LLM 理论进展

Stakeholder interviews are an important source of information for requirements elicitation, yet many relevant requirements remain implicit in such conversations. Stakeholders frequently describe workflows, challenges, and operational practices without explicitly articulating the software capabilities that could address them. Recent work has considered the use of LLMs to analyze conversational data and extract requirements from stakeholder interviews. Existing approaches, however, primarily focus on identifying explicitly stated requirements, leaving implicit opportunities largely unexplored. In this paper, we present LENS (LLM-Enabled Needs Discovery from Stakeholder Interviews), an approach that analyzes stakeholder interview transcripts to both extract explicit requirements and infer additional latent requirements. LENS performs this inference by reasoning over stakeholder statements together with contextual information about organizational tools and infrastructure. Both extracted and inferred requirements are represented as user stories and linked to transcript excerpts to ensure traceability. We conduct a preliminary evaluation of LENS using twelve stakeholder interview transcripts collected in an industrial setting involving cybersecurity operations. We show that LENS achieves an average F1-score of 84.4% for extracting explicit requirements, while, on average, 75% of the latent requirements identified by LENS were perceived as providing useful automation or time-saving potential by domain experts.

Was this person ever at that place, and if so, when? Answering such questions from noisy, multilingual historical documents is the central challenge of HIPE-2026, the third edition of the HIPE evaluation series. Moving from named entity recognition and linking (HIPE-2020, HIPE-2022) to reasoning about relationships between entities, HIPE-2026 targets two temporally grounded relation types: $at$, indicating that a person was present at a location at some point prior to a document's publication date, and $isAt$, indicating presence contemporaneous with that date. This paper presents the results of the evaluation campaign, which confronted 17 participating teams with the challenges of historical language variation, OCR noise, and indirect contextual cues across three languages: French, German, and English. The datasets include historical newspaper text from the nineteenth and twentieth centuries, as well as a surprise-domain generalization set drawn from early modern French literary texts. A distinctive feature of HIPE-2026 is its three-fold evaluation framework, which assesses predictive accuracy, computational efficiency, and cross-domain generalization, reflecting the practical demands of large-scale historical document processing in the cultural heritage domain. Across more than 40 submitted runs, results reveal a wide range of strategies, from state-of-the-art large language models to lightweight task-specific classifiers, and highlight the trade-offs between accuracy, efficiency, and robustness inherent to historical relation extraction at corpus scale. System descriptions, datasets, and findings are presented and discussed, offering a detailed picture of the current state of temporally grounded relation extraction for historical documents.

Large language models (LLMs) attain remarkable surface fluency on code, yet they neither formally guarantee the syntactic validity of their output nor leverage the hierarchical structure defining the target language. While existing constrained-decoding frameworks address the former, they operate under rigid assumptions that preclude critical lexical mechanisms — including context-sensitive lexing, maximal-munch tokenization, and keyword extraction — and only approximate vocabulary masking, sacrificing completeness. For the latter, code LLMs typically inject grammatical structure via predetermined policies rather than learning which structural information to expose. In this work, we introduce Weave of Formal Thought (WoFT), a paradigm uniting rigorous syntactic validation with learned structural representations. First, we present a formal engine and constrained decoder that is sound and complete with respect to the full Tree-sitter specification. By augmenting generalized LR (GLR) parsing with a speculative-lexing construction that maintains concurrent lexer-state hypotheses synchronized with a GLR graph-structured stack, our decoder admits every subword token extending to a valid program prefix and rejects all others. Second, we present a latent-variable fine-tuning method training the language model to interleave non-terminal grammar symbols directly into generation. Utilizing the reweighted wake-sleep (RWS) algorithm to optimize the importance-weighted evidence lower bound (IW-ELBO) of the surface text, the model learns to selectively retain formal derivations as an adaptive structural scratchpad. For Python, fine-tuning StarCoder2-3B with our RWS objective reduces per-token cross-entropy by 14.3% relative to a text-only SFT baseline, demonstrating that discretionary latent syntax recovers critical structural information that flat autoregressive training discards.

