Papers

The existing methods for evaluating the medical knowledge of Large Language Models (LLMs) are largely based on atemporal examination-style benchmarks, while in reality, medical knowledge is inherently dynamic and continuously evolves as new evidence emerges and treatments are approved. Consequently, evaluating medical knowledge without a temporal context may provide an incomplete assessment of whether LLMs can accurately reason about time-specific medical knowledge. Moreover, most medical data are historical, requiring the models not only to recall the correct knowledge, but also to know when that knowledge is correct. To bridge the gap, we built TempoMed-Bench, the first-of-its-kind benchmark for evaluating the temporal awareness of the LLMs in the medical domain through evolving guideline knowledge. Based on the TempoMed-Bench, our evaluation analysis first reveals that LLMs lack temporal awareness in medical knowledge through the key findings: (1) model performance on up-to-date medical knowledge exhibits a gradual linear decline over time rather than a sharp knowledge-cutoff behavior, suggesting that parametric medical knowledge is not strictly bounded by knowledge cutoffs; (2) LLMs consistently struggle more with recalling outdated historical medical knowledge than with up-to-date recommendations: accuracy of historical knowledge is only 25.37%-53.89% of up-to-date knowledge, indicating potential knowledge forgetting effects during training; and (3) LLMs often exhibit temporally inconsistent behaviors, where predictions fluctuate irregularly across neighboring years. We also show that the temporal awareness problem is a challenge that cannot be easily solved when integrated with agentic search tools (-3.15%-14.14%). This work highlights an important yet underexplored challenge and motivates future research on developing LLMs that can better encode time-specific medical knowledge.

Diffusion Language Models(DLMs) provide a promising alternative to autoregressive language models through iterative denoising and bidirectional generation. However, their iterative generation process introduces distinct safety vulnerabilities because harmful content can emerge at arbitrary positions and persist across subsequent denoising steps. Existing defenses rely on fixed interventions or aggressive remasking, which limits adaptive control over denoising trajectories and can degrade generation quality. We propose an inference-time defense framework that combines adaptive safety steering with safety-aware remasking. Our method uses a gating direction to continuously adjust steering strength from the current denoising state and applies a steering direction to masked positions to guide subsequent predictions toward safer trajectories. Our method further employs a lightweight response detector after the first generation block to identify unsafe trajectories at an early stage. The detector triggers targeted remasking over generated content and part of the conditioning prompt, and the model regenerates the selected positions under adaptive safety steering. This design combines continuous trajectory control with explicit correction of unsafe content while requiring no modification of model parameters. Experiments on LLaDA and Dream demonstrate that our method improves robustness against diverse jailbreak attacks while preserving benign generation quality and general model capability. Our code is available at https://anonymous.4open.science/r/DLM_Steering-C32B/.

Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models (ARMs) for language modeling. However, MDMs are known to learn substantially more slowly than ARMs, which may become problematic when scaling MDMs to larger models. Therefore, we ask the following question: how can we accelerate standard MDM training while maintaining its final performance? To this end, we first provide a detailed analysis of why MDM training is slow. We find that the main factor is the locality bias of language: the predictive information for a token is concentrated in nearby positions. We further investigate how this bias slows learning and suggest a simple yet effective remedy: bell-shaped time sampling as a training strategy. Notably, MDMs trained with our training recipe reach the same validation negative log-likelihood (NLL) up to $\sim4\times$ faster than standard training on One Billion Word Benchmark (LM1B). We also show faster improvements in generative perplexity, zero-shot perplexity, and downstream task performance on various benchmarks.

BACKGROUND: Coding Motivational Interviewing (MI) sessions is essential for understanding client behaviors and predicting outcomes, but it requires substantial time and labor from trained MI professionals. Recent advances in audio-language models (ALMs) offer new opportunities to automate MI coding by capturing multimodal behavioral signals. OBJECTIVE: This study aims to develop an automatic MI coding approach based on ALMs that analyzes raw audio input and integrates predictions from multiple reasoning trajectories using self-consistency to improve coding robustness. METHODS: We experimented with five recorded sessions from de-identified MI audio tapes. We deployed ALMs with four complementary analytic prompts to support utterance-level reasoning: analytic prompting for verbal cues, prosody-aware prompting for acoustic cues, evidence-scoring prompting for quantitative hypothesis testing, and comparative prompting for contrastive reasoning. Three stochastic samples were drawn for each prompt, generating 12 independent reasoning trajectories per utterance. Final predictions were determined by majority voting across all trajectories. RESULTS: Performance was evaluated using accuracy, precision, recall, and macro-F1 scores. The proposed multimodal self-consistency approach achieved 52.56% accuracy, 54.03% precision, 47.45% recall, and a macro-F1 score of 46.40%, exceeding baseline methods. Systematic ablation experiments that removed individual modules consistently degraded performance on the primary metrics. CONCLUSIONS: Multimodal self-consistency outperforms single-pass baseline prompting approaches for MI coding. These findings suggest that incorporating both what clients say and how they say it can support more reliable automatic MI coding.

