Papers

Large Language Models (LLMs) often face ethical tradeoffs in which several responses may be defensible but express different priorities, such as fairness, honesty, courage, or restraint. We introduce VirtueMap, a framework for describing these patterns through an Aristotelian virtue-ethics lens. Instead of asking for a single correct answer, VirtueMap asks humans or LLMs to rank all five responses to each of seven general, non-lethal, non-political, and non-religious ethical dilemmas. To define the reference orderings used for scoring, we first proposed, for each dilemma and virtue, an ordering of the five responses from most to least expressive of that virtue. We then collected more than 100 respondent evaluations per ordering and retained it as operational ground truth only when at least 95% confirmed it. Rankings are scored against these retained orderings using normalized Borda alignment, yielding profiles over Practical Wisdom, Justice, Truthfulness, Courage, and Temperance. We apply VirtueMap to nine LLM families in a repeated-run evaluation and find high mean rank consistency (90.3%), with the largest differences appearing on Courage, Temperance, and Justice. We also release an interactive website that computes profiles locally in the browser and compares respondents with measured LLM profiles.

Tool-using LLM agents increasingly read untrusted content while holding side-effecting tools such as payments, email, CRM, and infrastructure APIs, yet common framework defaults still conflate tool exposure with authorization. We audit whether LangChain/LangGraph, LlamaIndex, and the Stripe Agent Toolkit re-authorize each model-emitted call, with concrete argument values, before execution. Across pinned public-source commits, all three provide capability gating by default, but none provides a deterministic fail-closed per-call value authorization gate by default. We introduce ScopeGate, a five-stage PDP/PEP for agent tool calls: scope, authorization, money ceiling, idempotency, and default deny. Evaluation shows the identical unauthorized payout call executes under LangChain's default dispatch (with a companion LlamaIndex PoC) but is denied by ScopeGate; the tested control reports 0/48 static bypasses, 0/29 unauthorized attempts (40-iteration adaptive run), 0/10 benign false-denies, and Latam-GPT payment-agent containment at 10/10. ASR denotes attempted unauthorized action, containment is not a cure, deployment-tier claims are inference over measured model classes, and no CVE is asserted.

Predicting distant metastasis from the digital H & E slides of the primary tumor is a critical yet challenging task in computational pathology. Multiple Instance Learning (MIL) approaches can attend to subdomains in whole slide images (WSIs) that harbor features of pre-metastatic cancer regions. However, conventional MIL models largely treat tissue patches as unordered bags, discarding the spatial layout that defines how these regions are arranged and interact across the tissue. We propose that metastatic risk is shaped not only by local patch appearance, but also by the geometric organization of patches in the WSI and the interaction between the tissue compartments. To this end, we introduce Distance-aware Tissue Modeling for Multiple Instance Learning (DTMF-MIL), a spatial MIL framework that reinforces feature embeddings with explicit distance priors. By computing signed distance functions (SDFs) to capture regions with similar features, and representing each patch with radial-basis distance responses and local SDF statistics, DTMF-MIL learns positions of patches with respect to regional interiors and boundaries. The interactions between similar patches are contextualized across local tissue neighborhoods and used to guide slide-level attention while pooling patch feature evidence for metastasis prediction. We evaluate DTMF-MIL for prediction of distant prostate cancer metastasis in an internal prostate needle-biopsy cohort (IPC) from a large hospital system and on public TCGA-COAD and TCGA-KIRC datasets across multiple pathology foundation-model backbones. Across most dataset, backbone, and metric combinations, DTMF-MIL achieves the strongest results consistently.

Generative models in finance face the dual challenge of producing realistic data while satisfying strict regulatory and economic objectives, a requirement that standard tabular diffusion models cannot provide. To address this difficulty, we introduce Constrained Tabular Diffusion for Finance (CTDF), a novel integration of sampling-time feasibility operations with mixed-type tabular diffusion in financial applications. By incorporating a training-free feasibility operator into the reverse-diffusion sampling loop, CTDF enforces hard constraints for applications such as simulation, legal compliance, and extrapolation. Extensive experiments on large-scale financial datasets demonstrate zero constraint violations and improvement in scarce data utility. CTDF establishes a robust method for generating trustworthy and compliant synthetic data, opening new avenues for rigorous generative modeling and analysis in the financial domain.

