Papers

Large-scale population surveys are essential for generating robust social and scientific insights, yet they face significant challenges, including declining response rates, increasing data collection costs, long delays between data collection and data provision, and the risk of nonresponse bias. Advances in artificial intelligence (AI) have opened up new opportunities for AI-supported survey infrastructures where the goal is to overcome these challenges without limiting the data quality. A promising AI-enabled survey infrastructure for which we build a first pilot is a hybrid panel. A hybrid panel is a longitudinal AI-enabled survey which allows to iteratively improve the alignment between large language models (LLMs) and the population they aim to simulate and use the errors to inform the design and implementation of the next survey wave (e.g., inform the participant recruitment, assignment of questions to participants). It incorporates both human participants and LLMs as fundamental elements of its design. In this research note, we introduce the concept of a hybrid panel by providing a definition and outlining an overarching framework, spanning data collection to data validation. We detail results from a first pilot study to illustrate (open) challenges that we identify for hybrid panels.

Multi-agent large language model (LLM) systems are designed to improve reasoning by decomposing tasks across multiple agents with specialized functions, but the presence of multiple agents does not inherently guarantee coherent reasoning or outputs that align with task objectives. This paper introduces a quantitative ethnographic (QE) approach for diagnosing and redesigning multi-agent LLM systems based on the discourse produced through agent interactions. We test this approach using automated essay scoring as an example context, applying Epistemic Network Analysis (ENA) to model a five-agent multi-agent debate system and examine differences between debates that produced correct versus incorrect scoring decisions. Results show that, in the initial system, correct scoring decisions were characterized by rubric-grounded justification, agreement, and elaboration. Incorrect scoring decisions, in contrast, were characterized by extended proposition-challenge-response exchanges that were less consistently tied to rubric criteria. We then used the findings to revise the agents' prompts. The revised system improved exact scoring accuracy from 27.78% to 40.28% and shifted the discourse of incorrect debates toward the rubric-grounded pattern of correct ones, making the two nearly indistinguishable. Based on these results, we argue that QE can support a diagnostic-to-redesign loop for AI reasoning by tracing how patterns of agent interaction relate to system performance, informing prompt redesign, and evaluating whether those redesigns change both outcomes and interaction patterns.

Rubrics aim to make language-model evaluation transparent by decomposing response quality into interpretable criteria. However, natural-language rubrics are often ambiguous, require LLM judges, and typically assume criteria aggregated through linear weighted sums, limiting their ability to capture dependencies, alternatives, penalties, and override conditions. We propose ExecRubrics, a framework for representing rubrics as compact executable programs. ExecRubrics encodes evaluation logic as verifiable Python scoring functions, giving natural-language rubric intent an operational semantics: a fixed decision procedure that can be inspected, executed, and edited. On three long-form response benchmarks – HealthBench, HelpSteer, and ArgQuality – we show that ExecRubrics can recover substantial preference signal without an LLM judge at evaluation time. On ArgQuality and HelpSteer, the strongest executable variants are within 1.1 and 4 percentage points, respectively, of the direct GPT-5.5 agentic baseline. Executable rubrics are also considerably faster, achieving a 192x average speedup. We show that incorporating external logic and resources from text processing libraries such as NLTK and spaCy can further improve preference accuracy. Our results suggest a novel way of approaching automated evaluation, by offering a faster, more explainable, and less ambiguous alternative to black-box rubric evals, particularly in high-stakes domains such as healthcare and banking where precision and auditability are critical.

Existing LLM unlearning methods struggle with robustness: unbounded forget losses degrade model coherence, fixed-weight balancing cannot adapt as retain difficulty shifts mid-training, and methods that work on one benchmark falter under scaling or repeated application. We propose BLADE, a constrained bilevel framework whose three mechanisms give smooth, predictable control over the optimization landscape: a clamped-entropy forget loss whose gradient is exactly zero once a token reaches sufficient uncertainty; an asymmetric augmented Lagrangian that permanently ratchets retain protection after any violation; and a bilevel structure confined to LoRA adapters that repairs retain damage before each forgetting step. BLADE dominates across three benchmark families, improving average composite scores over the strongest baselines by $6$% on TOFU, $9$% on MUSE Books, and $7$% on KnowUndo, and it remains stable under $4\times$ scaling and $4$ sequential unlearning steps on MUSE News where the best competing method collapses entirely.