Safety evaluation of large language models (LLMs) is commonly performed by querying models with unsafe or jailbreak prompts and judging whether their outputs violate a safety policy. Although useful, output-level evaluation is expensive, sensitive to judge choice, and easily tied to fixed question banks. We propose **SafeVec**, a white-box evaluation procedure that measures safety from internal representations rather than generated answers. **SafeVec** first extracts layer-wise refusal directions from a safety-aligned reference model, then selects stable layer windows where safe and unsafe behaviors are separable, and finally scores a target model by measuring whether its hidden states align with these refusal directions under unsafe and jailbreak prompts. The resulting metric, **RAS** (**R**efusal **A**lignment **S**core), maps representation-level refusal alignment to a calibrated 0-100 safety score. Across `Llama`, `Gemma`, and `Qwen` model families, RAS separates aligned models from uncensored and abliterated variants, tracks output-level attack success rate, and is substantially faster than judge-based evaluation. These results suggest that refusal alignment provides a compact and efficient signal for white-box LLM safety evaluation.

Software vulnerability remediation is a cognitively demanding task that requires specialized security expertise often lacking in general developers. In the meantime, Large Language Models (LLMs) assisted tools show potential in vulnerability detection, location, and repair tasks. [Hypothesis:] While LLM-assistance is hypothesized to accelerate patching, it also risks introducing hallucinations or insecure code, leading to a higher likelihood of generating superficial repairs that bypass the standard functionality checks but fail the security validation. [Objective:] We aim to present an empirical experiment, unveiling the capability of LLM-assisted vulnerability patching compared to manual debugging on human participants in real-world scenarios. [Method:] We plan to conduct a controlled experiment using a Balanced Crossover design. For that, we have developed a WebApp for code execution and integrated hidden Ghost Tests to verify patch integrity beyond visible functional requirements. The experiment involves training and evaluation scenarios. The remediation speed, remediation efficacy for both standard functionality tests and security tests, and participant perception will be evaluated. [Pilot Study:] A pilot experiment with a small sample of participants has been conducted, providing insights for the following study.

Large Language Models (LLMs) are increasingly used in software engineering to generate and refine code. In practice, developers often continue from an initial code generation request with follow-up refinement instructions, such as requests to improve style, restructure implementation, or change the execution strategy while preserving the intended behaviour. However, existing benchmarks generally omit this multi-turn code refinement dialogue setting and therefore cannot evaluate whether LLMs maintain functional correctness, i.e., whether the refined code still passes the test suite for the original task. To address this limitation, we introduce CodeChat-Eval, an evaluation framework that constructs evaluation sessions from multi-turn code refinement dialogues using a dynamic instruction selection algorithm. Our empirical study on open-weight and proprietary LLMs observes a statistically significant decrease ranging from 19.2% (GPT-5 Nano) to 69.2% (Llama 3.1 8B) in functional correctness over multi-turn refinement. The largest correctness drops are associated with logic-level refinements and additive change requests. These findings indicate that LLMs struggle to maintain functional correctness during multi-turn code refinement dialogues, and highlight the need for benchmarks that evaluate functionality-preserving refinement beyond single-turn generation.

Standard benchmarks for multimodal large language models (MLLMs) score each item on one canonical ordering and miss whether order-irrelevant shuffling changes the answer, a baseline reliability property called for by emerging AI evaluation guidelines. We introduce Facet-Probe, a five-facet audit (option, evidence-chunk, document-rank, image-set, and mixed-modality ordering) of 18 frontier and open-weight MLLMs. A Bayesian item-response model separates ordering noise from per-facet bias, and a same-ordering control estimates the decoder-stochastic floor for observed flips. We find that none of the 18 MLLMs we audit are order-invariant: screened per-facet panel-mean flip rates span 24-50%. A Gemini same-ordering control at temperature 0 estimates a substantial ordering excess over a same-input decoder-noise floor in verified cells. Capability predicts but does not eliminate flips; the best model still flips on 13.4% of trials. In our Gemini mitigation tests, training-free prompt changes are modality-conditional and do not transfer from text to visual reasoning. These results suggest that prompt-level mitigation alone is unlikely to provide general order robustness, motivating future work on training-time and architectural approaches. We propose cross-ordering flip rate as a standard reporting axis for MLLMs.