We introduce MACROCAST, a lightweight Time Series Foundation Model (TSFM) for real-time macroeconomic forecasting. Existing TSFMs suffer from data leakage in two forms: temporal contamination, as the model may have seen the realized values of the series it forecasts, and revision bias, as training on fully revised data diverges from the preliminary, vintage-specific releases available to real-time forecasters. MACROCAST is, to our knowledge, the first TSFM that rules out both forms of leakage entirely: at no stage of training is the model exposed to information that would not have been available to a forecaster in real time. We train MACROCAST first on purely synthetic time series in approximately one GPU-day and then fine-tune it on synthetic time series drawn from Bayesian VARs, dynamic factor models, and ARIMA specifications estimated on vintage-specific ALFRED data. Because pretraining uses only simulated data and fine-tuning uses only real-time vintages, no observed future or revised value ever enters the model; each fine-tuning run takes nine minutes. Evaluated on the FRED-MD database in a genuine real-time out-of-sample exercise, MACROCAST improves on the AR(1) benchmark for roughly 80% of series-horizon pairs, matches or surpasses Chronos-2 – the strongest currently available TSFM – and outperforms the Bayesian VAR and dynamic factor model benchmarks, all in a data-leakage-free manner.

Agent-based AI has enabled the automation of tasks by exposing application tools and resources to large language models (LLMs). However, to improve scope and accuracy, agents are often given access rights that exceed those of ordinary users, introducing significant security risks. AI is routinely integrated into applications with a disregard to security, risking data exposure and breaching regulations. This paper applies the AI Trust, Risk, and Security Management (TRiSM) framework to a medical report-generation application to demonstrate how an insecure agent workflow can be transformed into security-conscious agentic workflow. Both workflows were evaluated across five LLMs (Claude Haiku 4.5, GPT-4.1-nano, GPT-4.1-mini, GPT-5.4-mini, and Gemini 2.5 Flash) on two report types, totalling 800 generations and 500 attack scenarios including RAG poisoning, data-field injection, and client-side network injection. The TRiSM-guided agentic workflow reduced mean attack success rates from 31% to 10% for RAG poisoning and from 42% to 25% for data-field injection, while eliminating the network injection vector entirely through server-side prompt construction. Furthermore, report accuracy increased by 14 percentage points (72.5% to 86.5%) with the agentic workflow, demonstrating a secure design which provides more reliable outputs. This paper contributes to knowledge by demonstrating least-privilege, defence in depth agentic workflows improving security and accuracy, while also highlighting model choice is a necessary architectural consideration.

The flatness hypothesis suggests that flatness of the loss landscape, as measured by the eigenvalues of the loss Hessian, correlates with better neural network generalization. While various algorithms reduce these eigenvalues, most focus on procedural design, leaving it unclear how data distributions and NN parameters structurally determine directions toward flat minima. Characterizing these directions analytically is generally intractable. To overcome this mathematical difficulty, recent studies derived the Wolkowicz-Styan (WS) upper bound on the maximum eigenvalue of the cross-entropy loss Hessian in three-layer NNs. Although this upper bound is differentiable, its gradient was not derived. Therefore, we analytically derive the gradient of the WS upper bound to characterize directions leading to flat minima. Based on this, we propose Hessian Spectral Range (HSR) Regularization, which updates parameters along the steepest descent direction of the WS bound. Experiments demonstrate that HSR Regularization narrows the Hessian eigenvalue spectrum, avoids sharp minima and saddle points, and promotes convergence to flat minima. Although the applicability of this method is currently limited to cross-entropy loss and three-layer architectures, to the best of the authors' knowledge, this is the first study to report a closed-form gradient that promotes convergence to flat minima without numerical approximations. Therefore, the theoretical analysis of this gradient is expected to contribute to the further development of NNs.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

