arXiv:2602.19633v1 Announce Type: new Abstract: Language Model (LM) agents have demonstrated remarkable capabilities in solving tasks that require multiple interactions with the environment. However, they remain vulnerable in environments where a single error often leads to irrecoverable failure, particularly under strict feasibility constraints. We systematically analyze existing agent frameworks, identifying imperfect planning and stochastic execution as the primary causes. To address these challenges, we propose Tool-guided Adaptive Planning with constrained Execution (TAPE). TAPE enhances planning capability by aggregating multiple plans into a graph and employing an external solver to identify a feasible path. During execution, TAPE employs constrained decoding to reduce sampling noise, while adaptively re-planning whenever environmental feedback deviates from the intended state. Experiments across Sokoban, ALFWorld, MuSiQue, and GSM8K-Hard demonstrate that TAPE consistently outperforms existing frameworks, with particularly large gains on hard settings, improving success rates by 21.0 percentage points on hard settings on average, and by 20.0 percentage points for weaker base models on average. Code and data available at here.

arXiv:2602.19914v1 Announce Type: new Abstract: Existing benchmarks for AI reasoning provide limited insight into how closely these capabilities resemble human reasoning in naturalistic contexts. We present an adaptation of the Watson & Holmes detective tabletop game as a new benchmark designed to evaluate reasoning performance using incrementally presented narrative evidence, open-ended questions and unconstrained language responses. An automated grading system was developed and validated against human assessors to enable scalable and replicable performance evaluation. Results show a clear improvement in AI model performance over time. Over nine months of 2025, model performance rose from the lower quartile of the human comparison group to approximately the top 5%. Around half of this improvement reflects steady advancement across successive model releases, while the remainder corresponds to a marked step change associated with reasoning-oriented model architectures. Systematic differences in the performance of AI models compared to humans, dependent on features of the specific detection puzzle, were mostly absent with the exception of a fall in performance for models when solving longer cases (case lengths being in the range of 1900-4000 words), and an advantage at inductive reasoning for reasoning models at early stages of case solving when evidence was scant.

arXiv:2602.18464v1 Announce Type: cross Abstract: A growing body of research assumes that large language model (LLM) agents can serve as proxies for how people form attitudes toward and behave in response to security and privacy (S&P) threats. If correct, these simulations could offer a scalable way to forecast S&P risks in products prior to deployment. We interrogate this assumption using SP-ABCBench, a new benchmark of 30 tests derived from validated S&P human-subject studies, which measures alignment between simulations and human-subjects studies on a 0-100 ascending scale, where higher scores indicate better alignment across three dimensions: Attitude, Behavior, and Coherence. Evaluating twelve LLMs, four persona construction strategies, and two prompting methods, we found that there remains substantial room for improvement: all models score between 50 and 64 on average. Newer, bigger, and smarter models do not reliably do better and sometimes do worse. Some simulation configurations, however, do yield high alignment: e.g., with scores above 95 for some behavior tests when agents are prompted to apply bounded rationality and weigh privacy costs against perceived benefits. We release SP-ABCBench to enable reproducible evaluation as methods improve.

arXiv:2602.19193v1 Announce Type: cross Abstract: As one of the simplest non-prehensile manipulation skills, pushing has been widely studied as an effective means to rearrange objects. Existing approaches, however, typically rely on multi-step push plans composed of pre-defined pushing primitives with limited application scopes, which restrict their efficiency and versatility across different scenarios. In this work, we propose a unified pushing policy that incorporates a lightweight prompting mechanism into a flow matching policy to guide the generation of reactive, multimodal pushing actions. The visual prompt can be specified by a high-level planner, enabling the reuse of the pushing policy across a wide range of planning problems. Experimental results demonstrate that the proposed unified pushing policy not only outperforms existing baselines but also effectively serves as a low-level primitive within a VLM-guided planning framework to solve table-cleaning tasks efficiently.