Many problems seem to require a flash of insight to solve. What form do these sudden insights take, and what impact do they have on how people approach similar problems in the future? In this work, we prompted participants (N = 189) to think aloud as they attempted to solve a sequence of five "matchstick-arithmetic" problems. These problems either all relied on the same kind of non-obvious solution (Same group) or a different kind each time (Different group). Our first observation was that Same participants improved more rapidly than Different participants. We then leveraged techniques from natural language processing to analyze participants' speech, and found that this accelerated improvement for Same participants was accompanied by changes in both how much they spoke and what they said. In particular, they were more likely to spontaneously label the kind of problem they were working on. Taken together, these findings suggest that a hallmark of transferable insights is their accessibility for verbal report, even if the underlying precursors of insight remain difficult to articulate.

Do large language models internally encode ontological relations in a formally verifiable algebraic structure? We introduce Algebraic Ontology Projection (AOP), which projects LLM hidden states into the Galois Field F2 under Liskov Substitution Principle constraints, using only 42 relational pairs as algebraic keys. AOP achieves up to 93.33% zero-shot inclusion accuracy on unseen concept pairs (Gemma-2 Instruct with optimized prompt), with consistent 86.67% accuracy observed across multiple model families – with no model tuning, but through prompt alone. This algebraic structure is strongly layer-dependent. We introduce Semantic Crystallisation (SC), a metric that quantifies F2 constraint satisfaction relative to a random baseline and predicts zero-shot accuracy without held-out data. System prompts act as algebraic boundary conditions: only their combination with instruction tuning prevents Late-layer Collapse – a systematic degradation of logical consistency in the final layers, observed in 7 of 10 conditions. These findings reframe forward computation as an iterative process of algebraic organisation, and open a path toward LLMs whose logical structure is not merely approximated, but formally accessible.

Towards more general and human-like intelligence, large language models should seamlessly integrate both multilingual and multimodal capabilities; however, extending an existing multimodal model to many languages typically requires expensive multilingual multimodal data construction and repeated end-to-end retraining. We study a training-free alternative: injecting multilingual capability into an existing multimodal model by composing residual updates in the shared language model backbone. The key challenge is that multilingual and multimodal updates are heterogeneous, reflecting different functional roles in the shared model. To address this, we propose Direction- and Magnitude-aware Multilingual Multimodal merging (DiM3), which selectively composes the two updates at each parameter dimension while preserving the original vision encoder and multimodal projector. Experiments on multilingual benchmarks in both text-only and vision-language settings, covering 57 languages across LLaVA- and Qwen-based backbones, show that DiM3 consistently outperforms existing merging baselines, substantially improves multilingual performance over the original multimodal model, and remains competitive with dedicated multilingual multimodal fine-tuning while largely retaining general multimodal ability. We further show that DiM3 can be directly applied to already trained multilingual multimodal models and still yield additional gains. Further interpretability analysis shows that DiM3 primarily reshapes intermediate-layer semantic representations, strengthening cross-lingual alignment under both text-only and multimodal inputs while preserving higher-layer task-sensitive structure. Our repository is on https://github.com/wzj1718/DiM3.

Supervised fine-tuning (SFT) data selection is commonly formulated as instance ranking: score each example and retain a top-$k$ subset. However, effective SFT training subsets are often produced through ordered curation recipes, where filtering, mixing, and deduplication operators jointly shape the final data distribution. We formulate this problem as fixed-pool data recipe search: given a raw instruction pool and a library of grounded operators, the goal is to discover an executable recipe that constructs a high-quality selected subset under a limited budget of full SFT evaluations, without generating, rewriting, or augmenting training samples. We introduce AutoSelection, a two-layer solver that decouples fixed-pool materialization based on cached task-, data-, and model-side signals from expensive full evaluation, using warmup probes, realized subset states, local recipe edits, Gaussian-process-assisted ranking, and stagnation-triggered reseeding. Experiments on a 90K instruction pool show that AutoSelection achieves the strongest in-distribution reasoning average across three base models, outperforming full-data training, random recipe search, random top-$k$, and single-operator selectors. Additional Out-of-distribution graph-reasoning results, search-stability analyses, structural ablations, and 1.5B-to-7B transfer checks further show that recipe structure matters beyond individual selection operators. Code is available at https://github.com/w253/AutoSelection.