Deformable 3D Gaussian Splatting (3DGS) has emerged as an efficient approach for rendering dynamic scenes in a wide range of 3D applications. However, existing deformation field-based approaches largely lack explicit object-level modeling, often resulting in inconsistent Gaussian deformations within individual objects and unwanted coupling between different objects. To address this limitation, we introduce a semantics-guided framework that enforces dynamic regularization at the object level, aiming to achieve spatially consistent object-wise deformation. Specifically, we first extract segmentation masks using the Segment Anything Model (SAM) and derive semantic features from input images. An object-ID map is then constructed via feature relevance matching with a predefined object dictionary. Guided by this object-ID map, we identify the pixel-wise top-k contributing Gaussians for each object and impose consistency regularization on their deformation parameters, including position, scale, and rotation. Unlike prior methods that learn deformation fields without explicit object-level constraints, our approach incorporates semantic cues to guide deformation behavior at the object level. Experimental results demonstrate that our semantics-aware regularization improves object-level deformation consistency and outperforms baseline methods in rendering quality, achieving higher PSNR and SSIM and lower LPIPS in dynamic 3DGS rendering. Our project page is available at https://dyn-reg-3dgs.github.io/.

Referring Expression Segmentation (RES) aims to generate pixel-wise segmentation masks from complex and implicit textual queries. While recent advances in Multimodal Large Language Models (MLLMs) have substantially boosted RES performance, their prohibitive computational overhead remains a critical bottleneck, which, however, is rarely explored. To fill this gap, we first evaluate typical token compression methods on this task and observe a surprising performance degradation. In this paper, we aim to understand this phenomenon for a solution. By extensive experiments, we find that token compression for RES requires preserving the original position embeddings and local neighboring spatial structures, indicating that visual token position information is far more critical than in other tasks. Building on this insight, we ask: Can we design the token compression method purely based on the position information? Therefore, we propose PAYN, a plug-and-play, training-free token compression method that relies solely on position information. PAYN retains tokens that are adequately distributed in every local neighboring region while strictly preserving original positional indices, thereby maintaining spatial relational consistency. Experiments on multiple RES benchmarks demonstrate that our method outperforms existing token compression methods, verifying that position is indeed all you need for token compression in the MLLM-based RES task. Codes are avaliable at https://github.com/YuhanLiu231/PAYN.

Large Language Models demonstrate remarkable syntactic fluency, yet the optimization dynamics governing their acquisition of deep semantic dependencies remain poorly understood. We propose a mechanistic framework that models this learning process as a competition between Surface Statistics and Deep Semantics. Our theoretical analysis identifies a “Gradient Starvation" phenomenon where the error signals for sparse semantic dependencies are actively suppressed during early optimization. This suppression impedes the learning of structural reasoning and causes its emergence to manifest as a sudden phase transition. Furthermore, this framework offers a mechanistic basis for the effectiveness of Chain-of-Thought (CoT) strategies. By externalizing intermediate reasoning steps into concrete tokens, CoT effectively bypasses the suppression regime inherent to implicit reasoning. We validate these findings across scales ranging from toy transformers to production models (Llama-3.1-8B, Qwen2.5-Coder-7B). Finally, guided by this theory, we propose a topology-aligned contrastive objective that explicitly rectifies the gradient geometry. Experiments on variable binding tasks demonstrate that our method achieves an improvement that is over 2x larger than that obtained via standard cross-entropy fine-tuning.