arXiv:2602.19313v1 Announce Type: cross Abstract: While Vision-Language-Action (VLA) models have seen rapid progress in pretraining, their advancement in Reinforcement Learning (RL) remains hampered by low sample efficiency and sparse rewards in real-world settings. Developing generalizable process reward models is essential for providing the fine-grained feedback necessary to bridge this gap, yet existing temporal value functions often fail to generalize beyond their training domains. We introduce TOPReward, a novel, probabilistically grounded temporal value function that leverages the latent world knowledge of pretrained video Vision-Language Models (VLMs) to estimate robotic task progress. Unlike prior methods that prompt VLMs to directly output progress values, which are prone to numerical misrepresentation, TOPReward extracts task progress directly from the VLM's internal token logits. In zero-shot evaluations across 130+ distinct real-world tasks and multiple robot platforms (e.g., Franka, YAM, SO-100/101), TOPReward achieves 0.947 mean Value-Order Correlation (VOC) on Qwen3-VL, dramatically outperforming the state-of-the-art GVL baseline which achieves near-zero correlation on the same open-source model. We further demonstrate that TOPReward serves as a versatile tool for downstream applications, including success detection and reward-aligned behavior cloning.

arXiv:2602.19631v1 Announce Type: cross Abstract: Recent advances in text-to-image (T2I) diffusion models have seen rapid and widespread adoption. However, their powerful generative capabilities raise concerns about potential misuse for synthesizing harmful, private, or copyrighted content. To mitigate such risks, concept erasure techniques have emerged as a promising solution. Prior works have primarily focused on fine-tuning the denoising component (e.g., the U-Net backbone). However, recent causal tracing studies suggest that visual attribute information is localized in the early self-attention layers of the text encoder, indicating a potential alternative for concept erasing. Building on this insight, we conduct preliminary experiments and find that directly fine-tuning early layers can suppress target concepts but often degrades the generation quality of non-target concepts. To overcome this limitation, we propose High-Level Representation Misdirection (HiRM), which misdirects high-level semantic representations of target concepts in the text encoder toward designated vectors such as random directions or semantically defined directions (e.g., supercategories), while updating only early layers that contain causal states of visual attributes. Our decoupling strategy enables precise concept removal with minimal impact on unrelated concepts, as demonstrated by strong results on UnlearnCanvas and NSFW benchmarks across diverse targets (e.g., objects, styles, nudity). HiRM also preserves generative utility at low training cost, transfers to state-of-the-art architectures such as Flux without additional training, and shows synergistic effects with denoiser-based concept erasing methods.

arXiv:2602.19679v1 Announce Type: cross Abstract: Joint reconstruction of 3D human and object from a single image is an active research area, with pivotal applications in robotics and digital content creation. Despite recent advances, existing approaches suffer from two fundamental limitations. First, their reconstructions rely heavily on physical contact information, which inherently cannot capture non-contact human-object interactions, such as gazing at or pointing toward an object. Second, the reconstruction process is primarily driven by local geometric proximity, neglecting the human and object appearances that provide global context crucial for understanding holistic interactions. To address these issues, we introduce TeHOR, a framework built upon two core designs. First, beyond contact information, our framework leverages text descriptions of human-object interactions to enforce semantic alignment between the 3D reconstruction and its textual cues, enabling reasoning over a wider spectrum of interactions, including non-contact cases. Second, we incorporate appearance cues of the 3D human and object into the alignment process to capture holistic contextual information, thereby ensuring visually plausible reconstructions. As a result, our framework produces accurate and semantically coherent reconstructions, achieving state-of-the-art performance.

arXiv:2602.20051v1 Announce Type: cross Abstract: 3D human pose estimation (HPE) is characterized by intricate local and global dependencies among joints. Conventional supervised losses are limited in capturing these correlations because they treat each joint independently. Previous studies have attempted to promote structural consistency through manually designed priors or rule-based constraints; however, these approaches typically require manual specification and are often non-differentiable, limiting their use as end-to-end training objectives. We propose SEAL-pose, a data-driven framework in which a learnable loss-net trains a pose-net by evaluating structural plausibility. Rather than relying on hand-crafted priors, our joint-graph-based design enables the loss-net to learn complex structural dependencies directly from data. Extensive experiments on three 3D HPE benchmarks with eight backbones show that SEAL-pose reduces per-joint errors and improves pose plausibility compared with the corresponding backbones across all settings. Beyond improving each backbone, SEAL-pose also outperforms models with explicit structural constraints, despite not enforcing any such constraints. Finally, we analyze the relationship between the loss-net and structural consistency, and evaluate SEAL-pose in cross-dataset and in-the-wild settings.