A long-video answer is evidence-supported only when the frames decoded from the video cover every event the answer depends on. Existing evaluations score final-answer correctness or predicted evidence intervals, but the frames a method decodes before answering are rarely audited, so correct answers can still rest on incomplete observation. We introduce VES-Bench, a 600-question benchmark of Temporal Ordering and Event Counting items over 348 public long videos. Each item carries a jointly necessary set of evidence intervals, letting us audit at three strictness levels whether a method's decoded frames cover every one of them. We also propose TRACE, a training-free agent that grounds answers in raw visual clips, builds an evidence bundle round by round, and stops only when the answer stabilises as the bundle grows and a final pass over the same clips returns the same answer. Under a same-backbone audit, TRACE answers 50.7% of questions correctly with at least two decoded frames inside every evidence interval, at 98.7 frames per question: over 10 points above uniform decoding at 128 frames (40.2%), and within 2.6 points of uniform decoding at 256 frames at 0.39x its frame cost, while reaching the highest answer accuracy in the audit (63.5%). TRACE also stays competitive on Video-MME (86.1), LVBench (75.6), and LongVideoBench (75.1).

Claim-level Uncertainty Quantification (UQ) aims to mitigate the lack of reliability of Large Language Models (LLMs) by evaluating the factuality of each claim in their outputs. We introduce Kernel Token Contradiction (KTC), a lightweight approach to compute claim-level UQ under realistic white-box conditions. KTC represents the candidate tokens involved in LLM generation as a positive semi-definite kernel that integrates both the LLM's conditional distribution and a token contradiction score. We then use the Von Neumann entropy to quantify the uncertainty of this kernel. To estimate token contradiction, we develop a new approach based on frequency statistics from the Wikipedia corpus. Although CPU-only, our approach achieves over an 8.2x speedup compared to state-of-the-art GPU-accelerated methods based on cross-encoders, and over a 65x speedup compared to CPU-only methods with comparable performance. Our evaluation spans two benchmarks across four European languages and 16 different models. KTC not only matches the average performance of existing methods but also outperforms them in high-precision regimes. This combination of computational efficiency and accuracy makes real-time monitoring of LLM outputs practical in production.

Safety alignment helps models adhere to human values, but it often reduces response utility. We ask a critical but understudied question: Does safety alignment impose the cost equally across language groups? To answer this, we introduce a rigorous protocol to measure the utility loss imposed solely by safety alignment, which we term Safety Cost. Through direct pairwise comparisons between safety-aligned models and their unaligned counterparts, we find a systematic inequity: non-English users consistently bear a higher Safety Cost than English users. We further identify three underlying patterns. First, multiple languages lie in a double-penalty zone, experiencing both weaker safety protection and larger utility loss. Second, certain languages exhibit apparent utility gains that are in fact a consequence of safety filters failing to engage. Third, even high-resource languages pay a larger Safety Cost than English to reach the same level of safety. We show that these disparities arise from both explicit refusals and implicit qualitative differences across multiple dimensions. By accurately measuring the disparate effects of safety alignment, our findings expose a systematic disparity in current safety alignment practices.

Large Language Models (LLMs) increasingly support decision-making in high-stakes domains, but they often hallucinate and express confidence that is misaligned with factual correctness. Response-level confidence is a coarse signal: a single generation can mix correct and incorrect statements, so a single number is not actionable for users that must accept, reject, or verify individual pieces of information. We study claim-level confidence calibration as a decision-relevant uncertainty signal: each response is decomposed into atomic, verifiable claims, and each claim is assigned a calibrated confidence using inference-time signals from consistency across samples and self-verification. Our framework operates in closed-box settings (no logits, no fine-tuning) and applies post-hoc calibration directly at the claim level, enabling selective intervention such as evidence retrieval or human review for low-confidence claims. Across TriviaQA and TruthfulQA we evaluate seven baselines on six recent models (Llama-3.1, Mistral, Qwen2.5, DeepSeek-R1, GPT-4, GPT-4o), and show that claim-level decomposition combined with post-hoc calibration reduces expected calibration error on factual questions while exposing failure modes on adversarial false-premise questions where decision-makers most need reliable uncertainty estimates.