We propose a novel framework for the analysis of multimodal data – encompassing visual, auditory, and spatial stimuli – foregrounding the role of complexity in embodied perception and interaction in dynamic, naturalistic settings. Grounded in theories of embodied cognition and active vision, we argue that embodied perceptual complexity emerges from an agent's dynamic engagement with the environment and must be analyzed holistically, as a combination of qualitative and quantitative attributes pertaining to, for instance, visuospatial and auditory features. Building on previous work on visual complexity, we expand this into a categorization of diverse complexity attributes – quantitative, structural, dynamic, auditory, and interactional – that together characterize multimodal complexity. We demonstrate how this model provides a theoretical framework for characterizing aspects of visuospatial complexity and their interactions, specifically in the context of everyday driving. We also discuss practical applications of the proposed model for creating and evaluating benchmark datasets (e.g., in driving) that centralize cognitive human factors, as well as applications aimed at systematically investigating the effect of visuospatial complexity on human active vision from the viewpoint of visual perception research. The proposed framework lays the foundation for automated methods that interpret complexity in 3D dynamic environments from a human-centered perspective, serving as a semantic template for explainable computational analysis of visuospatial complexity with a categorical focus on cognitive human factors.

This report summarizes the work conducted on the Game of Hidden Rules (GOHR), focusing on reinforcement learning agents trained to infer hidden rules from trial-and-error feedback, representation design, rule difficulty analysis, transfer learning, generalization, and pseudo-bot-assisted human learning analysis. The report focuses on the Transformer-based A2C framework, Feature-Centric and Object-Centric representations, experimental findings, and classification of human learning data.

Post-training papers, model cards, and blog posts often treat scores on a small set of coding benchmarks (e.g., SWE-bench and LiveCodeBench) as evidence of broad coding capability, both for research artifacts and user-facing systems. We argue that optimization for these benchmarks leads to measuring task-specific performance, creating a meaning gap between measured scores and claims of general coding ability. We examine this gap with a Django-based case study benchmark suite we create. Evaluating foundation models and checkpoints post-trained on SWE-bench trajectories, we find that benchmark rankings frequently fail to generalize. Post-trained checkpoints show little cross-task transfer, and SWE-bench optimization yields limited or no gains on our tasks or on LiveCodeBench. Similarly, fine-tuning on individual Django modalities fails to transfer. We conclude that a small number of benchmarks is insufficient for evaluating diverse models under benchmark optimization pressure. We encourage the community to use differentiated evaluation - holistic assessment for frontier models, multi-task suites for research, and human-in-the-loop studies for narrow task applications. Finally, we argue for creating a capability taxonomy and sustained benchmark maintenance, rather than one-off benchmark releases. Without reliable evaluation standards, engineers and researchers using LLMs and agents have to rely on insufficient evidence to make research, development, and deployment decisions.

Music education has never been a static discipline. Each major technological shift has forced educators and institutions to reconsider what they teach, how they teach it, and why. We now stand at what may be the most consequential of such turning points. Three deeply intertwined transformations have been converging simultaneously: 1. The very nature of music has changed: how it is made, distributed, consumed, and valued; 2. The public for music has changed: listening habits are now shaped by streaming algorithms and the boundary between consumer and creator has blurred; 3. Music-making itself has changed: digital audio workstations (DAWs) have for two decades been reshaping compositional practice. In addition, generative AI has now irrupted, capable of producing complete, stylistically coherent musical pieces from a short text prompt. These changes are not independent of one another, and they all bear directly on musical education - both the content that must be taught, and the pedagogical tools and methods available to teach it. This paper attempts to map these challenges and to consider how education might adapt. Section 2 surveys the changes in various aspects of music (nature, production, public, economics). Section 3 examines the implications for education before Section 4 concludes.

As AI capabilities advance, AI systems will pose greater risks to national security and potentially humanity as a whole. Governments may eventually conclude that these risks warrant restraining AI development. This motivates the question: will governments still be able to restrain AI development in the future, should they want to do so? In this paper we analyze which world events and changes to the state of AI development would make future governance more difficult or even effectively impossible. Our analysis surfaces likely pathways that would lead to these difficulties, including hardware proliferation, continued algorithmic progress, and the release of catastrophically dangerous AI models. Due to the field's lack of understanding of AI development, it may be difficult or impossible to know when we will hit a "point of no return", and we therefore recommend a conservative approach. The window may be closing, but governments currently have an opportunity to preserve their optionality if they act soon. Our policy recommendations would enable governments to restrain AI development in the future, while imposing relatively small costs today.

