LLM 理论进展

Large language models solve grade-school math word problems with high accuracy, yet a single irrelevant clause inserted into the problem can collapse it. We reconcile these observations with a mechanistic account. We show that the model's internal computation decomposes into a four-stage sequential pipeline, Schema Abstraction, Operation Planning, Operand Binding, and Computation, each stage producing a distinct intermediate representation in an identifiable band of layers. Using the same scaffold to diagnose distractor-induced failure, we localize the corruption to a single stage, Operation Planning, implemented by a set of attention heads whose causal role we validate bidirectionally. In short, we provide a mechanistic interpretation of math word problem reasoning in LLMs, and their failure when distracted.

Large language models can produce fluent answers when their factual support is weak. This paper introduces Chain-of-Self-Questioning (CoSQ), a prompt-only framework that makes answer commitment conditional on an explicit assessment of the information required to answer a question. We evaluate three CoSQ variants under seventeen conditions on the 817-item TruthfulQA multiple-choice validation set using eleven open-weight and hosted model families. In the final balanced-option protocol, Grounded-CoSQ at τ=0.90 reduces the mean unconditional wrong-commitment rate from 13.1% under chain-of-thought prompting to 8.9%, a 32.1% relative reduction, while increasing answered accuracy from 86.9% to 89.7% and answering 87.6% of questions. Both improvements hold for all eleven models and at every evaluated threshold. Critical-CoSQ and Adaptive-CoSQ provide neighboring operating points with 88.6% and 86.5% coverage, respectively, while remaining more reliable than the baseline. A secondary Natural Questions Short-Answer evaluation provides convergent open-form evidence. These findings show that self-assessment can support explicit, tunable answer-or-abstain decisions when an unsupported commitment is more costly than referral or review.

Large language models can easily generate explanations for their own outputs, but such self-explanations are not necessarily faithful to the model's behavior. We study this issue through counterfactual self-explanations, where a model minimally edits an input so that its own prediction changes. Across sentiment analysis and natural language inference, we evaluate ten instruction-tuned models from the LLaMA-3 and Qwen-2.5 families, measuring faithfulness, minimality, and alignment with human-annotated rationales. Our results show that model scale is the strongest determinant of explanation quality: larger models are substantially more likely to generate counterfactuals that flip their own predictions and target decision-relevant evidence. In contrast, the rationale-guided condition produces edit-minimal counterfactuals that are also more human-aligned. However, it does not consistently improve faithfulness. Overall, counterfactual self-explanations can provide useful behavioral evidence about model decisions, but their reliability depends strongly on model capacity and should be empirically validated rather than assumed.

Stage lighting education requires instructors to bridge abstract concepts, technical operations, and learner-understandable representations. While Virtual Reality (VR) removes physical constraints, existing systems provide limited support for live instruction. We present LumiNote, an LLM-assisted VR system that transforms spoken pedagogical intent into instructor-reviewable spatial annotations, executable demonstrations, and linguistic support. In an exploratory study with 3 instructors and 24 students, we examined how instructors incorporated LumiNote into familiar lighting topics and how students received the resulting representations. We found LLM assistance most valuable for expressive, under-specified goals, but requiring greater expert intervention for fixture-specific or spatial configuration requests. Instructors engaged with generated suggestions as a controllable refinement process, shifting effort from manual setup toward pedagogical expression. However, representations that externalized expert reasoning did not always align with novice comprehension. These findings characterize LLM-assisted VR instruction as a domain-grounded mediation process among expert expression, executable operations, and learner-facing representations.

High-capacity associative memories based on Kernel Logistic Regression (KLR) exhibit a "Ridge of Optimization" characterized by extreme stability and a highly skewed weight spectrum. However, the dynamical process by which learning converges to this critical regime has remained unclear. This paper provides a geometric analysis of the learning trajectories on the statistical manifold of a KLR-trained Hopfield network. By comparing the paths of Gradient Descent (GD) and Natural Gradient Descent (NGD), we elucidate the mechanisms governing the optimization process. Our analysis reveals that learning on the Ridge proceeds in two distinct phases. We show that the extreme curvature of the Ridge causes standard GD to follow a highly oscillatory, non-geodesic path. In stark contrast, NGD explicitly corrects for this geometry, following the ideal geodesic path and completely overcoming the instabilities faced by GD. We demonstrate experimentally that NGD not only converges significantly faster but also achieves a solution with superior generalization performance. These results establish that the highly structured geometry of the Ridge is optimally suited for information-geometric optimization, providing a new perspective on the interplay between learning dynamics and emergent representation geometry.