Retrieval-augmented generation (RAG) enables LLMs to access external knowledge for answering knowledge-intensive questions. For complex multi-hop questions, multi-turn retrieval-augmented reasoning extends RAG into an iterative process that repeatedly searches for and integrates evidence across documents. However, existing reinforcement-learning (RL) approaches for agentic RAG are typically optimized with final-answer rewards, which provide sparse supervision and overlook whether the model actually retrieves the required evidence chain. We present GTA-RAG, a graph-trajectory-augmented RL framework for multi-turn retrieval-augmented reasoning. From an entity–document graph, we sample connected document paths, synthesize multi-hop QA trajectories, and validate them with the deployed retriever to obtain executable trajectory-level supervision. We then optimize the retrieval policy with Group Relative Policy Optimization (GRPO) and a trajectory-guided reward that encourages both accurate answers and acquisition of target evidence documents, followed by answer-reward training on natural QA instances. Experiments on three multi-hop and two simple QA benchmarks show that \method{} consistently outperforms RL-based RAG baselines with both Qwen2.5-3B and Qwen2.5-7B backbones, while substantially improving evidence-chain coverage. Our code is available at https://github.com/cjcj46262/GTA-RAG.

Function calling represents the core capability of agentic large language models (LLMs). Existing research has focused on enhancing LLMs function-calling accuracy through fine-tuning, reinforcement learning (RL), and multi-agent frameworks, particularly for native function-calling LLMs. This work demonstrates that LLMs achieve superior accuracy in function calling in instruction-following contexts (i.e., standard user-assistant interactions) rather than a tool calling context. We introduce Instruction-Followed Function Calling (IFFC), a novel framework that decouples function-calling logic from the primary LLM and delegates it to a dedicated smaller model operating within the instruction-following paradigm. Our method consistently outperforms both native function calling (NFC) and prompt-based function calling (PFC) baselines, with particularly strong gains on reasoning-oriented LLMs. Furthermore, we demonstrate that IFFC maintains robust performance under aggressive quantization, enabling efficient on-device deployment without significant accuracy degradation. This work establishes a new paradigm for reliable, resource-efficient function calling in edge-computing scenarios.

Surprisal, the negative log-probability a language model assigns to a word given its preceding context, reliably predicts adult reading times. Does it contribute as much to explaining when children acquire individual words? Frequency reflects a learner's cumulative exposure to a word, whereas surprisal reflects how predictable a single occurrence is given its context. We investigate this question across two corpus-based studies of Spanish. In Study 1, we modeled age of acquisition (AoA) for 225 Spanish nouns using lexical frequency and contextual diversity from child-directed speech, plus surprisal from three language models differing in architecture and training language (BETO, BERTIN, mGPT). Frequency strongly predicted AoA (r=-.597, p<.001); surprisal added little beyond frequency and word length, including in a naturalistic-context analysis. In Study 2, we modeled adult fixation durations in the Chilean Spanish subsample of the Multilingual Eye-movement Corpus (MECO Wave 2), using mGPT surprisal alongside two independent frequency measures. Surprisal robustly predicted longer fixation durations after controlling for frequency and word length, consistent across both frequency sources. A matched word-type-level comparison showed the surprisal-behavior association was stronger in reading than in acquisition (z=3.63, p<.001). The findings suggest cumulative lexical exposure and contextual predictability play different roles across the language trajectory: frequency is particularly informative about when early lexical representations are acquired, whereas surprisal captures moment-to-moment processing difficulty in an already-established linguistic system. We discuss this pattern in relation to usage-based and entrenchment-based accounts of lexical development and to the evaluation of language models as models of human language behavior.

