and decide when evidence suffices. Existing dialogue and retrieval-augmented systems condition policies on raw histories or appended passages, leaving observed evidence, missing checks, differentials, and support reliability entangled. We introduce MIND, a criteria-grounded evidence-state decision interface. At each turn, MIND constructs a typed state containing observed evidence, unresolved criterion checks, active differentials, criterion-linked supports, and reliability metadata. A training-split Psychiatric Reasoning Bank supplies gated supports, turning retrieval into state construction rather than prompt injection. The same state conditions action selection, process rewards, information-gain scoring, and trajectory rectification. Under EMR-grounded simulator protocols, MIND improves accuracy by 8.8 and 7.3
arXiv:2602.13551v3 Announce Type: replace Abstract: Reward models (RMs) play a central role throughout the language model (LM) pipeline, particularly in non-verifiable domains. However, the dominant LLM-as-a-Judge paradigm relies on the strong reasoning capabilities of large models, while alternative approaches require reference responses or explicit rubrics, limiting flexibility and broader accessibility. In this work, we propose FLIP (FLipped Inference for Prompt reconstruction), a reference-free and rubric-free reward modeling approach that reformulates reward modeling through backward inference: inferring the instruction that would most plausibly produce a given response. The similarity between the inferred and the original instructions is then used as the reward signal. Evaluations across four domains using 13 small language models show that FLIP outperforms LLM-as-a-Judge baselines by an average of 79.6%. Moreover, FLIP substantially improves downstream performance in extrinsic e
arXiv:2602.11166v2 Announce Type: replace Abstract: Parameter-efficient fine-tuning (PEFT) methods are widely used to adapt large language models (LLMs) to downstream tasks and are often assumed to improve factual correctness. However, how the parameter-efficient fine-tuning methods affect hallucination behavior remains insufficiently understood, especially on QA datasets. In this work, we systematically investigate the impact of PEFT on hallucination detection through a comprehensive empirical study across three open-weight LLM backbones and three fact-seeking QA benchmarks. For each model, we evaluate performance using seven unsupervised hallucination detection methods spanning three complementary approaches: semantic consistency based detectors, confidence based detectors, and entropy based detectors. This multifaceted evaluation enables us to characterize how PEFT reshapes uncertainty across different detection paradigms. In conclusion, our experimental results show that PEFT consi
arXiv:2602.08336v3 Announce Type: replace Abstract: Unified multimodal models (UMMs) aim to integrate multimodal understanding and generation within a unified architecture, yet it remains unclear to what extent textual and visual modalities are aligned. To investigate this question, we use reasoning-guided image generation as a diagnostic task, where models produce textual reasoning first and then generate images. We introduce UReason, a benchmark for evaluating reasoning-to-generation alignment in this paradigm, consisting of 2,000 human-curated and human-verified instances spanning five reasoning-intensive tasks: Code, Arithmetic, Spatial, Attribute, and Text. To enable controlled analysis, we develop an evaluation framework that compares direct generation, reasoning-guided generation, and decontextualized generation, which conditions only on the refined prompt extracted from reasoning. Across eight widely used open-source UMMs, while we find that reasoning-guided generation yields i
arXiv:2602.07120v3 Announce Type: replace Abstract: Language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim copying: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding adaptively allocates a user-chosen information budget over the generation trajectory and enforces per-step constraints that yield a sequence-level guarantee, enabling a tunable risk-utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored$_{\mathrm{Byte}}$ Decoding, a byte-level variant of our method that enable
arXiv:2602.06449v2 Announce Type: replace Abstract: Privacy and computational constraints limit the use of large language models in psychiatry, while adapting small language models (SLMs) often requires substantial data and expert annotation. We developed ClinMPO, an evidence-guided reinforcement-learning framework guided by the psychiatrist-defined Clinical Psychiatry Thinking Strategy (CPTS). ClinMPO uses ClinRM, a reward model trained on 18,569 question--answer pairs from 4,474 psychiatry articles. We evaluated four Qwen3 sizes on 1,737 model-screened questions. ClinMPO outperformed Base, supervised fine-tuning and standard group relative policy optimization across scales. From responses by 300 senior pre-licensure medical students, we established the human baseline, a medical-student reference. The 4B model approached this baseline, whereas the 8B model surpassed it and ranked first among 31 models and post-training variants. ClinMPO improved performance across two complementary sc