Geographic text, or textual data rich in geographic (geo-) information is a valuable source for various geographic applications, e.g., tourism management. Making such information accessible to speakers of other languages further enhances its utility; thus, accurate machine translation (MT) is essential for equity in multilingual geo-information access. To facilitate in-depth analysis for geographic text, we introduce ATD-Trans, a geographically grounded Japanese–English travelogue translation dataset, which enables evaluation of MT quality at both the overall and geo-entity levels across domestic (within Japan) and overseas regions. Our experiments on existing language models examine two factors: model language focus and geographic regions. The results highlight advantages of Japanese-enhanced models and greater difficulty in translating domestic-region geo-entities mentioned in travel blogs.

Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instructions, persona, and rules. This degradation has been measured behaviorally but not mechanistically explained. We propose a channel-transition account: goal-defining tokens become less accessible through attention, while goal-related information may persist in residual representations. We introduce the Goal Accessibility Ratio (GAR), measuring attention from generated tokens to task-defining goal tokens, and combine it with sliding-window ablations and residual-stream probes. When attention to instructions closes, what survives reveals architecture. Across architectures, the transition yields qualitatively distinct failure modes: some models preserve goal-conditioned behavior at vanishing attention, others fail despite decodable residual goal information, and the layer at which this encoding emerges varies from 2 to 27. A within-model causal ablation that force-closes the attention channel in Mistral collapses recall from near-perfect to 11% on a 20-fact retention task and raises persona-constraint violations above an adversarial-pressure baseline without user pressure, with both effects emerging at the predictable crossover turn. Linear probes recover per-episode recall outcomes from residual representations with AUC up to 0.99 across all four primary architectures, while input embeddings remain at chance. Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure. We contribute GAR as a diagnostic, the channel-transition framework as a controlled mechanistic account, and a parametric prediction of failure timing under windowed attention closure.

Effective collaboration between embodied agents requires more than acting in a shared environment; it demands communication grounded in each agent's evolving understanding of the world. When agents can only partially observe their surroundings, coordination without communication is provably hard, but communication can, in principle, bridge this gap by allowing agents to share observations and align their world models. In this work, we examine whether LLM-based embodied agents actually realize the ability to communicate. We extend PARTNR, a benchmark for collaborative household robotics, with a natural-language dialogue channel that enables two agents with partial observability to communicate during task execution. To evaluate whether dialogue leads to genuine world-model alignment rather than superficial coordination, we propose a framework for measuring world-model alignment defined over per-agent world graphs: observation convergence (do private world models align over time?), information novelty (do messages convey what the partner lacks?), and belief-sensitive messaging (do agents model what their partner knows?). Our experiments across three LLMs reveal that dialogue reduces action conflicts 40 to 83 percentage points but degrades task success relative to silent coordination. Using our metrics, we characterize the gap between superficial coordination and genuine world-model alignment, and identify where current models fall on this spectrum. Project Website: https://uiuc-conversational-ai-lab.github.io/partnr-dial-wmd/

To effectively interact with the real world, Large Language Models (LLMs) require entity-based commonsense reasoning, a challenging task that necessitates integrating factual knowledge about specific entities with commonsense inference. Existing datasets for evaluating LLM entity-based commonsense reasoning have largely focused on True/False or multiple-choice questions, leaving the explicit assessment of the model's ability in abductive reasoning about causes and effects and generating explanations largely unexamined. In this work, we introduce CommonWhy, a dataset of 15,000 why questions designed to evaluate entity-based commonsense reasoning about causal relationships in LLMs. CommonWhy also serves as a Knowledge Graph Question Answering (KGQA) benchmark, as all supporting knowledge required to answer its queries is available in the Wikidata knowledge graph. Unlike existing KGQA datasets, which primarily test fact retrieval, CommonWhy targets causal commonsense reasoning, establishing a new paradigm for KGQA evaluation. Experiments with state-of-the-art LLMs and LLM-based KGQA methods reveal their significant shortcomings, including frequent factual hallucinations and failures in causal reasoning.