Metaphors are figurative use of words for conceptual mapping. Metaphor detection in the legal context has been crucial as metaphors are persuasive juridical means of creating legal meaning and concepts resulting in significant consequences. Metaphorical framing in legal discourse by judges, lawyers, and legislators brings about real-time implications upon individuals and influences judicial decision-making, argumentation and interpretation of laws. This is crucial in Human Rights infringement cases where language determines severity of punishment, public perception and judicial outcomes. While automatic metaphor detection in major languages like English, Spanish, Polish, Lithuanian have aided in understanding inherent intentions of metaphorical use of language, there is no such attempt in low-resource languages like Hindi. The dearth of annotated legal corpora in Hindi makes it difficult to develop NLP models and detect metaphors in judicial proceedings. In the Indian context, Convolutional Neural Networks (CNNs) have been used for classification of bail judgements, however there are no existing models designed for metaphor detection. We present a Hindi Legal Metaphor Corpus (HiLeMe) by isolating judgements from Hindi Legal Data Corpus (HLDC). Legal experts annotated HiLeMe to classify metaphorical constructions using the MIPVU schema. We downstreamed an mBERT on Hindi legal metaphor detection task. We built a transformer-based architecture for metaphor detection that are known to outperform traditional models in legal classification tasks. This model provides insights into the judicial psyche for decoding judicial decisions. Our research contributes to advancing automated models in legal discourse in low-resource languages like Hindi and envisages adoption into 22 Indian schedule languages.

The unit of AI safety evaluation is still the individual model, yet language-model agents are increasingly deployed in interacting populations that read and write one another's decisions. This raises a question no single-agent audit can answer: an agent that is well-calibrated on its own may still be pulled toward a different decision by the agents around it. We study this on a security-triage task, where populations of language-model monitors decide whether to escalate or dismiss alerts, and into which we can inject a committed minority that always pushes one way. We find that two alerts a single agent judges almost identically on its own can drive collective behavior far apart, so auditing any one member need not reveal what the population will do. Yet that collective behavior can be predicted in advance. From a population's benign, adversary-free operation alone, we calibrate a response function that forecasts, before any attack is run, how far a committed minority will later move it. We then ask what shifts the outcome and find that letting agents see each other's reasoning neutralizes a weak attack, while only delaying it against a strong one, turning the question from whether the population converges on the adversaries' choice into when. Finally, we exclude the hypothesis of capture being an irreversible trap: once the committed agents are removed, the population drifts back toward where it began, so capture is a temporary state. Alignment in isolation is not alignment in a population, yet what a population will do under attack can be read in advance, from how it behaves before any adversary arrives.

Multilingual LLM judges produce different evaluator-backbone rankings depending on the prompt language: on an eight-language Agent-as-a-Judge benchmark, the top-ranked backbone alternates across English, Arabic, Chinese, Hindi, Japanese, Spanish, Turkish, and Swahili, and 7 of 15 backbone pairs show statistically significant pairwise rank reversal. We treat this as a measurement problem. The multilingual judge score decomposes additively into task difficulty, backbone skill, and a language-backbone interaction term, the last of which is recoverable without human labels by double-centering the cell-mean score matrix. We make this estimator (Consensus-Based Calibration, CBC) explicit, give an $O(1/\sqrt{n})$ finite-sample concentration bound with variance constant $(1-\tfrac{1}{m})(1-\tfrac{1}{k})$, and show that it is unbiased even when task-language interactions are present. Across 7{,}920 judge runs (6 backbones, 8 languages, 55 tasks, 3 frameworks), CBC raises held-out cross-task rank consistency $\tau$ from 0.650 to 0.902 and agrees with the held-out additive-model oracle in 100% of per-language decisions versus 68.5% raw; these are consistency diagnostics, not human-grounded correctness measures. On a separately collected M-RewardBench panel (7 languages, 1{,}500 items per language, 10{,}500 language-item instances, 5 evaluators), panel agreement with the public human gold preferences rises from 68.7% to 76.6% (gain 7.9 percentage points, 95% CI $[6.0, 9.9]$), our strongest external evidence of downstream usefulness. The estimator is the standard two-way ANOVA interaction-recovery operation under sum-to-zero contrasts; our contribution is its application as a label-free post-hoc calibrator for multilingual LLM judges, an explicit finite-sample concentration bound, and an unbiasedness result that holds even under task-language misspecification.