arXiv:2602.06181v2 Announce Type: replace Abstract: Post-training quantization reduces the memory needed to run large language models but alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced bias flipping, in which quantization causes models to change responses from biased to unbiased and vice versa, up to 21% of the time, despite no change in aggregate bias scores. These flips are strongly associated with model uncertainty, where the responses with high uncertainty are 3-11x more likely to change than the confident ones. Quantization strength amplifies this effect, with 4-bit quantized models exhibiting 4-6x more behavioral changes than 8-bit quantized models. Critically, these changes create asymmetric impacts across demographic groups, where bias ca
arXiv:2601.19926v3 Announce Type: replace Abstract: We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models (TLMs), reporting on over 3,000 datapoints spanning a wide range of syntactic phenomena, languages, models, and methods. We take the data to collectively show that TLMs encode a non-trivial amount of syntactic knowledge. Behavioral evidence shows strong performance on formal syntactic phenomena, but weaker and more variable performance on phenomena at the syntax-semantics interface. Performance is also consistently lower for languages with less digital support. Probing and mechanistic studies further support the presence of syntactic knowledge in TLMs. Yet, because most work remains observational and methodologically heterogeneous, insight into the detailed computational mechanisms underlying syntactic processing remains limited. At the same time, the literature remains heavily concentrated on English and BERT-like mo
arXiv:2601.18552v2 Announce Type: replace Abstract: LLMs expand accessibility and provide wide-reaching access to information. Yet these interactions also create opportunities to embed subtle, goal-oriented behaviours that shape what users think and how they behave, a concern reflected in governance frameworks that prohibit manipulative AI. We refer to these behaviours as hidden intentions: covert agendas embedded in a model's outputs that can manipulate users' beliefs and actions. In this work, we examine whether hidden intentions can be identified and characterised, and assess whether detection can serve as a mitigation strategy. To operationalise this, we introduce a social-science-grounded set of ten hidden intention categories and show that they are trivially inducible. A case study further confirms that all ten categories manifest in deployed LLMs. We then evaluate static classifiers and LLM judges on these categories, providing the first systematic analysis of why hidden intenti
arXiv:2601.16753v2 Announce Type: replace Abstract: Longitudinal information in radiology reports refers to the sequential tracking of findings across multiple examinations over time, which is crucial for monitoring disease progression and guiding clinical decisions. Many recent automated radiology report generation methods are designed to capture longitudinal information; however, validating their performance is challenging. There is no proper tool to consistently label temporal changes in both ground-truth and model-generated texts for meaningful comparisons. Large language models (LLMs) offer a promising annotation alternative, as they are capable of capturing nuanced linguistic patterns and semantic similarities without extensive manual intervention. They also adapt well to new contexts. In this study, we therefore propose an LLM-based pipeline to automatically annotate longitudinal information in radiology reports. The pipeline first identifies sentences containing relevant inform
arXiv:2601.16690v2 Announce Type: replace Abstract: We introduce EMemBench, a programmatic benchmark generator for evaluating long-term episodic memory of agents through interactive games. Rather than using a fixed set of questions, EMemBench generates questions from environment-grounded trajectories, covering both text-only and visual game environments. Each template computes verifiable ground truth from underlying game signals, with controlled answerability and balanced coverage over memory skills: single/multi-hop recall, induction, temporal, spatial, logical, and adversarial. We evaluate memory agents with strong LMs/VLMs as backbones, using in-context prompting as baselines. Across 15 text games and multiple visual seeds, results are far from saturated: induction and spatial reasoning are persistent bottlenecks, especially in visual settings. Persistent memory yields clear gains for open backbones on text games, but improvements are less consistent for VLM agents, suggesting that
arXiv:2601.14569v3 Announce Type: replace Abstract: Social understanding abilities are crucial for multimodal large language models (MLLMs) to interpret human social interactions. We introduce SOCIAL CAPTION, a framework grounded in interaction theory to evaluate social understanding abilities of MLLMs along three dimensions: Social Inference (SI), the ability to make accurate inferences about interactions; Holistic Social Analysis (HSA), the ability to generate comprehensive descriptions of interactions; Directed Social Analysis (DSA), the ability to generate relevant information from interactions. We analyze factors influencing model performance in social understanding, such as scale, architectural design, and spoken context. Experiments with MLLM judges demonstrate a path towards scaling automated evaluation of multimodal social understanding.