When a large language model (LLM) answers a question about a company, it grounds the answer in retrieved web sources, and those sources decide what the model says. Most analysis of AI brand visibility looks at the answer text. This study looks one step earlier, at the citations. We merge three Rankfor.AI datasets covering 128 brands across 12 home markets and 13 languages, and analyse 167,551 URL-grounded citations (189,974 total attribution rows). We classify each citation by domain and source type and measure where AI gets its brand information, by language and by market. Four patterns hold. First, AI grounds brand answers overwhelmingly in third-party sources: 85.7% of citations point to sites the brand does not own, against 14.3% owned. Second, the source base is concentrated and long-tailed: 80% of citations come from about 18% of domains, fitting a Zipf law (alpha = 0.86, R^2 = 0.983). Third, one reference site dominates almost everywhere: Wikipedia is the most-cited domain in 11 of 12 languages, the exception being Lithuanian, where the business daily vz.lt edges it (4.38%). Fourth, the source mix is market-specific at the margin: for 46 Polish national brands the most-cited domain is YouTube, and four HR and careers portals supply 637 citations against 297 for Polish Wikipedia, about twice as many.

Multi-modal Large Language Models (MLLMs) offer powerful reasoning for forensic tasks, yet existing approaches utilizing exogenous segmentation decoders often suffer from suboptimal localization. The reliance on stitched pipelines introduces information bottlenecks during backpropagation, which dilutes spatial signals and is limited by semantic priors of the segmentor. To address these limitations, we propose ForensicsTok, which reformulates image manipulation localization as an autoregressive sequence generation task. ForensicsTok directly generates spatially grounded token sequences, enabling precise mask prediction without intermediary supervision. Specifically, we introduce a Token Splatting Decoder (TSD) to map tokens to binary masks via codebook-aware code smoothing, which mitigates sharp gradients from deterministic detokenizers. Furthermore, to capture diverse tampering clues, we propose a Hierarchical Expert Fusion (HEF) module that injects multi-scale features from a forensic expert model. This unified architecture effectively compensates for the lack of forensic priors in standard MLLMs. Extensive experiments on six benchmarks show that ForensicsTok substantially improves over existing MLLM-based baselines and slightly improves over strong forensic expert baselines, while exhibiting stronger robustness to perturbations.

General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88%$ balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first $\sim20%$ of thinking, despite giving the appearance of deliberation at the text level ($\sim74%$ of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.

Jailbreak attacks reveal a persistent weakness in aligned Large Language Models: carefully crafted prompts can elicit policy-violating responses despite safety training. While most defenses operate at the prompt or output level, it remains unclear how harmful intent is encoded within the model's internal representations. We investigate this question by analyzing token-level predictive entropy trajectories across layers of a frozen LLM using the logit lens. We find that static aggregate statistics of prompt-level entropy (e.g., mean, variance) carry little discriminative signal, whereas features capturing how entropy evolves across token positions, such as monotonic rank-based trend scores, are substantially more informative. Importantly, this signal is not uniform across model depth: it is concentrated in intermediate layers and degrades at the final layer, indicating that jailbreak-relevant structure is most pronounced in mid-network representations rather than at the output head. Across multiple models (Llama, Qwen, Gemma) and adversarial benchmarks, these entropy dynamics provide architecture-consistent separation without additional training. Together, our findings show that jailbreak behavior is reflected in structured intermediate uncertainty dynamics, clarifying both which entropy-derived features encode harmful intent and where in the network that signal is most pronounced.