Large language models are increasingly consulted at moments of distress, yet single-turn benchmarks neither test sustained exchanges nor distinguish between users. We built a personality-aware evaluation in which four widely used models advised several synthetic help-seekers, each given a psychometrically specified profile, in an acute crisis: a caregiver learning of a relative's dementia diagnosis. Auditors blind to the profile prompt recovered the specified bands from dialogue alone with high agreement on every instrument (ICC(2,4) = 0.91; 0.79-0.96 by instrument; band-score r = 0.78), as expected for the Big Five but equally for coping style, coping self-efficacy, resilience and reactance, which the lexical approach never covered. Such evaluation therefore reaches beyond the Five Factor Model to motivational, regulatory and self-appraisal dispositions. The four models were not distinguishable on emotion stabilisation and failed alike, sharing three modes: verbosity, a talk-to-listen ratio above one, and problem-solving before the situation had been explored.

Large language models (LLMs) are increasingly used to support text analysis in qualitative research, yet evidence on their performance in inductive content analysis remains limited. This study compares human and LLM-based inductive coding of open-ended survey responses from 903 answers across six variables from a European PhD student survey. Five human coders performed inductive content analysis following a standardized coding scheme, while an LLM (GPT-5.4) conducted the same task using an established prompting procedure. Agreement between human and LLM outputs was assessed using the Adjusted Rand Index (ARI). Results showed an alignment between humans and the LLM, with ARI values of 0.61 for coding and 0.54 for theme generation. These values were close to the internal consistency of coding and theme results within humans (ARI = 0.68) and the LLM (ARI = 0.76). Agreement varied widely across variables, with low within-entity consistency consistently linked to low between-entity agreement, underscoring the role of data characteristics and individual performance in reliability. Overall, the findings suggest that LLMs can approximate human coding in this case-specific setting, particularly at the coding level, and may serve as a scalable support tool for inductive qualitative analysis.

Social interaction increasingly takes place in multicultural settings, where individuals may draw on multiple cultural influences and adapt their communicative behavior across contexts. Despite recent advances in equipping Large Language Models (LLMs) with social understanding capabilities, existing approaches often model culture as a static demographic attribute, limiting their ability to accommodate hybrid and dynamically expressed communicative preferences. Therefore, in this paper, we propose DyCAC, a training-free framework that achieves fluid social alignment by incorporating \underline{Dy}namic \underline{C}ultural \underline{A}daptation with continuous \underline{C}ognitive tracking. Rather than inferring a fixed cultural identity, DyCAC models culturally relevant communicative preferences as a time-varying mixture of population-level cultural reference profiles. This reference-based representation is further calibrated using dialogue-style signals observed in the ongoing interaction, enabling the model to capture both composite cultural influences and turn-level shifts in communicative behavior. In parallel, a memory module driven by Theory of Mind (ToM) continuously tracks the cognitive states of the interlocutor. Extensive experiments on interactive social and cultural benchmarks demonstrate the superiority of our approach. The proposed framework outperforms existing baselines, exhibiting enhanced social intelligence and broad adaptability across varied multicultural contexts.

Large language models (LLMs) have demonstrated strong potential in graphical user interface (GUI) generation, but reliable evaluation remains challenging due to uncontrolled data distributions, noisy annotations, and limited layout scenario coverage. To address this, we propose SchemaGUI, a template-based benchmark for controllable GUI generation evaluation. By synthesizing paired natural language instructions and deterministic function-call references from parameterized interface schemas, SchemaGUI can generate thousands of deterministically annotated tasks in seconds without human labeling. Based on 1,000 evaluated instances per scenario and language across six representative bilingual scenarios, we benchmark five mainstream models, including the Qwen3.5 family, Qwen3-Coder-30B, and DeepSeek-R1. Our extensive analysis reveals three key insights. First, precise geometric spatial control remains an important bottleneck; while scaling Qwen3.5 from 4B to 27B improves Schema Feasibility from 91.56% to 99.63%, the Geometry score improves more modestly (from 67.05% to 75.30%). Second, generation difficulty is highly sensitive to layout complexity, with current LLMs excelling at simple sequential arrangements but suffering severe coordinate drift in dense grids and multi-region compositions. Third, thinking mode increases token consumption while generally reducing GUI Score, particularly for smaller models.