arXiv:2602.20113v1 Announce Type: cross Abstract: Voice style conversion aims to transform an input utterance to match a target speaker's timbre, accent, and emotion, with a central challenge being the disentanglement of linguistic content from style. While prior work has explored this problem, conversion quality remains limited, and real-time voice style conversion has not been addressed. We propose StyleStream, the first streamable zero-shot voice style conversion system that achieves state-of-the-art performance. StyleStream consists of two components: a Destylizer, which removes style attributes while preserving linguistic content, and a Stylizer, a diffusion transformer (DiT) that reintroduces target style conditioned on reference speech. Robust content-style disentanglement is enforced through text supervision and a highly constrained information bottleneck. This design enables a fully non-autoregressive architecture, achieving real-time voice style conversion with an end-to-end latency of 1 second. Samples and real-time demo: https://berkeley-speech-group.github.io/StyleStream/.

arXiv:2509.19800v3 Announce Type: replace Abstract: There are two primary approaches to solving Markov decision problems (MDPs): dynamic programming based on the Bellman equation and linear programming (LP). Dynamic programming methods are the most widely used and form the foundation of both classical and modern reinforcement learning (RL). By contrast, LP-based methods have been less commonly employed, although they have recently gained attention in contexts such as offline RL. The relative underuse of the LP-based methods stems from the fact that it leads to an inequality-constrained optimization problem, which is generally more challenging to solve effectively compared with Bellman-equation-based methods. The purpose of this paper is to establish a theoretical foundation for solving LP-based MDPs in a more effective and practical manner. Our key idea is to leverage the log-barrier function, widely used in inequality-constrained optimization, to transform the LP formulation of the MDP into an unconstrained optimization problem. This reformulation enables approximate solutions to be obtained easily via gradient descent. While the method may appear simple, to the best of our knowledge, a thorough theoretical interpretation of this approach has not yet been developed. This paper aims to bridge this gap.

arXiv:2602.17053v3 Announce Type: replace Abstract: Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for reasoning faithfulness, defined by two testable conditions: stance consistency (a coherent stance linking reasoning to answer) and causal influence (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present RFEval, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can reduce reasoning faithfulness, even when accuracy is maintained. Crucially, accuracy is neither a sufficient nor a reliable proxy for faithfulness: once controlling for model and task, the accuracy-faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process. Our code and dataset can be found at project page: https://aidaslab.github.io/RFEval/

arXiv:2502.08941v3 Announce Type: replace-cross Abstract: This paper analyzes multi-step temporal difference (TD)-learning algorithms within the ``deadly triad'' scenario, characterized by linear function approximation, off-policy learning, and bootstrapping. In particular, we prove that $n$-step TD-learning algorithms converge to a solution as the sampling horizon $n$ increases sufficiently. The paper is divided into two parts. In the first part, we comprehensively examine the fundamental properties of their model-based deterministic counterparts, including projected value iteration, gradient descent algorithms, which can be viewed as prototype deterministic algorithms whose analysis plays a pivotal role in understanding and developing their model-free reinforcement learning counterparts. In particular, we prove that these algorithms converge to meaningful solutions when $n$ is sufficiently large. Based on these findings, in the second part, two $n$-step TD-learning algorithms are proposed and analyzed, which can be seen as the model-free reinforcement learning counterparts of the model-based deterministic algorithms.

arXiv:2506.05850v3 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable reward (RLVR) has been instrumental in eliciting strong reasoning capabilities from large language models (LLMs) via long chains of thought (CoT). During RLVR training, we formalize and systemically study an empirical phenomenon whereby a multilingual model's CoT reverts to its dominant pre-training language (e.g., English) even when prompted in another language, which we term Cross-lingual Collapse. Because the long-CoT regime magnifies exposure to linguistic priors, the underlying trade-off between maximizing reasoning depth and preserving target-language fidelity has remained under-characterized. To examine this trade-off, we train LLMs with Group-Relative Policy Optimization (GRPO) on translated versions of math datasets widely used to elicit long-CoT reasoning. Throughout training, we track both task accuracy and the language consistency of reasoning chains. Our experiments yield three findings: (i) under RLVR, CoT in LLMs systematically drifts toward the pre-training dominant language as reasoning performance rises; (ii) English-centric priors, long-CoT GRPO optimization, task difficulty, and high-entropy decoding jointly amplify this drift, and the pattern persists beyond mathematics; and (iii) interventions that favor target-language traces--via a language-consistency reward, decoding-time controls, or more balanced backbones--mitigate collapse but reveal a persistent performance-fidelity trade-off.