Understanding the features of large language models (LLMs) is a central goal of interpretability. LLMs are commonly assumed to use superposition to represent more features than they have dimensions. They may not only represent features in superposition but also perform computation in superposition. Theory predicts that computing in superposition requires error correction that privileges feature directions over generic ones, but this prediction has not been tested empirically. We propose an empirical test of error correction in LLMs based on activation perturbations. Perturbing residual-stream activations, we find that they are robust to small perturbations--forming activation plateaus consistent with error correction--but less robust along candidate feature directions ("pure" directions, constructed from contrastive prompt pairs) than along mixtures of two such directions, indicating that the pure directions are privileged. We quantify this privilegedness by modeling the perturbation effect as a function of the $L^p$-norm of its decomposition into feature components. For $p=2$ the response is a quadratic form with at most as many nonzero eigenvalues as the residual-stream dimension, which cannot privilege the many feature directions superposition requires. $p>2$ lifts this constraint and is consistent with feature-specific error correction. We find $p>2$ for contrastive, MELBO, and SAE-decoder directions, and $p\approx2$ for random and PCA directions (controls). These results replicate across Gemma-2-9B, Qwen3-1.7B, Llama-3.1-8B, Mistral-7B-v0.3, Aya-Expanse-8B, and Yi-1.5-9B. We further validate our method on a toy model of error correction with known ground-truth features, recovering $p>2$ for true feature directions, degrading toward $2$ as we rotate away from them.

Neural scaling laws describe how pre-training loss decays as power laws with training time, model size, and compute. This position paper argues that the exponents of these power laws are fixed by generic mechanisms: a one-third time scaling due to the strong nonlinearity of Softmax, an inverse width scaling due to representational superposition, and an inverse depth scaling due to ensemble averaging of Transformer layers. These mechanisms are robust to a wide range of data structures and architectural details, placing current large language models in a universality class with fixed exponents. The coefficients, however, are expected to be sensitive to data and architecture details, and directly determine practical quantities such as the optimal model shape and the compute-optimal frontier. We therefore argue that understanding the coefficients is the key to near-term performance improvements, and that a closer examination of the current universality class may reveal pathways to better universality classes.

Large language models have transformed code generation, raising concerns around authorship, assessment integrity, and software trust. SemEval-2026 Task 13 Subtask A operationalizes detection as binary classification over code snippets, with a particular emphasis on out-of-distribution (OOD) generalization across unseen programming languages and application domains. We propose a SALSA-style formulation, Single-pass Autoregressive LLM Structured Classification, that maps each class to a dedicated output token and trains the model to emit a single-token label in a structured response. Rather than engineering hand-crafted features or decision rules, this formulation delegates the authorship decision to the model. To improve OOD robustness, we combine balanced sampling across languages with parameter-efficient fine-tuning and conservative training (low learning rate, single epoch) to avoid overfitting to the training domain. Our best system achieves OOD $F_1 = 0.789$ on the official leaderboard, substantially outperforming the CodeBERT baseline ($F_1 = 0.305$).

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise — that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal — a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it — 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

Neural scaling laws for transformer language models predict smooth improvements in pretraining loss with increasing parameters, but downstream capabilities such as in-context learning are known to emerge abruptly past a certain model scale. In this paper, we show that emergent capabilities arise stochastically throughout training, with larger models acquiring them earlier on average. We demonstrate that the emergence of capabilities such as pattern completion and indirect object identification corresponds to the abrupt learning of task-relevant attention patterns. To isolate this phenomenon, we train transformer models on synthetic linear map and cellular automata datasets, and we show that the difficulty of learning attention patterns depends on context length and pattern sparsity. Moreover, scaling the number of attention heads improves learning efficiency on our synthetic tasks, while increasing the head dimension yields diminishing returns past a minimum capacity. We additionally investigate architectures with alternative attention mechanisms, showing that MLP-Mixer outperforms a transformer on linear map tasks with complex attention patterns. Our findings provide a mechanistic insight into emergence, showing that downstream capabilities arise abruptly due to the intrinsic difficulty of learning sparse attention patterns in transformer models.