Propaganda detection includes several related prediction levels, ranging from sentence-level decisions to technique classification and span identification. However, it remains unclear how supervision at these levels interacts when learned jointly across Arabic and English. We present ProBel, an Arabic and English resource that aligns binary labels, multi-label annotations over 23 propaganda techniques grouped into six coarse categories, technique-labeled spans, and reference explanations for the same news sentences. It includes a substantially larger English collection and supports matched binary, coarse-grained, multi-label, and span-level tasks in both languages. We evaluate zero-shot prompting, task-specific fine-tuning, and joint training under a shared setup. A single bilingual multi-task model achieves the best overall performance and remains competitive across tasks and languages. Cross-task analysis shows that transfer depends on the supervision level. Joint classification training preserves binary performance, whereas span-only training can weaken sentence-level prediction. Joint bilingual training yields the most stable results, while monolingual fine-tuning can reduce transfer to the other language. We will release the data, code, and evaluation scripts.

Scientific papers may relate by problem, method, result, or contribution, but document-level retrievers collapse these into a single similarity score without saying why they are related. Citation- and similarity-based retrieval alone also confines search to the neighbourhood of what is already known, whereas generative retrieval generates document identifiers directly, enabling the exploratory retrieval that scientific discovery depends on. We connect papers in a graph whose edges are typed by these four facets, derived from facet items and citation signals, and distil it into a generative retriever whose identifiers are the papers' own facet text. Two graph properties do not survive naive distillation. First, because every training pair is an edge, naive enumeration indexes just 84% of the corpus. Coverage-aware distillation makes every paper learnable through a reverse-neighbour fallback, a minimum-coverage threshold, and edge-importance weighting. Second, constrained decoding guarantees that every generated identifier is a valid paper, but not that the graph connects it to the query. Graph-weighted reciprocal rank fusion scales each candidate's rank term by its query-candidate edge weight, dropping unsupported ones. On LitWeave, our constructed corpus of 11,359 NLP papers, Graft recovers 91% of its graph teacher's Recall@20 with no nearest-neighbour index or encoder at inference, and outperforms the graph teacher on query papers outside the corpus. It reproduces the graph's own facet labels at 0.922 precision, so every returned paper arrives labelled with the facet that surfaced it rather than an opaque score.

Large language models generate tokens sequentially, but can they execute multiple tasks concurrently while forming each token? Broader attention allocation may provide a mechanism for such task concurrency. Existing approaches to scaling inference primarily rely on longer generations, more samples, or additional verification stages, while attention dispersion is often treated as a signal of interference or error. Task concurrency within serial generation therefore remains underexplored. We propose the Model Hyper-Threading Hypothesis and evaluate its predictions using multiple coordinated tasks that share state within the same problem. We design three conditions (Baseline, Serial Functional Scheduling, and Concurrent Functional Loading) and evaluate their benefits and costs using accuracy, output-token distributions, and attention metrics. On an AIME 2025 development set, Concurrent Functional Loading achieves the highest accuracy. Relative to Serial Functional Scheduling, its typical output length is similar and it is shorter on most problems, while exhibiting greater attention dispersion and higher task-relevant coverage, albeit with a heavier output-length tail. Within-step concurrency and its causal mechanism still require direct tests. Our results show that more dispersed attention can coexist with higher accuracy, providing preliminary behavioral and correlational evidence for the hyper-threading hypothesis. These findings motivate a shift in perspective on inference scaling from "generating more tokens" toward "having each generation step carry more tasks," pointing to a new avenue for improving reasoning performance.

Diffusion multimodal large language models (dMLLMs) frequently produce long-form outputs marred by semantic drift and repetition, with quality generally degrading as output length increases. We identify two structural deficiencies in existing decoding methods as primary drivers of these failures: confidence-based scoring ignores decoded-neighbor support, and block partitioning prevents access to high-readiness semantic anchors, together causing tokens to be committed before their local context is sufficiently established. We propose \ours{} (Context-Aware Cluster Decoding), a training-free decoding method that scores each masked position by a multiplicative composite of softmax confidence and neighbor proximity, promoting contextually ready tokens above isolated candidates while suppressing low-confidence positional noise, operating block-free to keep high-readiness anchors globally accessible. \ours{} further applies architecture-aware calibration to handle confidence heterogeneity induced by diverse visual integration strategies. Experiments on three dMLLMs across four benchmarks demonstrate consistent quality gains and hallucination reduction over Original, with larger gains in several longer generation settings, highlighting the importance of neighbor support and visual integration strategy for future dMLLM decoding method design. Our code is openly available at https://github.com/zhaoyk-sysu/CACD-dMLLM.