The task-based framework in economics models occupations as bundles of tasks. It is the standard lens for understanding how technology affects work: a new technology changes the cost or time each task requires and these task-level effects aggregate to occupation-level effects. We study how tasks should be weighted in this aggregation. Prior work has relied on idiosyncratic or ill-justified choices for task weights. While recent work suggests weighting tasks by time spent, existing time shares are either based on coarse ONET data not intended for this purpose or estimated via black-box language models. We address this gap by proposing a principled method for estimating time shares for nearly 18,000 tasks that constitute nearly all U.S. jobs. Our estimates factor a task's time into (i) the expected frequency of the task, derived from ONET, and (ii) the time to complete a single instance of it. To estimate the latter, we solve a constraint satisfaction problem based on pairwise comparisons elicited from language models about which tasks are longer per instance. We validate our estimates by characterizing the solution space of the constraint satisfaction problem and collecting data from workers for multiple occupations. We apply our time shares to analyze how AI exposes U.S. occupations and find that some prior results are sensitive to time weights. Accounting for the share of working time exposed to AI, rather than the share of tasks like prior work, widens the gap between the least and most exposed jobs: it lowers measured exposure for most occupations but raises it for the most exposed. Re-weighting by time also reshuffles 11 of the 25 occupations widely reported as most exposed to AI, shifting the top of the list away from clerical work and toward analytical roles. Time shares can serve as a general primitive for research and policy on the labor economy and the economics of technology.

As Large Language Models scale to increasingly long contexts, the memory I/O and computational overhead of the Key-Value (KV) cache during decoding emerges as the primary throughput bottleneck. To address this, we propose GLIDE, a Guided Layerwise Hybrid Attention that strategically integrates sliding-window softmax attention with linear recurrent aggregation. GLIDE is motivated by layer-wise heterogeneity: early layers exhibit high sensitivity to softmax removal, while deeper layers demonstrate redundancy and tolerate aggressive replacement by linear alternatives. Leveraging this insight, GLIDE introduces a layer-wise adaptive mechanism wherein each layer balances an efficient linear recurrence with a variable-sized softmax window. Unlike uniform hybrid approaches, GLIDE non-uniformly compresses the softmax footprint across the model, reducing aggregate KV cache I/O while preserving expressive power where most vital. Empirical evaluations demonstrate the GLIDE achieves superior performance-efficiency tradeoffs, reducing end-to-end latency for long-context generation without compromising quality.

Estimating origin-destination (OD) flows under disruptive events is important for disaster response and urban resilience. Existing deep OD models trained on routine mobility often degrade when extreme events abruptly alter regional functions and population activities, while retraining a new generator for each event is impractical under limited event-time supervision. We propose EventOD, an event-adaptive OD generation framework that steers a pretrained OD generator using structured event semantics. EventOD first uses a large language model to infer region-level functional and demographic control vectors from coarse event observations. It then learns two lightweight adaptation modules, AlphaNet and BetaNet, to calibrate the magnitude of these semantic shifts, and further introduces a retrieval-augmented fallback pathway for scenarios with sparse supervision. The resulting event-conditioned features are injected into a pretrained graph diffusion OD model through input-level modulation, enabling event-aware adaptation without updating generator parameters. Experiments on hurricane- and pandemic-induced mobility across U.S. counties show that EventOD consistently improves both reconstruction accuracy and distributional fidelity over strong baselines. Source code is available at https://anonymous.4open.science/r/EventOD-5C11/.