arXiv:2510.01650v2 Announce Type: replace-cross Abstract: Neural network pruning is a promising technique to mitigate the excessive computational and memory requirements of large language models (LLMs). Despite its promise, however, progress in this area has diminished, as conventional methods are seemingly unable to surpass moderate sparsity levels (50-60%) without severely degrading model accuracy. This work breaks through the current impasse, presenting a principled and effective method called $\texttt{Elsa}$, which achieves extreme sparsity levels of up to 90% while retaining high model fidelity. This is done by identifying several limitations in current practice, all of which can be traced back to their reliance on a surrogate objective formulation. $\texttt{Elsa}$ tackles this issue directly and effectively via standard and well-established constrained optimization techniques based on ADMM. Our extensive experiments across a wide range of models and scales show that $\texttt{Elsa}$ achieves substantial improvements over existing methods; e.g., it achieves 7.8$\times$ less perplexity than the best existing method on LLaMA-2-7B at 90% sparsity. Moreover, we show that $\texttt{Elsa}$ remains stable even at extreme sparsity (e.g., 95\%), yielding up to $\times$3.98 inference speedup and $\times$7.80 memory compression over its dense counterpart. We also present $\texttt{Elsa}_{-L}$, a quantized variant that scales to extremely large models (27B), and establish its theoretical convergence guarantees.These results highlight meaningful progress in advancing the frontier of LLM sparsity, while promising that significant opportunities for further advancement may remain in directions that have so far attracted limited exploration.

arXiv:2510.07231v3 Announce Type: replace-cross Abstract: Socio-economic causal effects depend heavily on their specific institutional and environmental context. A single intervention can produce opposite results depending on regulatory or market factors, contexts that are often complex and only partially observed. This poses a significant challenge for large language models (LLMs) in decision-support roles: can they distinguish structural causal mechanisms from surface-level correlations when the context changes? To address this, we introduce EconCausal, a large-scale benchmark comprising 10,490 context-annotated causal triplets extracted from 2,595 high-quality empirical studies published in top-tier economics and finance journals. Through a rigorous four-stage pipeline combining multi-run consensus, context refinement, and multi-critic filtering, we ensure each claim is grounded in peer-reviewed research with explicit identification strategies. Our evaluation reveals critical limitations in current LLMs' context-dependent reasoning. While top models achieve approximately 88 percent accuracy in fixed, explicit contexts, performance drops sharply under context shifts, with a 32.6 percentage point decline, and falls to 37 percent when misinformation is introduced. Furthermore, models exhibit severe over-commitment in ambiguous cases and struggle to recognize null effects, achieving only 9.5 percent accuracy, exposing a fundamental gap between pattern matching and genuine causal reasoning. These findings underscore substantial risks for high-stakes economic decision-making, where the cost of misinterpreting causality is high. The dataset and benchmark are publicly available at https://github.com/econaikaist/econcausal-benchmark.

arXiv:2602.13540v1 Announce Type: cross Abstract: Large language models (LLMs) are widely deployed as general-purpose problem solvers, making accurate confidence estimation critical for reliable use. Prior work on LLM calibration largely focuses on response-level confidence, which estimates the correctness of a single generated output. However, this formulation is misaligned with many practical settings where the central question is how likely a model is to solve a query overall. We show that this mismatch results from the stochastic nature of modern LLM decoding, under which single-response correctness fails to reflect underlying model capability. To address this issue, we introduce capability calibration, which targets the model's expected accuracy on a query. We formally distinguish capability calibration from response calibration and show that the two differ both theoretically and empirically. We establish an empirical evaluation setup and study a range of confidence estimation methods. Our results demonstrate that capability-calibrated confidence improves pass@$k$ prediction and inference budget allocation, establishing a foundation with potential for diverse applications.