A cognitive architecture is more than the module that reasons: it must also decide how long to think and what deserves the effort. We built a minimal but complete system - a recurrent reasoner with adaptive halting, a homeostatic control field, and a value module - and asked of each part: does this function emerge from gradient descent, or must it be computed? Competence emerges. Stopping appears to emerge too, and to be worth more than everything decidable in advance, but that appearance is instrumentation: payoff at matched mean compute climbs from 0.467 (uniform) through 0.546 (difficulty) to 0.698 (ex-ante value), and the further climb to 0.921 (posterior self-observation) does not survive audit. PonderNet-style halting returns a halting-weighted mixture of hidden states while forced-depth baselines return one, and the language head is trained on the mixture alone; equalizing the readout annihilates the apparent advantage of native execution (residual +0.000 [0.000, 0.000]). Value does not emerge: trained couplings capture zero of a payoff an explicit allocator captures completely (+0.151, routing correlation +0.79), so the second-order decisions that pay must be computed, at least where value is orthogonal to content, as here by construction. On a frozen LLM actuator the same instruments show self-consistency voting to be a measured bound (+0.0236 [+0.0150, +0.0326]) and inter-sample agreement nearly worthless as a stopping signal, its mass concentrating on wrong answers. Every null we assert carries a mechanism and a positive control, and the protocol is part of the contribution. Executing our own falsifiable prediction, value under commitment pays +0.1312 [+0.1124, +0.1502] in a cliff-cost family, some seven times the smooth-family estimate - not because the cliff shifts information ex ante, but because it multiplies the attainable range fivefold (5.1x [3.4, 8.2]).

Extracting skills from past successes is critical for the efficient evolution of Large Language Model (LLM) agents. Prevailing agent self-evolution paradigms typically rely on a core assumption: equipping LLMs with skill memories derived from successful trajectories will monotonically improve their problem-solving capabilities. However, probe analyses reveal that extracting skills solely from successful trajectories traps the model in a Skill Imitation Trap. For tasks that resemble past successes but require different tools, retrieving more skills paradoxically increases the model's confidence in wrong tool calls—procedure skills raise the wrong-tool margin by $47\%$ over a memory-free baseline. To overcome this limitation, we propose Boundary-Aware Skill Memory (BASM), which augments each skill with explicit boundary fields—applicability conditions, risk cues, avoidance rules, and recovery notes. These fields transform each retrieved skill from an unconditional action template into state-conditioned guidance: the agent applies the skill when its conditions hold, suppresses inapplicable tool calls when they do not, and issues targeted repairs when execution fails. Across three agent benchmarks and four model scales, BASM consistently outperforms success-distilled skill-memory baselines: it improves task success rate by up to $23.8\%$ on AppWorld, accuracy by up to $5.0\%$ on BFCL, and reduces attack success rate by $4.6\%$ on AgentDojo, while simultaneously reducing average AppWorld steps by up to $6.6\%$ relative to the memory-free baseline.

Bengali is the seventh-most-spoken language globally, yet LLM safety evaluation remains overwhelmingly English-centric. We introduce BanglaSafe, a benchmark of 879 Bengali prompts combining 309 natively authored prompts with 570 expert-reviewed prompts, spanning 17 culturally grounded harm categories and five prompting conditions that vary language, writing style, and authority framing. Evaluating 18 frontier LLMs, we find that over half of all responses are unsafe or partially unsafe (53.6%) while 14.7% contains strictly harmful content, and that the strongest observed effect is not the switch from English to Bengali but the choice of writing style within Bengali: the same harmful request phrased as a formal newspaper investigation succeeds 17 percentage points more often than the same request phrased as a casual message, with no adversarial engineering involved. We further show that existing safety classifiers struggle to reliably evaluate Bengali content, with even frontier models failing on nearly half of all cases.