Long video understanding remains a daunting challenge for Multimodal Large Language Models (MLLMs) due to the excessive computation and memory footprint. Thus, keyframe selection is often adopted to mitigate this shortcoming, which however still suffers from low flexibility and high noise due to its hard sampling principle. In this paper, we define video frame selection as a problem of Quasi-Gaussian Sampling, and propose an adaptive and training-free approach termed AdaQ. Inspired by the 3-$σ$ rule of Gaussian distribution, the objective of AdaQ is to achieve the optimal 3-$σ$ interval for different examples, i.e., a smaller 3-$σ$ interval for the local query and a larger one for the global query, thereby facilitating robust and adaptive frame sampling. To validate AdaQ, we apply it to four MLLMs with three embedding models. The extensive experimental results not only show its obvious performance gains over the default MLLMs and the SOTA keyframe selection methods, e.g., helping Qwen3-VL-8B outperform GPT4o by 15.8% on average by using only 64 frames, but also confirm its superior robustness and high efficiency for long-video understanding, e.g., only 1 hyper-parameter needs to be set.

Language models are running out of high-quality training data, and even aggressively deduplicated corpora retain some amount of repetition. Earlier controlled studies predated Chinchilla-style scaling laws and could only measure the cost of repetition indirectly. We revisit repetition in the Chinchilla era, using a fitted no-repetition scaling law to report Compute-Equivalent Gain and Compute-Equivalent Loss. We show that under this modernized paradigm, repetition damage is systematic in three ways. First, holding compute allocated to repeated data constant, eval loss peaks at an intermediate repeat count $\Rep$; repeating a moderately sized subset a moderate number of times damages performance more than repeating a large subset a few times or a small subset many times. Second, the location of this peak is well-fit by a power law in model size; this scaling law reveals that the most damaging number of repeated data grows more quickly than compute. Finally, when repeated documents consume 10% of the FLOPs budget in a controlled exact-document repetition setting, the compute-equivalent loss can be large: on FineWeb-Edu-Dedup, the most damaging repeat count for a Qwen3-style 344M-parameter model at $\OT=1$ matches the loss of a no-repetition run using 67% of the FLOPs. We demonstrate that these phenomena are not language-model-specific, and can be analytically understood in a simple statistical model: a misspecified linear regression with verbatim duplicates reproduces the same qualitative loss peak, quantifying how such peaks can arise from a statistical tradeoff between memorization and generalization. Our findings add precision to the study of duplication in language models, allowing practitioners to quantify the wasted compute incurred by the presence and repeat structure of duplicates in pretraining corpora.

Large Language Models (LLMs) are traditionally viewed as autoregressive generators. However, from the perspective of collective computation, they function as high-dimensional Dense Associative Memories that store complex reasoning patterns as latent attractors. In this work, we investigate the energy landscape of mathematical reasoning. We posit that correct reasoning chains correspond to deep, wide attractor basins ("flat minima") in the model's output distribution, whereas hallucinations manifest as sharp, unstable local minima. To exploit this geometry, we introduce a retrieval mechanism based on a Gibbs measure of the trajectory's spectral entropy. By sampling multiple reasoning paths and weighting them by their inverse energy ($P \propto e^{-βE}$), we approximate the equilibrium distribution of the associative memory, effectively ``relaxing'' the system into a robust solution. Empirically, this physics-inspired mechanism improves Microsoft Phi-3.5 performance on GSM8K by 5.38% (84.7% $\to$ 90.1%), demonstrating that inference is better modeled as a dynamic settling process into an attractor basin rather than greedy next-token prediction.

Unified multi-modal large language models (MLLMs) have achieved strong text-to-image generation quality, but still struggle with structure-aware prompt following, where object counts, spatial relations, attribute bindings, and coarse layouts must be preserved. We attribute this limitation in part to the entanglement of structural planning and appearance rendering within a single conditioning stream. To address this issue, we propose Implicit Visual Chain-of-Thought (IV-CoT), a latent visual reasoning framework for query-conditioned image generation. IV-CoT decomposes the visual conditioning queries into a structural-to-semantic cascade, where structural queries first form a latent visual plan and semantic queries then render appearance conditioned on this plan. To guide the structural queries, we introduce training-only sketch supervision, which encourages them to capture structure from sketches without requiring sketch extraction or intermediate decoding at inference time. IV-CoT performs implicit CoT reasoning in a single forward pass and achieves superior results on GenEval and T2I-CompBench. Visualizations and analyses demonstrate that the learned structural and semantic queries play complementary roles in structure-aware generation.