arXiv:2602.14233v1 Announce Type: cross Abstract: Large Language Models (LLMs) are increasingly integrated into financial workflows, but evaluation practice has not kept up. Finance-specific biases can inflate performance, contaminate backtests, and make reported results useless for any deployment claim. We identify five recurring biases in financial LLM applications. They include look-ahead bias, survivorship bias, narrative bias, objective bias, and cost bias. These biases break financial tasks in distinct ways and they often compound to create an illusion of validity. We reviewed 164 papers from 2023 to 2025 and found that no single bias is discussed in more than 28 percent of studies. This position paper argues that bias in financial LLM systems requires explicit attention and that structural validity should be enforced before any result is used to support a deployment claim. We propose a Structural Validity Framework and an evaluation checklist with minimal requirements for bias diagnosis and future system design. The material is available at https://github.com/Eleanorkong/Awesome-Financial-LLM-Bias-Mitigation.

arXiv:2508.13213v3 Announce Type: replace Abstract: Strategic decision-making requires balancing immediate opportunities against long-term objectives: a tension fundamental to competitive environments. We investigate this trade-off in chess by analyzing the dynamics of human and AI gameplay through a network-based metric that quantifies piece-to-piece interactions. Our analysis reveals that elite AI players sustain substantially higher levels of strategic tension for longer durations than top human grandmasters. We find that cumulative tension scales with algorithmic complexity in AI systems and increases linearly with skill level (Elo rating) in human play. Longer time controls are associated with higher tension in human games, reflecting the additional strategic complexity players can manage with more thinking time. The temporal profiles reveal contrasting approaches: highly competitive AI systems tolerate densely interconnected positions that balance offensive and defensive tactics over extended periods, while human players systematically limit tension and game complexity. These differences have broader implications for understanding how artificial and biological systems navigate complex strategic environments and for the deployment of AI in high-stakes competitive scenarios.

arXiv:2509.25260v2 Announce Type: replace Abstract: The extent to which decoder-only language models (LMs) engage in planning, that is, organizing intermediate computations to support coherent long-range generation, remains an important question, with implications for interpretability, reliability, and principled model design. Planning involves structuring computations over long horizons, and considering multiple possible continuations, but how far transformer-based LMs exhibit them without external scaffolds, e.g., chain-of-thought prompting, is unclear. We address these questions by analyzing the hidden states at the core of transformer computations, which capture intermediate results and act as carriers of information. Since these hidden representations are redundant and encumbered with fine-grained details, we develop a pipeline based on vector-quantized variational autoencoders that compresses them into compact summary codes. These codes enable measuring mutual information and analyzing the computational structure of the underlying model behavior. Using this framework, we study planning in LMs across synthetic grammar, path-finding tasks, and natural language datasets, focusing on two planning properties: (i) the planning horizon of pre-output computations, and (ii) the extent to which the model considers alternative valid continuations. As a separate downstream use of the same pipeline, we also analyze how decision-relevant information is distributed across layers and earlier prefix blocks when producing next-token predictions. Together, these analyses advance our understanding of planning in LMs and provide a general-purpose pipeline for inspecting internal model dynamics. Our results reveal that the effective planning horizon is task-dependent, that models implicitly preserve information about unused correct continuations, and that predictions draw most on recent computations, though earlier blocks remain informative.

arXiv:2511.11079v3 Announce Type: replace Abstract: We present ARCTraj, a dataset and methodological framework for modeling human reasoning through complex visual tasks in the Abstraction and Reasoning Corpus (ARC). While ARC has inspired extensive research on abstract reasoning, most existing approaches rely on static input-output supervision, which limits insight into how reasoning unfolds over time. ARCTraj addresses this gap by recording temporally ordered, object-level actions that capture how humans iteratively transform inputs into outputs, revealing intermediate reasoning steps that conventional datasets overlook. Collected via the O2ARC web interface, it contains around 10,000 trajectories annotated with task identifiers, timestamps, and success labels across 400 training tasks from the ARC-AGI-1 benchmark. It further defines a unified reasoning pipeline encompassing data collection, action abstraction, Markov decision process (MDP) formulation, and downstream learning, enabling integration with reinforcement learning, generative modeling, and sequence modeling methods such as PPO, World Models, GFlowNets, Diffusion agents, and Decision Transformers. Analyses of spatial selection, color attribution, and strategic convergence highlight the structure and diversity of human reasoning. Together, these contributions position ARCTraj as a structured and interpretable foundation for studying human-like reasoning, advancing explainability, alignment, and generalizable intelligence.