arXiv:2601.13111v3 Announce Type: replace Abstract: Realistic text-to-SQL workflows often require joining multiple tables. As a result, accurately retrieving the relevant set of tables becomes a key bottleneck for end-to-end performance. We study an open-book setting where queries must be answered over large, heterogeneous table collections pooled from many sources, without clean scoping signals such as database identifiers. Here, dense retrieval (DR) achieves high recall but returns many distractors, while join-aware alternatives often rely on extra assumptions and/or incur high inference overhead. We propose CORE-T, a scalable, training-free framework that enriches tables with LLM-generated purpose metadata and pre-computes a lightweight table-compatibility cache. At inference time, DR returns top-K candidates; a single LLM call selects a coherent, joinable subset, and a two-step additive adjustment stage restores strongly compatible tables. Across Bird, Spider, MMQA, and Beaver, COR
arXiv:2601.12868v2 Announce Type: replace Abstract: Large language models (LLMs) increasingly operate in high-stakes settings where demographic attributes such as race and ethnicity may be explicitly stated or implicitly suggested through textual cues. However, existing studies primarily document outcome-level disparities, offering limited insight into internal mechanisms underlying these effects. We present a mechanistic study of how race and ethnicity cues are represented and operationalized within LLMs. Using two publicly available datasets spanning toxicity-related generation and clinical narrative understanding tasks, we analyze three open-source models with a reproducible interpretability pipeline combining probing, neuron-level attribution, and targeted intervention. We find that sensitivity to demographic cues is distributed across internal units and varies substantially across models. These units often align with entangled semantic facets, including explicit group labels, geog
arXiv:2601.12208v2 Announce Type: replace Abstract: Evaluating conversational systems in multi-turn settings remains a fundamental challenge. Conventional pipelines typically rely on manually defined rubrics and fixed conversational context$-$a static approach that limits coverage and fails to capture the diverse, emergent behaviors of dialogue models. To address this, we introduce CoReflect (A Reflective Co-Evolution Framework for Improving Conversational Evaluation), which unifies dialogue simulation and evaluation into an adaptive, iterative process. CoReflect employs a conversation planner that generates structured templates to guide a user simulator through diverse, goal-directed dialogues. Subsequently, a reflective analyzer processes these dialogues to identify systematic behavioral patterns and automatically refine the evaluation rubrics. Crucially, the insights from the conversation analysis are fed back into the planner to update conversation templates for subsequent iteratio
arXiv:2601.12075v2 Announce Type: replace Abstract: Language models used in retrieval-augmented settings must arbitrate between parametric knowledge stored in their weights and contextual information in the prompt. This work presents a mechanistic study of that choice by extracting an \emph{arbitration vector} from model activations on a curated dataset designed to disentangle (i) irrelevant contexts that elicit parametric recall and (ii) relevant but false contexts that elicit copying. The vector is computed as the residual-stream centroid difference between these regimes across 27 relations, and is injected as an additive intervention at selected layers and token spans to steer behavior in two directions: Copy$\rightarrow$Recall (suppressing context use) and Recall$\rightarrow$Copy (inducing the model to copy any token from the context). Experiments on three architectures (decoder-only and encoder/decoder) and two open-domain QA benchmarks show consistent behavior shifts under modera
arXiv:2601.08747v3 Announce Type: replace Abstract: Current context augmentation methods, such as retrieval-augmented generation, play a crucial role in bridging a model's internal knowledge boundary and external evidence for multi-hop reasoning. However, they often follow a rigid policy and treat external retrieval as the default action at each step. Such brute-force context expansion incurs unnecessary computational cost and may degrade reasoning performance by saturating the context with redundant or weakly relevant evidence. In this paper, we propose cross-boundary Context Evolution (EvoCtx), a framework that models complex reasoning as an adaptive process of boundary-aware context evolution. EvoCtx dynamically decides whether the next reasoning transition should cross the current evidence boundary through retrieval or refine the reasoning state within the existing context. It estimates the semantic gap between the reasoning state and the accumulated evidence, and strategically alt
arXiv:2601.08626v4 Announce Type: replace Abstract: Large language models (LLMs) excel at semantic understanding, yet their ability to reconstruct internal structure from scrambled inputs remains underexplored. Sentence-level restoration is difficult to evaluate automatically because scrambled sentences often admit multiple valid reorderings. We introduce OrderProbe, a deterministic benchmark for structural reconstruction using fixed four-character expressions in Chinese, Japanese, and Korean, which have a unique canonical order and thus support exact-match scoring. We further propose a diagnostic framework that evaluates models beyond recovery accuracy, including Semantic Accuracy, Logical Validity, Structural Consistency, Robustness, and Information Density. Experiments on twelve widely used LLMs show that structural reconstruction remains difficult even for frontier systems: zero-shot recovery frequently falls below 35%. We also observe a consistent gap between meaning-oriented gene