Multi-step reasoning with Chain-of-Thought (CoT) prompting remains fragile: logical errors or hallucinations in early steps silently propagate, producing confident but incorrect conclusions. This paper presents VeryTrace, a zero-shot verification-and-repair framework that formalizes natural-language reasoning traces into a structured, compilable representation. VeryTrace introduces a Domain-Specific Language (DSL) that (i) makes step dependencies explicit, (ii) mechanizes quantitative content as executable expressions, and (iii) structures semantic inferences via deduction schemas. Our hybrid verifier combines deterministic checks for computational correctness, dependency resolution, and constraint satisfaction with targeted LLM audits for non-mechanizable semantic judgments, enabling step-level error localization and repair. Across three diverse domains-competition mathematics (AIME 2025), robotics planning (LLM-BabyBench), and kinship reasoning (CLUTRR), VeryTrace improves accuracy over zero-shot baselines on state-of-the-art LLMs without requiring domain-specific training or in-context examples, demonstrating that formalized trace verification achieves both precision and generalization.

Although text-to-motion generation has achieved strong progress in synthesizing realistic single-person motions from language, extending it to text-driven 3D human-human interaction (HHI) remains non-trivial, as HHI requires modeling the underlying social structure that governs phase progression, actor roles, and inter-actor coordination. In this paper, we formulate HHI generation as a social structure modeling and grounding problem: the model must first infer how an interaction unfolds and how the two actors coordinate their roles, and then realize this structure as continuous, physically plausible, and partner-aware 3D motion. To study how such structure should be modeled, we first examine the capability boundary of large language models (LLMs) for HHI generation. Our analysis shows that LLMs can think by recovering phase decompositions and partner-aware roles, but cannot directly move, as they fail to generate dynamic, physically plausible, and interaction-aware motion. This motivates our planner-executor paradigm, Think with LLM, Move with Motion Skill. The LLM planner converts implicit interaction semantics into motion-aligned social supervision by decomposing interactions into phases, assigning partner-aware actor roles, and aligning them with motion sequence. The motion executor then grounds the planned social structure into coordinated two-person motion by adapting a pretrained solo motion model with LoRA, previous-phase self-conditioning, and ego-relative partner conditioning. Together, our Solo-to-Social framework bridges social organization and motion realization, producing 3D HHI with improved phase consistency, role alignment, and partner-aware coordination.

Reasoning language models are increasingly asked not only to answer difficult questions, but also to estimate their likelihood of success. Existing methods typically elicit confidence only once: either before thinking or after answering. We argue that confidence in reasoning models is state-dependent: before thinking, confidence should estimate the chance of the model correctly solving the prompt, while after thinking it should predict whether the realized answer is likely to be correct. This distinction determines the appropriate supervision target: prompt-level success should supervise confidence estimates made after seeing the prompt, while individual answer-level correctness should supervise confidence estimates made after answering. We introduce CALIBER (Calibration Before and After Reasoning), which elicits both estimates and supervises each with the target matched to its information state. Under this unified protocol, CALIBER reduces Expected Calibration Error (ECE) by 52.5% over the strongest single-confidence baseline on BigMathDigits for the 7B model, while achieving the best Brier score and AUROC, and remains within 2.1 points of the best accuracy. Further, on a larger 30B model, CALIBER achieves the best ECE on BigMathDigits while remaining competitive in Brier score and AUROC. Out of distribution, it achieves the best ECE and Brier score on GPQA and TriviaQA, and remains competitive on SimpleQA. Ablations further show that this position-target alignment is most beneficial under distribution shift where it consistently reduces calibration error across all out-of-distribution benchmarks.

While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.