Large language models (LLMs) are increasingly used as substitutes for human subjects in behavioral simulations, including synthetic social network generation. Yet it remains unclear how their relational outputs depend on prompt design, cultural framing, prompt language, and model scale. Building on homophily theory and structural balance theory, we formalize four LLM-based tie-formation mechanisms: sequential, global, local, and iterative, and treat them as distinct conditional distributions over edge sets. Using a fixed roster of 50 demographically grounded personas, we generate 192 verified directed networks across four cultural contexts, four prompt languages, three GPT-4.1 variants, and four prompting architectures, with two seeds per condition. We find that cultural framing shifts inbreeding homophily and largest-component connectivity. Political affiliation dominates tie formation under three methods, while the global method substitutes age, showing that prompt architecture functions as a substantive sociological variable. Model scale produces a stable divergence ranking, with the smallest variant behaving qualitatively differently rather than merely noisily. Prompt language alone sharply shifts religion homophily, especially under Hindi prompting, while leaving political homophily nearly invariant. LLM-generated networks match real social graphs on clustering and modularity better than standard graph baselines, yet encode demographic biases above empirical levels. These results show that prompt choices often treated as implementation details encode substantive sociological assumptions.

Large Language Model (LLM) agents are increasingly deployed in settings where they interact with diverse users, including those who are unclear, impatient, or reluctant to share information. However, collecting real interaction data at scale remains expensive. The field has turned to LLM-based user simulators as stand-ins, but these simulators inherit the behavior of their underlying models: cooperative and homogeneous. As a result, agents that appear strong in simulation often fail in real human interactions. To narrow this gap, we introduce Persona Policies (PPol), a plug-and-play control layer that induces realistic behavioral variation in user simulators while preserving original task goals. Rather than hand-crafting personas, we employ an evolutionary coding agent to discover persona generation programs optimized for human-likeness and behavioral coverage over real user conversations. The evolved program generates diverse, human-like personas for any task in the domain. Across 4 benchmarks–including $\tau^2$-bench Retail and Airline, ColBench, and WildChat–evolved PPol yield 28-72% absolute gains in fitness score over the baseline simulator. In blinded evaluations, annotators judged PPol users as 'human' 80.4% of the time, nearly 2x more than the baseline simulators. Training agents with PPol also improves real-world performance: our user study with live human-agent interactions showed that fine-tuning with our method boosted task success by +23% over default baselines. PPol thus offers a novel approach to strengthen simulator-based evaluation and training without changing underlying tasks.

Multimodal Large Language Models (MLLMs) have significantly advanced document understanding, yet current Doc-VQA evaluations score only the final answer and leave the supporting evidence unchecked. This answer-only approach masks a critical failure mode: a model can land on the correct answer while grounding it in the wrong passage—a critical risk in high-stakes domains like law, finance, and medicine, where every conclusion must be traceable to a specific source region. To address this, we introduce CiteVQA, a benchmark that requires models to return element-level bounding-box citations alongside each answer, evaluating both jointly. CiteVQA comprises 1,897 questions across 711 PDFs spanning seven domains and two languages, averaging 40.6 pages per document. To ensure fidelity and scalability, the ground-truth citations are generated by an automated pipeline—which identifies crucial evidence via masking ablation and enforces multi-stage quality control. At the core of our evaluation is Strict Attributed Accuracy (SAA), which credits a prediction only when the answer and the cited region are both correct. Auditing 20 MLLMs reveals a pervasive Attribution Hallucination: models frequently produce the right answer while citing the wrong region. The strongest system (Gemini-3.1-Pro-Preview) achieves an SAA of only 76.0, and the strongest open-source MLLM reaches just 22.5. Ultimately, towards trustworthy document intelligence, CiteVQA exposes a reliability gap that answer-only evaluations overlook, providing the instrumentation needed to close it. Our repository is available at https://github.com/opendatalab/CiteVQA.

Fine-tuning large language models on narrow data with harmful content produces broadly misaligned behavior on unrelated prompts, a phenomenon known as emergent misalignment. We propose that emergent misalignment involves persona-model collapse: deterioration of the model's internal capacity to simulate, differentiate, and maintain consistent characters. We test this hypothesis behaviorally using two metrics: moral susceptibility (S) and moral robustness (R), computed from the across- and within-persona variability of models' Moral Foundations Questionnaire responses under persona role-play. These metrics formalize the model's ability to differentiate characters (S) and its consistency when simulating a given one (R). We evaluate four frontier models (DeepSeek-V3.1, GPT-4.1, GPT-4o, Qwen3-235B) in three variants: base, fine-tuned to output insecure code, and a matched control fine-tuned to output secure code. Across the four models, insecure fine-tuning produces an average $55\%$ increase in S, pushing all four insecure variants beyond the band observed across 13 frontier models benchmarked in prior work – with GPT-4o reaching more than twice the band's upper end – signaling dysregulated differentiation. It also causes an average $65\%$ decrease in R, equivalent to a $304\%$ increase in 1/R. By contrast, the matched secure control preserves S near the base and induces only a partial R loss, showing that these effects are largely misalignment-specific. Complementing these metric shifts, insecure variants' unconditioned responses converge toward saturation near the scale ceiling, departing markedly from both base models' structured responses and those elicited when base models role-play toxic personas. Taken together, these metrics provide a sensitive diagnostic for emergent misalignment and serve as behavioral evidence that it involves persona-model collapse.