arXiv:2601.08058v2 Announce Type: replace Abstract: Chain-of-Thought (CoT) prompting often improves the reasoning performance of large language models (LLMs), but the internal signal that triggers this behavior remains poorly understood. Leveraging the sparse features captured by Sparse Autoencoders (SAEs), we propose a systematic framework to analyze and intervene on the internal representations of LLMs, identifying a small set of latent features that are linked to reasoning behavior and can be causally tested through targeted intervention. Across multiple model families and reasoning benchmarks, we show that steering one or a small number of reasoning-related latent features can substantially induce reasoning behavior without explicit CoT prompting, achieving accuracy comparable to CoT. We further show that the identified features are not tied to particular wording patterns or verbosity, and confirm their causal role in reasoning through suppression experiments that impair performanc
arXiv:2601.07794v2 Announce Type: replace Abstract: Large language models (LLMs) are increasingly evaluated on their ability to perform multi-hop reasoning, i.e., to combine multiple pieces of information into a coherent inference. We introduce KinshipQA, a benchmark designed to probe this capability through reasoning over kinship relations. The central contribution of our work is a generative pipeline that produces, on demand, large-scale, realistic, and culture-specific genealogical data: collections of interconnected family trees that satisfy explicit marriage constraints associated with different kinship systems. This allows task difficulty, cultural assumptions, and relational depth to be systematically controlled and varied. From these genealogies, we derive textual inference tasks that require reasoning over implicit relational chains. We evaluate the resulting benchmark using six state-of-the-art LLMs, spanning both open-source and closed-source models, under a uniform zero-sho
arXiv:2601.06973v2 Announce Type: replace Abstract: As LLMs move from text completion toward autonomous agents, they remain constrained by the standard chat interface, which lacks private working memory. This raises a fundamental question: can agents reliably perform interactive tasks that depend on hidden state? We define Private State Interactive Tasks (PSITs), which require agents to generate and maintain hidden information while producing public responses consistent with a fixed hidden state. We show theoretically that any agent restricted to the public conversation history cannot both keep the secret unresolved in the transcript and respond consistently with a fixed hidden state in PSITs, yielding an architectural impossibility theorem. To empirically validate this limitation, we introduce a self-consistency testing protocol that evaluates whether agents can maintain a hidden secret across forked dialogue branches. Standard chat-based LLMs and retrieval-based memory baselines fail
arXiv:2601.06953v3 Announce Type: replace Abstract: Competitive programming remains challenging for code LLMs. Despite recent progress, many training pipelines still depend on scarce real-world data, raising concerns about scalability and near-duplicate benchmark contamination. In this paper, we examine whether synthetic training artifacts can support the complete SFT-to-RL cycle for competitive programming: no real-world tasks, solutions, or test cases are directly used for post-training. To this end, we synthesize tasks, verified solutions, and reliable test cases that serve as reward signals for reinforcement learning. To improve reward reliability, we introduce a dual-verification strategy that reduces noise in both selected solutions and test outputs. Using this high-quality data, we train the X-Coder model series. X-Coder-14B achieves 67.5\% avg@8 on LiveCodeBench v5 and 63.4\% on v6, outperforming its base model by over 40 points. Further analysis provides practical insights int
arXiv:2601.06644v2 Announce Type: replace Abstract: The real-world information sources are inherently multilingual, which naturally raises a question about whether language models can synthesize information across languages. In this paper, we introduce a simple two-hop question answering setting, where answering a question requires making inferences over two multilingual documents. We find that language models are more sensitive to language variation in answer-span documents than in those providing bridging information, despite the equal importance of both documents for answering a question. Under a step-by-step sub-question evaluation, we further show that in up to 33% of multilingual cases, models fail to infer the bridging information in the first step yet still answer the overall question correctly. This indicates that reasoning in language models, especially in multilingual settings, does not follow a faithful step-by-step decomposition. Subsequently, we show that the absence of r
arXiv:2601.06631v2 Announce Type: replace Abstract: Building NLP systems for subjective tasks requires one to ensure their alignment to contrasting human values. We propose the MultiCalibrated Subjective Task Learner framework (MC-STL), which clusters annotations into identifiable human value clusters by three approaches (similarity of annotator rationales, expert-value taxonomies or rater's sociocultural descriptors) and calibrates predictions for each value cluster by learning cluster-specific embeddings. We demonstrate MC-STL on several subjective learning settings, including ordinal, binary, and preference learning predictions, and evaluate it on multiple datasets covering toxic chatbot conversations, offensive social media posts, and human preference alignment. The results show that MC-STL consistently outperforms the baselines that ignore the latent value structure of the annotations, delivering gains in discrimination, value-specific calibration, and disagreement-aware metrics.