Data augmentation is a simple and model-agnostic approach for exploiting known invariances in learning problems. Given a group acting on the input space, one augments the training set with transformed copies of each sample. Because it exploits symmetries without modifying the underlying learning algorithm, data augmentation can be applied broadly across learning methods. However, this universality comes at a computational cost: when the group is large, full group-sized augmentation quickly becomes computationally infeasible. This raises a fundamental question: Can partial data augmentation achieve the same statistical benefits as full augmentation in terms of generalization and sample complexity? We develop a general framework for investigating this question using Fourier analysis and the representation theory of finite groups. We show that, for a broad class of classical learning problems, partial data augmentation based on a randomly sampled subset of group elements achieves the same minimax rates as full augmentation, up to an approximation error that vanishes as the subset size increases. Our results provide a theoretical explanation for why partial augmentation can retain the statistical benefits of full augmentation despite enforcing symmetry only approximately, and shed light on a recently raised question in learning with symmetries: whether statistically optimal learning under general group invariances can be achieved using computationally scalable methods. Moreover, we prove a complementary impossibility result: enforcing exact invariance via data augmentation requires averaging over the entire group, and cannot be achieved by any strict subset when the hypothesis space is sufficiently expressive. Together, these results provide a unified perspective on full and partial data augmentation, as well as exact and approximate symmetry enforcement.

Motivated by an application in machine learning optimization, this paper focuses on the challenges of sampling a matrix uniformly from the unit spectral norm ball. It is proven that all singular values of sampled matrices converge to 1 almost surely as the matrix dimensions increase. This result provides the theoretical justification for a proposed simple sampling method applicable for large dimension sizes matching matrices found in modern large language models. Experimental results demonstrate both the convergence of the singular values, as well as the exact and proposed approximate sampling methods.

Visual token reduction has emerged as an effective strategy for accelerating Multimodal Large Language Models (MLLMs). Many existing methods prune tokens by ranking text-visual attention scores. However, we show that attention is often dominated by a model-induced prior: even without textual instruction, MLLMs tend to focus on certain task-agnostic regions. Consequently, the attention scores of instruction-conditioned tokens are suppressed, increasing the risk that these tokens are discarded during pruning. To address this issue, we propose Prior-Corrected Token Reduction (PriorTR), a training-free token reduction method that explicitly separates task-conditioned attention from the model-induced prior. PriorTR estimates the attention map of the prior, and contrasts it with the task-conditioned attention distribution to measure the additional usable information contributed by each visual token. Importantly, PriorTR computes both the model-induced prior and the task-conditioned posterior within a single forward pass by introducing a null token that serves as an instruction-agnostic probe in the attention block. This design avoids duplicated propagation. Extensive experiments across multiple multimodal benchmarks and MLLMs demonstrate that PriorTR consistently improves the trade-off between accuracy and efficiency over strong training-free baselines, particularly under aggressive token budgets.

Contextual entrainment, which is a newly discovered phenomenon in large language models (LLMs), refers to the tendency of a model to assign higher probabilities to tokens that appear in its context. In this work, we extend this phenomenon from the token level to the sentence level by examining the per-token mean log-probability of a sentence instead of the probabilities of individual tokens. We investigate sentence-level contextual entrainment across 26 LLMs from seven families and two datasets, which cover both subjective and objective tasks. We find that sentence-level contextual entrainment exists. This means that the sentences in the prompt (even if they are counterfactual statements) can significantly increase their probability during model inference time. As the model size increases, contextual entrainment gradually decreases. We also find that contextual entrainment is controlled by 2% to 4% of the attention heads. Turning off these attention heads can effectively mitigate contextual entrainment without hurting the model's performance.

As large language models (LLMs) are increasingly used in media production from journalistm to filmmaking, what impact do they have on the stories being told? Prior work has shown LLMs to perpetuate social biases, including those related to gender. We complement existing literature on gender bias in LLM outputs by auditing the network structure of LLM-generated movie screenplays through automating the Bechdel test, a popular measure of women's representation in literary and film works. We also introduce the use of social network analysis measures to further analyze representational bias in LLM-generated scripts. We evaluate screenplays generated by three state-of-the-art LLMs (GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5) against 768 corresponding human-written screenplays, finding that human-written scripts are more likely to pass the Bechdel test. However, other network analyses, like centrality, homophily, and triadic relationships demonstrate that in some cases LLM-scripts have less bias, although all script types demonstrate some representational bias under most measures. We conclude by discussing the continued need for further quantitative assessments of media representations and AI-generated content.