Vehicle-to-Everything (V2X) communication systems are based on datasets that not only contain vehicle trajectory data but also wireless network parameters with a realistic level of fidelity, enabling the creation of prediction and optimization models. There is a very critical research infrastructure gap today, and publicly available datasets are likely to be limited to one of the two: mobility or network parameters, and rarely provide a single, integrated view that combines both. This paper introduces MINT-V2X, a comprehensive dataset generated by coupling SUMO traffic dynamics with OMNeT++/Simu5G network simulation. The validation framework is composed of 14 standardized tests based on 3GPP Release 14 (C-V2X), ETSI standards and Shannon capacity theory. The resulting dataset contains 9.87 million synchronized data points from 1,386 vehicles from 15 roadside units (RSUs) during 3 hours of urban traffic simulation. We demonstrate strict algorithmic consistency through network metric correlations (CQI-SINR: 0.993; SINR-PDR: 0.946). Finally, we demonstrate the value of the dataset by conducting an RSU load prediction case study, showing that using trajectory data yields better predictive performance than network-history-only baselines. The dataset, experiments, and complete SUMO configuration files are available in the GitHub repository to facilitate reproduction on alternative simulation stacks.

Large language models are increasingly used in recursive refinement workflows, where an initial draft is repeatedly revised by the same model. Despite their growing use, the long-term dynamics of such workflows remain poorly understood. Does repeated refinement continue to improve outputs indefinitely, or does it converge toward a stable textual form? We study recursive self-refinement as a dynamical process in which repeated LLM revision drives text toward a model-preferred soft fixed-point region. Using GPT-5.5, we generate 10-step refinement trajectories for 50 ICML 2025 abstracts under both default-temperature and deterministic decoding, and additionally evaluate 15 ICML 2020 abstracts. We analyze normalized edit distance, exact and approximate fixed points, word-count stability, exponential relaxation, and external LLM-as-a-judge evaluation. Across all settings, refinement trajectories rapidly saturate. Most edits occur within the first few iterations, after which trajectories enter a soft fixed-point region with only minor surface-level changes. Deterministic decoding reaches exact fixed points earlier and exhibits smaller residual fluctuations than default-temperature decoding, while both achieve universal approximate convergence. The average edit magnitude follows a consistent exponential relaxation pattern, suggesting convergence toward a model-preferred textual equilibrium rather than open-ended optimization. External evaluation indicates that converged abstracts improve clarity, conciseness, and scientific style while preserving technical meaning. These findings support a dynamical-systems view of LLM self-refinement and motivate practical stopping criteria based on edit-magnitude saturation.

Recent research has explored the integration of knowledge graphs (KGs) with large language models (LLMs) to enhance their performance on downstream knowledge-intensive tasks, particularly knowledge graph question answering (KGQA). Existing approaches primarily combine LLMs with KGs through retrieval-augmented generation (RAG)-based, agent-based, and SPARQL-based methods. Although these methods have achieved notable success, they still suffer from several limitations, including structural information loss, unfaithful reasoning, and limited flexibility and generalization. To address these challenges, this paper proposes KG2Code, a novel approach that transforms knowledge graphs into a code-based representation, preserving structural semantics while naturally aligning with the code-aware pretraining of modern LLMs. Based on KG2Code, KG2Code-QA is further introduced as a KGQA framework that formulates KGQA as a code generation task. This formulation enables the generation of verifiable reasoning traces and executable code, thereby substantially mitigating the impact of hallucinations. In addition, an automated pipeline is developed to construct a large-scale, high-quality code corpus for effectively training open-source LLMs on KG2Code-QA. After training, LLMs are able to perform KGQA in zero-shot scenarios. Extensive experiments demonstrate that the proposed approach significantly outperforms existing KG-enhanced LLM methods for KGQA, while exhibiting strong generalization to unseen KGs. The code and data are available at Github.

Large language models (LLMs) have enabled increasingly capable conversational agents, but reliably controlling their behavior in real-time interactive environments remains a significant challenge. Existing approaches often rely on model fine-tuning or alignment procedures that are difficult to adapt to changing interaction requirements. This paper introduces layered scenario-driven LLM control, a framework that enables runtime behavior control through structured prompting. By combining persistent context with scenario-specific constraints, the approach allows agent behavior to be modified during interaction without changing the underlying model. The framework is implemented in ARDena, a real-time multimodal embodied agent that integrates speech interaction, visual perception, tool use, and avatar-based response generation. The proposed approach is evaluated with respect to control effectiveness, response latency, and operational stability. The results demonstrate that scenario definitions alone can produce substantially different interaction behaviors while maintaining stable real-time operation, highlighting the effectiveness of scenario-driven prompting for controlling LLM agents.