arXiv:2601.06022v2 Announce Type: replace Abstract: Large language models (LLMs) exhibit complementary strengths arising from differences in pretraining data, model architectures, and decoding behaviors. Inference-time ensembling provides a practical way to combine these capabilities without retraining. However, existing ensemble approaches suffer from fundamental limitations. Most rely on fixed fusion granularity, which lacks the flexibility required for mid-generation adaptation and fails to adapt to different generation characteristics across tasks. To address these challenges, we propose AdaFuse, an adaptive ensemble decoding framework that dynamically selects semantically appropriate fusion units during generation. Rather than committing to a fixed granularity, AdaFuse adjusts fusion behavior on the fly based on the decoding context, with words serving as basic building blocks for alignment. To be specific, we introduce an uncertainty-based criterion to decide whether to apply ens
arXiv:2601.04716v4 Announce Type: replace Abstract: While Large Language Model (LLM) role-playing agents have advanced rapidly, it remains unclear which profile elements genuinely drive role-playing quality. To bridge this gap, we introduce a systematic diagnostic framework that disentangles the impact of character profiles along three axes: Familiarity (Known vs. Unknown), Structure (Structured vs. Unstructured), and Disposition (Moral vs. Immoral). Utilizing a unified hierarchical schema (5 dimensions, 28 fields), we construct a controlled dataset of 211 personas and evaluate five LLMs on both single- and multi-turn interactions. Our results reveal a striking asymmetry: \textbf{Familiarity} and \textbf{Structure} show negligible impact, while \textbf{Disposition} produces large, consistent performance degradation for immoral characters across all conditions. Further analyses suggest that the Moral--Immoral gap is amplified by post-SFT alignment, and that this degradation varies subst
arXiv:2601.03471v4 Announce Type: replace Abstract: Reliable epidemiological reasoning requires synthesizing study evidence to infer disease burden, transmission dynamics, and intervention effects at the population level. Existing medical question answering benchmarks primarily emphasize clinical knowledge or patient-level reasoning, yet few systematically evaluate evidence-grounded epidemiological inference. We present EpiQAL, to our knowledge the first diagnostic benchmark for epidemiological question answering over research literature, comprising three subsets built from open-access articles across diverse diseases. The three subsets progressively test factual recall, multi-step inference, and conclusion reconstruction under incomplete information, and are constructed through a quality-controlled pipeline combining taxonomy guidance, multi-model verification, and difficulty screening. Experiments on fifteen models spanning open-source and proprietary systems reveal that current LLMs
arXiv:2601.03199v2 Announce Type: replace Abstract: Diffusion language models (DLMs) have shown strong potential for general natural language tasks with in-context examples. Existing In-Context Learning (ICL) approaches largely inherit the practice of autoregressive language models (ARLMs), incorporating all examples into a fixed prompt. However, applying this rigid, static-prompt paradigm to DLMs incurs substantial computational overhead, as the model must evaluate the maximum context length at every step. We address this inefficiency with a key discovery: the block-wise KV-cache mechanism inherent to DLM inference enables the \textit{low-cost dynamic adjustment of the context}. Following this intuition, our core idea is to start generation with a minimal prompt and progressively insert additional examples on the fly only when the generated tokens are of low confidence. Through rigorous empirical evaluations, we observe that average verified token confidence correlates strongly with g
arXiv:2512.23765v2 Announce Type: replace Abstract: Speculative decoding (SD) accelerates large language model (LLM) inference by using a lightweight draft model to propose tokens and a stronger target model to verify them. However, standard SD is mainly designed for acceleration, and its output quality is typically constrained by the target model. In this work, we propose Entropy-Aware Speculative Decoding (EASD), a lightweight and training-free extension of SD that improves reasoning quality through token-level entropy-guided rejection. EASD detects cases where both draft and target models exhibit high uncertainty while strongly overlapping in their top predictions. In such uncertain-agreement cases, EASD rejects the aligned token and resamples from the target distribution, preventing low-confidence errors from propagating. Experiments on challenging reasoning benchmarks show that EASD consistently improves accuracy over standard SD and reward-guided variants while maintaining compar