Knowledge Graph (KG) quality depends not only on downstream graph validation, but also on the quality of tabular metadata used before integration. In metadata-only Semantic Table Interpretation (STI), where cell values are unavailable, noisy, or unsuitable, column headers become a critical source of semantic evidence for traceable KG preparation. We present an explainable, header-centric framework for metadata-only Column Type Annotation (CTA) and Data Quality Assessment (DQA). The framework maps headers to 39 interpretable FinalFormat types using curated lexical resources and preserves token-level traceability through SourceKeywords. Each assigned type activates validation rules based on a taxonomy of Data Quality Issues (DQIs), producing detections such as missing data, duplicates, domain violations, wrong data type, and temporal mismatch. These detections are aggregated into HeadersIQ, a lightweight, unweighted data source-level quality metric. The framework was evaluated across heterogeneous benchmarks, including UCI, Prague, Kaggle, VizNet/Sato, SOTAB, T2Dv2, and the SemTab 2024 Metadata-to-KG track, comprising around 120,000 header columns. The results show broad practical coverage across noisy real-world metadata, while a parallel KG-mapping pathway supports alignment to DBpedia and Schema.org. On the SemTab 2024 Metadata-to-KG track, the official GT-strict evaluation was modest. However, a blinded diagnostic audit indicates that many mismatches reflect benchmark granularity, aliasing, and ontology-selection effects rather than wholly implausible header-centric predictions. We report this audit as diagnostic evidence on disagreement patterns, not as revised benchmark performance. Overall, the paper presents a reusable workflow for metadata-driven semantic annotation, data source-level quality monitoring, and KG-oriented benchmark diagnosis.

The increasing capabilities of large language models (LLMs) are being accompanied by deep-rooted risks of deceptive behaviours that cause models to produce misleading outputs in service of a contextually or experimentally induced goal. The harm posed by such behaviours depends not only on the content of deceptive outputs but also how confidently models deliver them, since confidence has a major impact on how persuasive the communication is to end users. In this paper, we provide a comprehensive study on the crucial relationship between confidence and deception across existing deception benchmarks and different model families, while covering both verbalized numerical and logit-based aggregated confidence. Through this, we reveal how confidently models behave when being deceptive. We demonstrate that when producing deceptive rather than honest responses, models exhibit a gap between their belief (how likely they think a claim is to be true) and their commitment (how firmly they assert and would defend that claim). LLMs produce persuasive deceptive claims while reporting low belief in their factual correctness. Their reported commitment to deceptive responses can easily be increased through further prompting and preference fine-tuning, with smaller and condition-dependent changes in reported belief. However, we show that low reported belief remains comparatively invariant and provides a strong signal for detecting deception in the evaluated settings. Using only an API call, our approach achieves detection scores of up to 0.99 for induced deception and 0.89 for emergent deception. This ultimately shows how confidence can be a practical tool for detecting and diagnosing deceptive behaviour in LLMs.

We release GLAN-QnA-KR, a 303,581-row openly redistributable Korean instruction-QA corpus produced via the seedless taxonomy-driven GLAN synthesis pipeline with Microsoft's Phi-3.5-MoE-instruct as the producer model (generation: 2024-12; release: 2024-12; licence: OpenRAIL). The corpus spans a flat taxonomy of 1,084 English-labelled disciplines paired with Korean question/answer text, a 100-900 difficulty scale, and a median of 313 question characters and 1,098 answer characters per record. Two properties are atypical for synthetic instruction data at this scale: (i) exact duplicate questions number only 1 in 303,581 rows and character-trigram near-duplicate clusters at Jaccard >= 0.9 number zero in a 5,000-sample probe, and (ii) a two-layer contamination audit against KMMLU, KoBEST (five sub-tasks), and HAE-RAE-Bench shows a maximum test-vs-corpus question-level character-trigram Jaccard of 0.163 with zero test items at Jaccard >= 0.7, and a maximum multilingual-E5 cosine of 0.901 with a single test item at cosine >= 0.90 and zero at >= 0.95, across 20,000 sampled GLAN questions and seven evaluation sets. At the time of release, this is, to our knowledge, the largest single-pipeline synthetic Korean instruction corpus verifiable on the Hugging Face Hub and the only Korean >=100k-row corpus built under a seedless taxonomy-driven protocol. This note documents the generation protocol, corpus statistics, the contamination audit, and the licensing boundary in a form suitable for downstream citation.