Following complex instructions with multiple explicit constraints remains a fundamental challenge for large language models (LLMs). Existing alignment methods, such as DPO, optimize holistic reward signals that often underemphasize strict satisfaction of individual constraints, particularly under out-of-distribution or multi-constraint settings. In this paper, we propose STAIF, a stage-wise optimization framework that decouples the alignment of subjective (soft) constraints from the optimization of objectively verifiable (hard) constraints. Stage 1 applies preference optimization with multiple negative samples to sharpen sensitivity to soft constraints, while Stage 2 applies Reinforcement Learning with Verifiable Rewards (RLVR) to enforce strict compliance with hard constraints. To support this method, we construct STAINSTRUCT, a high-quality bilingual (English, Chinese) dataset of approximately 31,000 complex multi-constraint instructions. Extensive analyses validate the design of STAIF and show state-of-the-art performance on representative benchmarks against strong baselines, as well as genuine generalization.

We present and evaluate LeanFlow, an LLM agent system specialized for translating mathematical papers into buildable Lean projects. Recent verifier-in-the-loop systems show that large formal artifacts can be produced, but it remains unclear which runtime mechanisms affect completion, auditability, or efficiency in document-to-project formalization. We study this question through case studies on two previously unformalized mathematical papers in number theory and measure theory, using model, proof-workflow, and toolset ablations with Kimi2.6 and GPT5.5; we report task outcome, API calls, input tokens, and output tokens. With Kimi2.6, the full workflow completes both document-level projects within the 2000-call budget, while no-queue variants reach the budget limit; with GPT5.5, all document-level variants complete, and the full workflow has the lowest or tied-lowest input-token cost on both sources. As complementary calibration, LeanFlow reaches 75.7% BEq+ on the PFR slice of RLM25 and solves all five ICML 2026 AI for Math TCS challenge projects in our GPT5.5 runs.

Language models perform economically valuable work, yet they are not currently assessed for how well they perform every economically valuable task. We introduce EconEvals as an open-source evaluation suite to measure capabilities relevant to tasks, work activities, and occupations in the US labor economy. We ground the evaluation suite in real user queries to language models where possible, and supplement these with synthetic data. Our evaluations improve coverage over OpenAI's GDPval benchmark, which is the existing state-of-the-art that covers 5% of US occupations, at 500x lower cost. Alongside benchmarks, we also introduce a simulation-based exposure measure to estimate how much time current language model capabilities could save across all tasks belonging to all US occupations, with detailed accounting for each estimate. Our estimates indicate that current models could save workers substantial time on at least half of their tasks in 47% of occupations. However, for 79% of tasks where we predict substantial time savings, observed Claude usage is low, suggesting that existing usage lags potential. Beyond inherent constraints of language model chatbots, our data identifies privacy and proprietary systems as the principal bottlenecks limiting further time savings from AI. Overall, we introduce adaptable infrastructure that grounds inferences about language models' labor-market impact in their current capabilities, which can be continually updated as capabilities improve.

Off-policy distillation is now central to large language model pre-training, yet how training data, objective parameterization, and model capabilities interact remains poorly characterized. We studies top-$k$-truncated, temperature-scaled off-policy distillation by decomposing this problem into two questions: an objective-to-capability analysis of how the training objective shapes token-level supervision and downstream performance, and a data-to-objective analysis of how data heterogeneity should inform objective routing. We first show that the language-modeling objective ($L_{\mathrm{LM}}$) and the knowledge-distillation objective ($L_{\mathrm{KD}}$) induce systematically different capability profiles, and trace this divergence to a gradient-level tension between direct observed-token reinforcement and teacher-supported alternative supervision. To quantify this tension, we introduce diagnostic metrics – support coverage, observed-token probability mass, and teacher-distribution concentration – and show via controlled sweeps that the support size $k$ governs a coverage-sharpness trade-off, while distillation temperature controls within-support probability allocation. We then examine adaptive objective routing: a domain-level policy that applies $L_{\mathrm{LM}}$ to math and code and $L_{\mathrm{KD}}$ to general-domain data yields consistent gains over both single-objective baselines, whereas token-level routing based on observed-token probability mass or teacher entropy fails to consistently match the single-objective baseline. These results suggest that effective objective routing depends less on routing granularity than on the quality of the routing signal, reframing continued pre-training via off-policy distillation as a structured, data-conditional supervision-design problem rather than a global hyperparameter choice.