Software clones are fragments of code that are similar or functionally equivalent to each other. They pose significant challenges for maintenance, refactoring, and bug detection. Detecting Type-IV clones, which are semantically equivalent but may differ syntactically, is particularly difficult for traditional token- or syntax-based methods. Recent machine learning approaches rely on contrastive learning, which requires careful negative sampling and can introduce bias. In this paper, we propose LWVIC4Code, a non-contrastive representation learning approach specifically designed for Type-IV clone detection. Building on the Variance-Invariance-Covariance Regularization (VICReg) framework and prior layer-wise VICReg training, LWVIC4Code introduces cross-layer consistency regularization and depth-dependent layer weighting to progressively refine semantic information across transformer layers, producing robust and discriminative code representations. We conduct an empirical study comparing LWVIC4Code against a contrastive learning baseline and zero-shot large language models on Python (Kamino) and multi-language (GPTCloneBench) datasets. Results show that LWVIC4Code achieves competitive or superior performance without negative samples, benefits from layer-wise supervision, and generalizes effectively from Python to other languages, particularly Java and C#. These results demonstrate that non-contrastive, layer-wise representation learning is a promising direction for robust semantic code clone detection.

Large Language Models (LLMs) have achieved remarkable progress across diverse domains, but continual adaptation to evolving tasks and environments remains a key challenge. Existing memory-augmented approaches retrieve individual past examples as direct references, but do not explicitly synthesize actionable strategies from them, causing the same types of errors to recur. We propose Dynamic Retrieval-based Policy Generation (DRPG), a framework that integrates memory-based retrieval with a dynamic policy generator, leveraging historical data and environment feedback to produce task-specific policies for continual LLM improvement. We evaluate DRPG across six benchmarks spanning text-to-SQL, question answering, medical diagnosis, and Python programming, using seven LLMs from both proprietary and open-weight families. DRPG outperforms strong baselines across most datasets and models. Further analysis demonstrates that DRPG's policy generation is robust to retrieval strategy, operates effectively without prior policy continuity, and can leverage smaller or cross-family models as cost-efficient policy generators. We also find that the benefit of policy-level guidance depends on task characteristics, offering practical insights into when and under what conditions this mechanism is most effective.

Multimodal large language models generate natural-language responses from visual inputs, yet may mention objects absent from an image. In medication assistance, accessible perception, and environmental decision-making, such hallucinations can create real-world safety risks. We propose Semantic-Spatial Agreement Verification (SSAV), a training-free method for verifying object claims. A visually grounded claim should remain stable across semantically equivalent queries and repeatedly localize to the same image region. SSAV aggregates multiple prompts to estimate semantic support and reduce sensitivity to query wording. Query-Induced Regional Verification (QIRV) combines cross-query region persistence, spatial overlap, and relative candidate dominance to identify isolated high responses and dispersed localizations. A geometric mean fuses semantic and spatial evidence, lowering the verification score when either branch lacks support. Experiments on three base models and multiple evaluation protocols show that SSAV effectively mitigates object hallucination. On LLaVA-1.5-7B, accuracy averaged across COCO, A-OKVQA, and GQA improves by 1.81 and 3.17 percentage points under POPE Popular and Adversarial, respectively, while CHAIRs decreases from 49.40% to 32.80%. These results show that cross-query semantic stability and regional consistency provide interpretable external visual evidence for object claims.

People increasingly use frontier AI models for health advice, but via different access modes (e.g., ChatGPT, ChatGPT Health, APIs) with varying settings. Here, we find systematic differences across access modes. Because evaluations typically rely on APIs while consumers interact through chatbot interfaces, these discrepancies limit evaluation validity. Our findings underscore an urgent need for model providers to enable faithful replication of consumer experiences and settings for rigorous audits.