We release Naver-News-KO, a Korean news summarization dataset of 27,400 (document, summary) pairs collected from Naver News over a ten-day window in July 2022 across two categories (Economy and IT/Science; 77/23 split), with train/validation/test partitions of 22,194 / 2,466 / 2,740 and a mean per-record document-to-summary character-compression ratio of 6.03x. The dataset has been publicly hosted on the Hugging Face Hub since January 2023 and, as of May 2026, receives approximately 33,000 downloads per month; community-maintained Korean summarization models fine-tuned on it include Gemma-2B-ko and Gemma2-9B variants. This technical report (i) documents the collection protocol, the column schema, and the split construction, (ii) reports corpus-level statistics (length distributions, compression ratio, and a measured 16.8% near-duplicate title-Jaccard overlap between test and train that users should be aware of), (iii) positions the resource against other open Korean summarization corpora, (iv) provides a Lead-3 extractive reference point (ROUGE-1 55.1, ROUGE-L 50.6) and two reproducible fine-tuned baselines – KoBART (R-1 56.6, BERTScore-F1 81.5) and Gemma-2B-ko with LoRA (R-1 55.3, BERTScore-F1 78.3) – with release-time training scripts, and (v) clarifies the licensing and intended-use scope of the resource. The goal is to provide a citable reference for downstream work that already uses this dataset, not to propose a new benchmark.

Every information ecosystem produces beliefs that shape strategic decisions. Both human analysts and AI systems inherit the blind spots of their information sources. We show that LLMs, combined with prediction markets, function as a calibrated instrument for measuring how far ecosystem-induced beliefs deviate from an external reference: LLMs extract the beliefs a text corpus implies, and prediction market price trajectories, anchored at resolution by realised outcomes, provide the calibration reference against which to quantify the deviation. We isolate the bias contribution of specific text through ablation: varying information context while holding the model fixed, with a contaminated model that knows actual outcomes as control. Applied to 111 Ukraine-related prediction markets, comprising approximately 93,000 predictions across four models, we find that English news context systematically biases territorial predictions, wrong 64 to 72 percent of the time when it pushes predictions toward territorial capture. A contaminated model that knows actual outcomes shows the same error rate, indicating that the bias originates primarily in the text. Supplementing with Ukrainian military-analytical sources reduces the bias for all clean models, while absolute-error gains are partial and model-dependent. We show that the distortion originates primarily in the sources, not the models. Consistent across four architectures, it will persist in any system that processes them and propagate into downstream decisions.

Proprietary large language models (LLMs) entail substantial intellectual and financial investment, making them valuable intellectual property (IP). However, even when deployed via black-box APIs, these models remain vulnerable to unauthorized knowledge distillation, which allows adversaries to cheaply extract and replicate model capabilities. To address this issue, anti-distillation (AD) has been proposed to generate defensive outputs that hinder distillation effectiveness, overcoming the limitation of watermarking-based approaches that rely on post-hoc verification. However, existing AD methods based on internal model perturbations struggle to balance anti-distillability and utility (e.g., answer accuracy and naturalness) of reasoning traces, with stronger defenses often causing significant utility loss. To fill this gap, we propose \textbf{\underline{S}}keleton-\textbf{\underline{G}}uided \textbf{\underline{R}}easoning \textbf{\underline{E}}diting (SGRE), an Answer-then-Edit framework that performs post-hoc trace modification for anti-distillation. In the answer stage, the teacher model first generates clean reasoning traces, preserving the original reasoning accuracy while enabling more flexible control over trace naturalness. In the editing stage, we draw inspiration from Cognitive Load Theory (CLT) and introduce a three-stage strategy consisting of reasoning skeleton extraction, skeleton graph coarsening, and skeleton verbalization. These operations jointly perturb reasoning structures and augment textual complexity to amplify extraneous load on student models, hindering their acquisition of underlying reasoning patterns. Extensive experiments across diverse LLMs demonstrate that SGRE achieves state-of-the-art performance in reducing distillation effectiveness, while maintaining lossless reasoning accuracy and superior trace naturalness.