Gate-level netlists exhibit intrinsic structural properties that influence signal propagation independently of functional simulation. We define a topology-driven structural manipulability score that characterizes node-level structural flexibility using path participation, k-core embedding, symmetry, and centrality. Modeling netlists as directed graphs, we formulate node-level regression to learn this topology-derived score using graph neural networks (GNNs). Experiments on ISCAS85 and EPFL benchmarks evaluate how effectively different GNN architectures approximate this metric across held-out circuits, with hierarchical models yielding the most consistent rankings. Component-level and ablation analyses examine the contribution of individual factors. As an illustrative case study, analysis of Trojan-injected circuits using TrustHub templates reveals statistically distinguishable structural patterns, indicating that topology-based scoring provides complementary structural insight.

Training multi-turn evidence-reading agents with outcome-only reinforcement learning is unstable because intermediate turns receive little direct credit. In HotpotQA experiments with Qwen2.5-3B-Instruct, GRPO initially improves (standard F1 0.430) but subsequently collapses to 100% format-violating outputs. Training-log diagnosis reveals a zero-advantage lock-in mechanism: all sampled trajectories receive the minimum format penalty (-2.0), group-relative advantages vanish, and the policy-gradient loss becomes zero–an optimization deadlock. We propose a variance-injection strategy: by assigning per-turn rewards to intermediate evidence-reading turns, we prevent the group reward distribution from collapsing to a single value–preserving the variation that GRPO's group-relative advantage requires. Contextual Information-Gain Policy Optimization (CIGPO) implements this strategy using the marginal increase in the frozen reference model's log-likelihood of the ground-truth answer as the per-turn signal. With separate normalization of IG and F1 rewards and an IG-weight curriculum, CIGPO reaches a standard F1 of 0.518 on HotpotQA at the 3B scale (from 0.252 base; +105%), compared with 0.430 for the best GRPO checkpoint and 0.000 for the final GRPO checkpoint. CIGPO maintains meaningful reward variance and avoids zero-advantage lock-in throughout training. These results identify reward-variance collapse as a concrete failure mode of outcome-only GRPO and show that turn-level IG rewards can prevent it in this HotpotQA setting.

Multimodal industrial anomaly inspection assistants are a critical component of next-generation smart factories, enabling interactive vision-language-based querying. However, multimodal large language models remain impractical for on-site deployment due to prohibitive computational demands and privacy risks from cloud-based inference. Compact multimodal small language models (MSLMs) offer a deployable alternative, yet progress is constrained by the lack of comprehensive robustness analyses and meaningfully challenging benchmarks that reflect real-world industrial conditions. To address this gap, we develop RobustMAD, the first deployment-motivated benchmark, designed to comprehensively evaluate model robustness through diverse open-ended queries spanning object understanding, anomaly detection, unanswerable problems, and visual quality degradations. Contrary to conventional assumptions, top-performing MSLMs exhibit promising capabilities, surprisingly outperforming even the larger GPT-5 Nano. However, they still fall short of safety-critical requirements, and RobustMAD reveals critical robustness gaps that pose operational risks. In particular, three recurring failure modes emerge: (i) fragile multimodal grounding under fine-grained distinctions or degraded visual conditions, (ii) insufficiently comprehensive responses, and (iii) weak logical grounding on unanswerable or ill-posed queries, leading to hallucinated outputs. Grounded in these insights, we provide actionable guidance for the design of next-generation multimodal industrial inspection assistants that leverage their promising competence. Code is available at https://github.com/en-research/RobustMAD.