Large Language Models are now common in student assessment, but we know little about how student demographics affect their use. Sometimes, considering student demographics may be necessary — for example, to improve readability for users with lower educational levels. However, it also risks being a cause of discrimination, e.g., when assigning lower scores to students from lower socioeconomic backgrounds. We set up controlled prompts to test 1) explicit demographic effects, where we mention demographic details directly, and 2) implicit effects, where we use conversation history as a demographic signal. We test these settings in three tasks: Automated Essay Scoring, Formative Feedback, and Metalinguistic Question Answering. We test six state-of-the-art LLMs on these tasks. In both explicit and implicit cases, the models pick up on demographic cues and can change their scoring, feedback, and answers accordingly. We find that LLMs frequently adjust the readability of feedback to education levels when these are explicitly mentioned. On the other hand, implicit conditions produce unpredictable biases, such as in question answering, where responses from lower-education levels receive lower sentiment scores. Our results provide clear evidence of demographic sensitivity in LLMs for educational assessment tasks.

People increasingly reason with large language models (LLMs), yet complementary capabilities do not guarantee outperforming both components. In a between-subjects study, participants (N=535) solved a 40-item battery of matrix reasoning, mental rotation, syllogisms, and letter-string analogies, unaided or with GPT-5.6-Luna, Claude Opus 4.8, Gemini 3.6 Flash, or Kimi K3. Each assisted trial required consultation with the model. Each model answered every item alone 100 times under matched elicitation. The assisted-unaided accuracy difference increased with item-level LLM competence. Deference varied across tasks and increased with competence within tasks. Post-advice confidence distinguished correct from incorrect answers less strongly than unaided confidence. In a reference comparison, about half the increase in LLM accuracy carried through to assisted accuracy. How much of that accuracy gain reached participants differed across the models. These findings motivate evaluating LLMs in interaction with humans and designing support for selective deference that preserves independent reasoning.

State Space Models (SSMs) have emerged as a compelling alternative to Transformers, enabling sequence modeling with constant memory and linear compute. Although SSMs exhibit reasonable performance and favorable computational characteristics, they continue to lag behind Transformers on tasks that require in-context learning and precise retrieval, slowing their adoption for large-scale language modeling. In this work, we demonstrate that both the success and failure of SSMs in these domains can be explained by studying the role of the gating mechanism, a prevalent component in modern recurrent networks. Specifically, we show through theory and experiments that this gating mechanism causes SSMs to first learn an in-weights "memorization" solution, while delaying, or even preventing, convergence to a correct in-context learning solution. Importantly, this happens even in cases where there are no fundamental limitations due to the architecture or its memory capacity. On the other hand, we find that gating is often beneficial for improving generalization to long sequence lengths. Our results illuminate the crucial role of the gating mechanism in shaping both the training dynamics and generalization of SSMs, and provide a basis for understanding and improving linear-time models.

Quantifying organizational environmental action from publicly available web content remains a challenging environmental data science problem because relevant information can be dispersed across multiple webpages and is primarily communicated through unstructured text. We present a scalable computational framework for transforming organizational web content into structured measures of environmental action and demonstrate the approach using Jewish congregations in the United States. We constructed a national database of 4,964 congregations by integrating multiple geospatial, knowledge-base, directory, and manually reviewed sources. Of these, 2,657 had active websites that were successfully crawled, producing a corpus of 154,454 webpages. We compared three approaches for detecting environmental actions: keyword retrieval followed by large language model (LLM) classification, semantic vector retrieval followed by LLM classification, and direct LLM classification classification without preliminary retrieval. Agreement with an expert human reviewer was lowest for keyword retrieval ($κ$ = 0.26), higher for semantic vector retrieval ($κ$ = 0.42), and similar for direct LLM classification ($κ$ = 0.40). Although semantic retrieval achieved the highest agreement, its retrieval recall was 0.87, indicating loss of relevant content before classification. Applied to the complete corpus, direct LLM classification identified at least one environmental action at 1,398 congregations (53%), providing greater coverage than either retrieval-based approach. These results demonstrate that preliminary retrieval can reduce computational cost but may exclude relevant information before it reaches the classifier. The framework provides a reproducible approach for extracting organization-level environmental information from unstructured web content that can be adapted to other institutions.