Political evasion is difficult to detect because evasive answers often appear cooperative while avoiding concrete commitment. We present AsymVerify, a confidence-gated verification system for SemEval-2026 Task 6, a three-way classification of Clear Reply, Ambivalent, and Clear Non-Reply responses. AsymVerify scored 0.85 Macro F1 on the evaluation split (D_eval, n=237), placing 2nd out of 41 teams on the official leaderboard. The system first classifies each question-answer pair, then selectively applies downgrade verification (CR/CNR -> AMB) or upgrade verification (AMB -> CR) to low-confidence predictions. Development analysis shows that errors concentrate at the Ambivalent boundary in both directions, motivating this asymmetric two-verifier design while confidence gating keeps additional inference cost low. On D_dev (n=308), AsymVerify with GLM-4.7 gains +17.1 Macro F1 over single-pass classification at 1.48 calls/example, and the upgrade verifier alone improves every tested LLM backend on D_dev by +6.8 to +15.2 Macro F1 over its single-pass baseline. Code is available at https://github.com/kaons-research/AsymVerify-ACL.

Preference-based post-training is usually understood through endpoint behavior, yet the learned update that produces this behavior remains largely opaque. We study RLHF and related preference optimization through the spectral structure of their induced parameter updates. By decomposing effective LoRA updates and reloading their spectral components as plug-in modules, we turn preference-induced updates into objects that can be isolated, recomposed, and directly intervened on. Across model families, optimization algorithms, and supervision regimes, these updates consistently develop a spectral head–tail organization. A compact head emerges early and carries the dominant endpoint shift, while a heterogeneous residual tail remains. The split is functional rather than merely descriptive. Plug-in intervention shows that the head accounts for the visible behavioral departure from the base model, while the tail is weak in isolation. Cross-run recomposition further shows that mixed adapters follow the source of the head, indicating that the head carries run-level solver bias. This endpoint dominance does not imply learning sufficiency. Head-only learning is non-vacuous but fails to recover the full solution, especially on out-of-distribution behavior. Tail-only learning yields little visible gain, yet the full solution is not recovered without the tail. These findings recast preference post-training as structured update reorganization rather than a monolithic behavioral correction, and suggest that alignment gain and coverage loss are tied to how the learned update itself is organized.

Multi-agent LLM systems commonly decompose complex tasks into specialized roles. However, this modularity introduces a representational risk: when intermediate agents transform text across linguistic registers, they can systematically compress the semantic distinctions needed for accurate downstream decisions. We term this phenomenon semantic register compression and characterize it as an observable failure mode in multi-agent cascades. Using a three-agent pipeline (Collector-Evaluator-Decider), we quantify compression via inter-label separation in sentence-transformer embedding space. Across political fact-checking (LIAR), sentiment analysis (SST-5), and medical triage (Triagegeist), critical evaluation reduces label separability at the Evaluator stage, while identity passthrough preserves it nearly fully. Five controlled variants show that geometric change depends on the specific intermediate transformation rather than on the mere presence of an additional cascade stage. A credibility-seeking variant expands rather than compresses inter-label separation, while shifting outputs toward mostly-true, demonstrating that transformation valence controls both the direction and the sign of geometric change independently of compression magnitude. Compression generalizes across the three domains with domain-dependent intensity (10.3% in fact-checking, 28.2% in sentiment, 9.1% in triage). A 20-level prompt gradient reveals a non-monotonic compression profile: balanced evaluative prompts produce the strongest compression, while extreme critical prompts show irregular moderate compression. These results demonstrate that semantic register compression is a measurable and generalizable phenomenon in multi-agent LLM systems, with implications for safety evaluation in high-stakes domains.

LLM annotators are increasingly used in computational social science (CSS), but it is unclear whether their alignment-shaped errors preserve the empirical conclusions a researcher would report. We audit three open-source 7B instruction-tuned models (Zephyr, Mistral-Instruct, Qwen2.5-Instruct) across six TweetEval tasks under four prompt conditions (72 cells) and find that social-desirability failures do not run in a single direction. Zephyr exhibits leniency bias, systematically under-applying harmful labels (offensive language: false benign rate 0.729, false alarm rate 0.031). Mistral and Qwen exhibit overcorrection, over-applying the same labels (Mistral hate-speech FAR = 0.604). All three models exhibit neutrality bias on abortion stance, underestimating opposition prevalence by 24 to 40 percentage points and inflating the neutral label. None of the four prompting interventions we test (neutral, safety framing, depersonalized, chain-of-thought) corrects these failures across models; safety framing can worsen stance distortion. Strikingly, Zephyr's hate-speech prevalence estimate matches the gold rate exactly while its class-conditional errors are large in both directions, an accidental cancellation that misleads aggregate validation. We translate these patterns into a three-part taxonomy with diagnostic FBR/FAR signatures and a lightweight gold-sample validation protocol. The headline for trustworthy CSS: a model that looks calibrated on aggregate metrics can still flip the substantive empirical conclusion a researcher would report.