Large Language Models (LLMs) demonstrate remarkable problem-solving capabilities when guided by Chain-of-Thought (CoT) prompting, yet the internal mechanisms underlying these improvements remain poorly understood. In this work, we investigate where CoT-related causal effects emerge across the generated reasoning trajectory and which attention heads carry signals that contribute to final-answer computation. Because CoT reasoning unfolds over multiple generated tokens, standard activation patching at a single static token position is insufficient to characterize these temporally distributed effects. To address this limitation, we introduce a sequential activation patching framework that traces CoT-conditioned attention-head activations across token positions and aggregates their effects using Part-of-Speech-guided analysis. We further introduce Sequential Multi-Head Patching to evaluate the joint contribution of distributed head sets, together with cross-question and random activation controls. Targeted zero-ablation experiments show that the identified heads are functionally important for successful answer generation and affect several overlapping mechanisms, including reasoning-trajectory maintenance, answer anchoring, exemplar-target separation, and numerical generation. Overall, our results provide evidence for distributed reasoning-support sub-circuits associated with CoT-conditioned computation.

We audit measurement variability for 3 native tool-calling endpoints across 2 providers on the official BFCL multiple and parallel categories, using matched AST grading. At temperature 0, reruns are nearly deterministic across Groq endpoints and a thinking-enabled Gemini setting: ever-flip fractions are 0.7%, 2.0%, and 2.7%, with mean run correlations of 0.997, 0.966, and 0.961. Semantics-preserving prompt perturbations create the larger floor on all endpoints, with median perturbation paired SDs 11x to 58x larger than rerun paired SDs. The failure character also shifts: malformed-output failures account for 30%, 7%, and <1% of task failures, so marginal accuracy hides not only stability but also failure mode.

Multimodal time-series forecasting has emerged as a promising paradigm in which natural-language context is expected to improve predictive performance. Recent multimodal foundation models, including Aurora, as well as early- and late-fusion approaches such as MM-TSFlib and TaTS, report substantial gains over unimodal baselines on the Time-MMD benchmark, attributing these improvements to textual information. However, whether these models are actually sensitive to the semantic content of the text remains unverified. We address this question through controlled text perturbations, attribution analyses, and probes of Aurora's text pathway. On Time-MMD, swapping each row's text for any other real text (empty, constant, within-domain shuffled, or cross-domain) moves mean MSE by less than $0.5\%$ on all three architectures. The improvement reported in the literature is recovered when a co-shipped numeric column is removed without touching text. We conclude that, on this benchmark and within this family of frozen-encoder architectures, text content is not the operative signal behind the reported gains. To support future work on text integration in multimodal foundation models for structured data, we release our perturbation protocol and evaluation harness as a reusable diagnostic toolkit.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language interaction, yet their safety alignment remains vulnerable to jailbreak attacks. A key challenge is that safety behavior learned in the textual space does not reliably transfer to fused cross-modal representations, leaving multimodal inputs exploitable through latent semantic cues. We propose Text-Anchored Semantic Perturbation Attack (TA-SPA), a black-box jailbreak framework that optimizes transferable perturbations in a text-anchored semantic space. TA-SPA integrates Text-Anchored Semantic Factorization (TASF), which encourages the separation of cross-modal semantic factors from modality-specific residuals, with Semantic-Preserving Augmentation (SPA), which diversifies harmful target anchors while preserving semantic consistency. Experiments show strong attack effectiveness and transfer to commercial MLLMs, with competitive performance under representative defenses. Additional controls and probing support the intended factorization without implying perfect disentanglement, motivating representation-level safety alignment beyond input-level filtering.

Evaluation of large language models (LLMs) increasingly requires predicting how a model will perform on new questions or tasks before collecting large amounts of new annotations. This problem is challenging because question difficulty, scenario, and underlying capability demands can vary substantially. Simple retrospective averages may confound model ability with item characteristics. In this paper, we study a model-based evaluation framework that combines multidimensional item response theory model with question contexts to predict LLM performance on unseen questions. The framework represents LLMs through latent capability profiles while using question content to inform item characteristics, allowing information to transfer beyond previously observed items. Empirically, we find that for within-scenario evaluation, incorporating question embeddings improves prediction relative to model-free baselines, and that multidimensional latent structure provides a richer description of capability variation than unidimensional alternatives. At the same time, our results reveal an important limitation that the generalizability does not necessarily translate into reliable prediction under cross-scenario shift. These findings suggest that context-aware psychometric modeling is a promising direction for efficient and interpretable LLM evaluation, while also highlighting cross-scenario generalization as a central open challenge.