This paper addresses the challenge of making complex healthcare information more accessible through automated Plain Language Adaptation (PLA). PLA aims to simplify technical medical language, bridging a critical gap between the complexity of healthcare texts and patients' reading comprehension. Recent advances in Large Language Models (LLMs), such as GPT and BART, have opened new possibilities for PLA, especially in zero-shot and few-shot learning contexts where task-specific data is limited. In this work, we leverage the capabilities of LLMs such as GPT-4o-mini, Gemini-1.5-pro, and LLaMA for text simplification. Additionally, we incorporate Mixture-of-Agents (MoA) techniques to enhance adaptability and robustness in PLA tasks. Key contributions include a comparative analysis of prompting strategies, finetuning with QLoRA on different LLMs, and the integration of MoA technique. Our findings demonstrate the effectiveness of LLM-driven PLA, showcasing its potential in making healthcare information more comprehensible while preserving essential content.

Scaling Multimodal Large Language Models (MLLMs) to long-form video understanding is bottlenecked by the explosion of visual tokens, which saturates context windows and incurs prohibitive costs. Current solutions predominantly rely on auxiliary models for token reduction but face a fundamental dilemma: lightweight encoder-driven approaches often overlook critical semantic information, whereas heavyweight MLLM-driven reduction negates the efficiency gains. {In this work, we identify a more fundamental inefficiency underlying this dilemma: while fine-grained visual details are essential for detailed understanding, they are largely redundant for the preliminary task of selecting semantically relevant regions. } Motivated by this, we introduce VideoMM, which marks a paradigm shift from model-centric downsizing to adaptive perceptual granularity. Specifically, our framework {decouples selection from reasoning} by executing semantic filtering on a cost-effective Macro Proxy (derived from downscaled frames), and projecting the selected regions onto high-fidelity Micro Tokens for detailed understanding only when necessary. Extensive evaluations show that VideoMM significantly outperforms existing solutions. It achieves a 6.13$\times$ speedup and a 7.4% accuracy gain over full-context baselines on LongVideoBench, and further accelerates inference by 2.73$\times$ over current leading methods, establishing a highly scalable paradigm for long-video understanding. Our code is available at: https://github.com/adfh917k/VideoMM.

We introduce a simple architectural modification to decoder-only transformers: a persistent recurrent state that observes hidden representations via cross-attention, updates itself through a GRU, and modulates subsequent processing via gated addition. Inserted between the lower and upper halves of a 6-layer transformer, this module adds only 3.7% additional parameters while reducing evaluation loss from $2.438 \pm 0.004$ to $1.743 \pm 0.018$, corresponding to a 28.5% reduction on held-out language modeling data. The improvement is statistically significant across 5 random seeds ($p < 0.01$) and corresponds to reduced overfitting (generalization gap 0.12 vs 0.26). Through controlled ablations, we demonstrate that the improvement stems entirely from the persistent memory topology, not from auxiliary self-prediction objectives. A model with identical topology but no auxiliary loss performs equivalently, while a random auxiliary loss provides no benefit. Representation probing reveals that the persistent state encodes narrative position (52% vs 33% chance level)---information that standard attention maintains less efficiently. Our results suggest that bridging transformer layers with a lightweight recurrent memory is a simple, effective approach to improving generalization in small-scale language models.

Multimodal large language models (MLLMs) can answer knowledge-intensive visual questions by combining visual evidence from images with facts retrieved from external sources. However, MLLMs may overlook relevant evidence in both modalities, attending weakly to the textual sentences or visual regions needed for the correct answer. Recent efforts address this by highlighting retrieved text and marking visual regions before generation, but apply a fixed, one-shot policy that cannot adapt to three sources of variation: whether highlighting is necessary, how much evidence different examples require, and when different textual evidence becomes relevant as the answer unfolds. We introduce Adaptive Relevance-guided Evidence Allocation (AREA), a training-free inference-time method that formulates evidence highlighting as adaptive allocation. AREA generates a single probe token to read visual and textual relevance from fixed backbone layers, then makes three decisions: i) whether to intervene (controlled by natural attention coverage and visual sink contamination), ii) how much evidence to expose (determined by relevance entropy), and iii) when to refresh text during generation (triggered by causal context-attention peaks). Across four KB-VQA and seven standard multimodal benchmarks with nine frozen MLLM checkpoints, establishes the best performance among training-free highlighting methods.