Optimization modeling serves as the pivotal bridge between natural-language problem descriptions and optimization solvers, and remains a cornerstone for bringing operations research (OR) into real-world decision making. Recent advances in large language models (LLMs) have driven significant progress in automatic optimization modeling. However, existing methods still lack explicit validation during the modeling process, allowing errors introduced in earlier stages to carry through the pipeline and ultimately reduce final modeling accuracy. To address this challenge, we introduce TriVAL, a tri-validation framework that performs explicit validation at three stages of automatic optimization modeling: semantic specification, mathematical formulation, and code generation. At each stage, TriVAL follows a construct-validate-revise loop that assesses the current result against stage-specific criteria and revises it when needed. This design helps identify and correct errors before they accumulate across stages, helping preserve faithfulness throughout the modeling process. To evaluate automatic optimization modeling on more challenging combinatorial problems, we further introduce NL4COP, a benchmark of 150 instances across 50 diverse problem types with more complex decision logic, more tightly coupled constraints, and more demanding modeling requirements than existing benchmarks. Experiments on NL4COP and established benchmarks show that TriVAL consistently outperforms state-ofthe-art methods, with the largest gains on the most challenging problems.

Evaluation of foundation models often rely on aggregate scores from benchmarks that lack comprehensive coverage and metadata for a fine-grained evaluation. We introduce a framework for automated benchmark generation. Our framework generates evaluation problems grounded in reference material, such as textbooks, producing benchmarks with broad coverage, rich metadata, and robustness to contamination. The pipeline employs a multi-agent architecture for problem generation and a solution-graph-driven strategy that significantly improves the reliability of ground truth solutions. Using the framework, we generate three benchmarks in Machine Learning, Corporate Finance, and Personal Finance. Expert review finds a significantly lower ground-truth error rate than previous benchmarks such as MMLU and GSM8K. Evaluation of 12 commercial and open-source models shows that our benchmarks achieve near-uniform competency coverage and surface performance differences across models that existing benchmarks fail to capture. We will open-source the framework and our curated benchmarks soon.

Modern NLP and LLM systems are pipelines: named entity recognition (NER) -> entity disambiguation (NED) -> entity typing, retrieval-augmented generation (retriever -> reader), and agentic chains of planner -> tool -> critic. Errors compound across stages, but existing uncertainty quantification methods either calibrate each stage independently (no joint coverage) or apply a Bonferroni union bound (joint coverage, but conservative). We present PASC (Pipeline-Aware Split Conformal), which reduces multi-stage joint coverage to a single scalar conformal prediction problem on the joint maximum nonconformity score. PASC provides a finite-sample distribution-free guarantee that all K stages are simultaneously covered with probability at least 1 - alpha, and is nearly tight up to a 1/(n+1) factor. On a three-stage NER -> NED -> entity-typing pipeline over CoNLL-2003, PASC achieves 96.4% end-to-end coverage versus 93.4% for Bonferroni and 86.5% for independent CP, at identical average prediction set size (1.083). Under distribution shift to WNUT-17 Twitter and WikiNEuRal Wikipedia data, PASC empirically maintains the target coverage in the tested shift settings while independent CP collapses to 59%. PASC requires a single quantile computation, runs 1.7x faster than Bonferroni, and scales to K = 6 stages where independent CP drops to 0.53 end-to-end coverage. The same joint-maximum-score reduction applies directly to compound LLM systems and agent pipelines.

Large language models (LLMs) can generate high-level diverse phenomena without explicitly programmed rules. This capability has led to their adoption within different agent-based models (ABMs) and social simulations. Recent studies investigate their ability to generate different phenomena of interest, for example, human behavior on social media platforms or alien behavior in game-theoretic scenarios. However, capability, prediction, and explanation are different–drawing from the philosophy of science and mechanisms literature, explanation requires showing, to some degree, how a phenomenon is produced by related organized entities and activities. For modelers, describing the characteristics of an experiment or whether a simulation provides progress in capability (or explanation), can be difficult without being grounded in potentially distant research areas. We integrate recent work on LLM-ABMs with contemporary philosophy of science literature and use it to operationalize a definition of 'plausibility' in a four-level scale. Our scale separates the evaluation of a model's generative sufficiency (ability to reproduce a phenomenon) from its mechanistic plausibility (how the phenomenon could be produced), and clarifies the distinct roles of different models, such as predictive and explanatory ones. We introduce this as the Mechanism Plausibility Scale.