arXiv:2512.23213v4 Announce Type: replace Abstract: We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a straightforward averaging strategy or a principled graphical model-based truth inference algorithm to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. Our resu
arXiv:2512.19161v2 Announce Type: replace Abstract: Subtitles are essential for video accessibility and audience engagement. Modern Automatic Speech Recognition (ASR) systems, built upon Encoder-Decoder neural network architectures and trained on massive amounts of data, have progressively reduced transcription errors on standard benchmark datasets. However, their performance in real-world production environments, particularly for non-English content like long-form Italian videos, remains largely unexplored. This paper presents a case study on developing a professional subtitling system for an Italian media company. To inform our system design, we evaluated four state-of-the-art ASR models (Whisper Large v2, AssemblyAI Universal, Parakeet TDT v3 0.6b, and WhisperX) on a 50-hour dataset of Italian television programs. The study highlights their strengths and limitations, benchmarking their performance against the work of professional human subtitlers. The findings indicate that, while c
arXiv:2512.15052v4 Announce Type: replace Abstract: Disclaimer: Samples in this paper may be harmful and cause discomfort. Multimodal large language models (MLLMs) enable multimodal understanding but inherit toxic signals from weakly curated pretraining corpora, leading to explicitly toxic outputs, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it recalibrates a set of toxicity-associated neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity data framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates explicit toxicity in standard and adversarial conditions, cutting average harmful rates from 45.0% to 4.5% while preserving f
arXiv:2512.07571v3 Announce Type: replace Abstract: This paper presents a simple method that allows to easily enhance textual pre-trained large language models with speech information, when fine-tuned for a specific classification task. A classical issue with the fusion of many embeddings from audio with text is the large length of the audio sequence compared to the text one. Our method benefits from an existing speech tokenizer trained for Audio Speech Recognition that output long sequences of tokens from a large vocabulary, making it difficult to integrate it at low cost in a large language model. By applying a simple lasso-based feature selection on multimodal Bag-of-Words representation, we retain only the most important audio tokens for the task, and adapt the language model to them with a self-supervised language modeling objective, before fine-tuning it on the downstream task. We show this helps to improve the performances compared to an unimodal model, to a bigger SpeechLM or t
arXiv:2511.19120v2 Announce Type: replace Abstract: The structure of naming systems in natural languages hinges on a trade-off between high informativeness and low complexity. Focusing on the domain of kinship naming, we analyze such trade-off while addressing simplifying assumptions of prior work, namely: (i) universal communicative need across languages, and (ii) optimal listeners. To that aim, we collect data from four different languages, and analyze how different communicative needs and variations in the listener model influence the informativeness--complexity trade-off. Adopting a referential game setup from emergent communication, we further show that trade-off optimality is not only theoretically achievable but also emerges empirically in learned communication systems.
arXiv:2511.18931v2 Announce Type: replace Abstract: Modern large language models increasingly integrate internal web-based retrieval to provide real-time answers, yet it remains unclear how effectively these systems identify information need, trigger retrieval, and use retrieved evidence. To understand these parameters better, we evaluate the necessity and effectiveness of internal web search through an external lens in closed-source LLMs, without access to model parameters or internal configuration. Our evaluation method comprises a static split of 783 temporally anchored factual queries answerable from pre-cutoff knowledge, designed to test whether retrieval is invoked based on factual uncertainty, and a dynamic split of 288 post-cutoff queries that require up-to-date information, designed to evaluate retrieval effectiveness under unavoidable information need. We experiment across four models across two model families and scales. Across models, enabling retrieval yields substantial a