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state — the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values from 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.

Large language models (LLMs) are increasingly used as user simulators, yet it remains unclear whether their predictions faithfully reproduce the evolving decisions of individual users. We investigate this question in a controlled longitudinal paper-trading study with 80 participants, where user interactions, simulated transactions, virtual portfolio states, and point-in-time market information are aligned under a rolling next-day prediction protocol. We evaluate behavioral fidelity hierarchically, from trade occurrence to action structure, asset selection, and downstream portfolio consequences. Across 1,239 aligned user-days, no evaluated LLM reliably outperforms a simple recent-activity persistence baseline for predicting whether a user trades. Fidelity further deteriorates at finer levels: models struggle to recover buy--sell structure and traded assets, and similar activity-level predictions can lead to substantially different portfolio trajectories. Controlled evidence ablations show that recent trading history strongly governs activity prediction, whereas asset selection is substantially more sensitive to the available evidence. An observational analysis further finds that intensified ticker-specific research predicts imminent trading, but diagnostic tests do not support a causal interpretation. These findings suggest that current LLMs capture useful short-term behavioral regularities without yet recovering a stable individual decision mechanism.

Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning. Our approach is based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction. Rather than trying to hide the true refusal circuitry, DDO actively injects a high-magnitude, nonlinear decoy signal into the network's MLP neurons. When an attacker attempts to locate the refusal direction, the decoy corrupts their estimator, tricking them into ablating a harmless orthogonal feature while the actual safety mechanism remains intact. We prove a spectral bound formalizing this effect and evaluate DDO across six model families, achieving <10% ASR under standard RFA. On Llama-3-8B-Instruct, DDO remains comparable to trained defenses under adaptive multi-phase attacks (65% vs. 58% worst-case ASR) and reduces Heretic weight-level attack ASR from 88.7% to 18%, all at 30 to 450 times lower optimization cost per configuration than the trained baselines.

Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily on the products of science and optimized using outcome-level signals, providing limited supervision for these process-level shifts in scientific judgment. We investigate whether such judgment can be recovered from scientist interaction traces and used to control the internal computation of a frozen frontier model. Using contrastive interventions collected during real scientific research, we identify a coordinated, low-dimensional control structure within Kimi 2.6, a trillion-parameter mixture-of-experts model. Residual analysis, attention-weight subspace alignment, and cross-layer singular value decomposition converge on a mid-depth control surface spanning key layers. We introduce Metacognitive Steering, an inference-time controller that reads the model's cognitive regime and dynamically composes layer-specific interventions for exploration, procedural convergence, or critical reassessment without modifying model parameters. Behavioral analyses show that this control produces more sustained exploration, explicit pruning, and evidence-responsive synthesis. We operationalize the method in Columbus-1, an autonomous research system that identified eight independently reproduced, attacker-reachable vulnerabilities in BlueZ and directed the design, simulation, and fabrication of a ten-foot rocket intended to land propulsively using non-throttleable solid motors. Together, these results show that process-level scientific judgment can provide supervision for interpretable, dynamic control over a model's reasoning strategy.

The difficulty of training large language models (LLMs), together with their ubiquity, raises the threat of stegomalware, where malicious payloads are embedded into model weights. Recent work has demonstrated the use of permutation symmetry in model weights to mitigate these threats, but failed to show neutralization of stegomalware across all weights for LLMs. In this paper, we demonstrate the full potential of behavior-preserving symmetries as a defense against stegomalware, as well as the risks these symmetries pose when exploited by attackers. For stegomalware neutralization, we improve upon previous work, demonstrating that it is possible to select permutations which displace all model parameters. This contrasts with previous methods which left a significant percentage of weights unaltered in LLMs. When used in an attack, we show that permutation symmetries can encode malware into the weights of a model in a way that is theoretically lossless, requires no retraining after encoding, and needs no payload-specific information in the extraction script---a combination of characteristics not previously seen in any single method. While theoretically lossless, permutation can in practice alter model behavior due to the accumulation of numerical error. We therefore quantify the loss in model performance associated with applying these methods, for both attack and defense, showing it to be minimal.

Peer stories have been shown to boost self-efficacy in older adults' health behavior change. Despite their effectiveness, peer stories are difficult to deploy in health promotion at scale given the difficulty of matching the diverse health concerns and coping styles of heterogeneous older populations. Large language models (LLMs) have been shown to generate authentic narratives, yet how personalization and narrative affective style, such as humor, jointly shape older adults' responses remains unknown. We developed a theory-driven system that generates first-person peer health narratives varying in personalization and humor through a three-stage LLM pipeline grounded in self-efficacy mechanisms. Thirty-one older adults were invited to participate in a within-subjects lab study. Results showed that personalization increased perceived relatability and relevance of peer stories, especially for older adults with lower humor preference. These findings position individual differences in affective styles as a second dimension in designing personalization for LLM-assisted health communication.

Geospatial artificial intelligence (GeoAI) powered by large language models (LLMs) is expanding the capacity to query, generate, and interpret spatial information through natural-language interfaces and agentic autonomous GIS workflows. This capability creates governance challenges that general AI ethics discussions do not fully capture, including passive location inference from mobility traces, spatially structured bias amplification driven by spatial autocorrelation and scale effects, hallucinated spatial facts, and uncertainty compounding across multimodal geospatial inputs. This narrative review identifies eight recurring issues in LLM-enabled GeoAI: data provenance and consent, spatial privacy and inference risk, algorithmic bias and spatial inequity, spatial mechanisms as structural risk (spatial autocorrelation, the modifiable areal unit problem, and scale effects), LLM-specific technical risks, explainability, policy and regulatory gaps, and public enablement and workforce development. For each issue, we characterize the underlying mechanism, ground it in an illustrative example from the literature, and assess the current state of technical or institutional responses, ranging from largely unaddressed to actively debated or subject to emerging policy. Building on this synthesis, we propose a governance-aware architecture for LLM-enabled autonomous GIS that maps each issue to enforceable controls and auditable artifacts across the geospatial data lifecycle, illustrated through a worked flood-response routing scenario. The review highlights a persistent evidence gap: proposed responses remain largely conceptual, and field-tested evaluations of governance controls for LLM-enabled GeoAI remain limited. We close by outlining a research agenda emphasizing empirical validation, spatially specific interpretability tools, and workforce training aligned with these emerging risks.

Simulations based on large language models (LLMs) have proven to be powerful for understanding human behavior, making them valuable additions to the social scientific toolkit. However, LLMs are ultimately black boxes based on deep neural networks which limits their value for social science. This is because of a lack of (i) interpretability: i.e. the ability to assign clear mechanisms driving observed behavior; and a lack of (ii) steerability: i.e. the ability to mute or amplify specific theoretically meaningful mechanisms of action to drive specific model behavior. Here, we demonstrate how the black box could be opened up to further enrich LLM-based simulations. Specifically, we compare three types of methods: (1) prompt-based manipulation, (2) SAE-derived feature steering, and (3) probe-based direction steering and examine their utility for LLM-based social scientific simulations. We do so by interpreting and steering two foundational components of human behaviors, namely preferences (risk attitudes, altruism) and capabilities (divergent creativity, product innovation), operationalized using four classic economic and creative tasks implemented as natural-language interactions. Overall, our results show that SAE- and probe-based techniques often outperform basic prompt-based methods for steering LLM agents, although this advantage depends on the specific prompting strategy involved. Together, SAEs and probes constitute an effective pipeline for social scientists seeking to interpret and steer agents in social simulations: SAEs decompose agents' internal representations into human-readable features, after which probes can reliably shift agents' behaviors in specified directions. We discuss implications of these methods for future work using LLM agents for social scientific simulations.

Chain-of-thought reasoning has revolutionized natural language processing by enabling large language models (LLMs) to decompose problems into intermediate steps before answering. Yet confining reasoning to the textual domain presents limitations for tasks requiring direct manipulation of visual representations. Recent efforts augment multimodal LLMs with external visual expert tools such as depth estimation or object detection modules, but these remain fundamentally limited by their reliance on narrow, rigid operations that cannot flexibly generate or transform visual content. We propose ReImaGin, which leverages image generation models as a flexible visual reasoning mechanism for multimodal LLMs: unlike fixed-function tools, they accept natural language commands and can perform open-ended visual operations, like removing an occlusion or generating a floorplan from multiple disjoint views of a room. Across six diverse visual reasoning tasks including multi-view spatial reasoning and collision prediction, ReImaGin consistently outperforms both text-only reasoning and specialist vision-tool baselines, with gains of up to 25%, demonstrating the advantage of flexible, generative visual reasoning.

Tabular foundation models deliver strong zero-training predictive performance via in-context learning, but their high inference latency makes them impractical as hot-path decision backends in interactive agentic loops. We distill a TabPFN teacher into a compact feed-forward student across a business-decision simulation on UCI Adult and five OpenML benchmarks: the classification head compresses 53.2M parameters to 8,546 (6,220x); the deployed two-head loan pipeline compresses 111.4M parameters to 17,059 (6,532x). The student retains 95.4-100.5% accuracy and 96.8-100.0% AUC, with the lowest accuracy retention on credit-g at 95.4%; an alpha = 0 hard-label control shows that the teacher's soft targets provide a 2.1-7.0 AUC point gain.

Long-video understanding remains challenging for multimodal large language models (MLLMs) because densely encoding long frame sequences is computationally expensive, while uniform sampling under a limited visual budget can miss sparse yet decisive evidence. Recent training-free keyframe selection methods have enabled more efficient inference and yielded promising performance gains. However, many existing methods score frames largely in isolation without explicitly considering how each candidate complements the currently selected subset, potentially resulting in redundant selections and incomplete evidence coverage. To address this limitation, we propose MarKey, a training-free framework that formulates keyframe selection as subset-aware greedy optimization. At each iteration, MarKey scores each candidate using a tractable surrogate that jointly accounts for query relevance, marginal coverage gain, and context-dependent redundancy, and selects the frame with the highest utility. To make this iterative subset-aware evaluation efficient, MarKey uses a compact set of representative anchors to approximate full-video coverage and a bounded window of previously selected frames to limit context-dependent comparisons. Experiments on six benchmarks spanning holistic video understanding, human-centric video understanding, and open-ended video understanding demonstrate that MarKey consistently outperforms existing methods. Further analyses show robust gains across different MLLM backbones, model scales, and frame budgets.

Evaluating the factual correctness of large language models (LLMs) is vital for many applications. But are our evaluation tools themselves trustworthy? Despite the rise of factuality-based metrics, their sensitivity and reliability remain underexplored. This paper introduces a meta-evaluation framework that systematically tests these metrics using controlled corruptions of gold standard answers. Our method generates ranked outputs with known degrees of degradation to probe how metrics capture nuanced changes in truthfulness. Our experiments reveal that pipeline-based methods, such as the RAGAS's factual correctness metric, better track degradation than LLM-as-judge approaches. We also propose a new variant of the factual correctness metric that provides a competitive and cost-efficient.

Large language models can hallucinate even when the knowledge required for a correct answer is already available. We study this failure through a latent-key view of inference, where answer selection depends on competition among associations acquired during pretraining. We show that model predictions can be highly sensitive to individual query keywords, that these influential keywords exhibit entity-specific binding, and that their effects are systematically shaped by pretraining frequency. Multiple bindings can also compete and exhibit higher-order interactions within the same query. Based on this mechanism, we introduce a two-stage keyword-perturbation method for hallucination detection. By removing influential keywords and measuring how the model reorganizes its prediction, the method distinguishes errors caused by misleading key associations from correct decisions supported by diagnostic evidence. Across multiple models and benchmarks, perturbation provides a strong and transferable detection signal, reaching $.910$ AUROC on probe-known ScientistQA. Finally, we extend the same probabilistic framework to four hallucination regimes: knowledge deficit, wrong knowledge, context distraction, and unstable inference. Their operational distributions across benchmarks provide diagnostic context for why different detector families succeed in different settings.