Searcharxiv⌕ Search

arXiv subjects

Yaodong Yang

Publications and source records attributed to Yaodong Yang.

At least 19 recordsLinked to original sources
arXiv2026

Your Group-Relative Advantage Is Biased

Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we characterize conditional estimation bias in group-relative advantages among non-degenerate groups. For mean-centered advantages under binary rewards, our analysis shows negative conditional mean estimation errors for hard prompts and positive errors for easy prompts. At the response level, correct answers to hard prompts are underweighted, while incorrect answers to easy prompts receive insufficient penalties. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

cs.LG↗
arXiv2026

LLM Detoxification via Transition-Aware Reward-Guided Representation Sampling

Large Language Models (LLMs) can generate toxic content, motivating effective test-time detoxification. Existing representation steering methods often rely on static or coarse interventions and provide imprecise guidance because they inadequately model transitions between toxic and non-toxic generations. We propose Transition-Aware Reward-Guided Representation Sampling (TARGRS), a test-time framework that learns reward guidance over toxicity transitions in latent representation space. During exploration, progressive steering along the toxicity direction constructs trajectories from non-toxic to toxic states. Toxicity scores along these trajectories turn sparse annotations into dense, listwise supervision for a transition-aware token-level reward model. During inference, TARGRS uses this model to guide coarse-to-fine sampling along the opposite, non-toxic direction: coarse sampling identifies the intervention needed to reach non-toxic regions, and refinement selects a detoxified representation with minimal perturbation. Experiments on 10 LLMs across RealToxicityPrompts, BOLD, and AttaQ show that TARGRS consistently outperforms state-of-the-art baselines, reducing toxicity by up to 69.38% on RealToxicityPrompts. TARGRS runs nearly as fast as the original LLM while preserving its core capabilities with minimal impact. By intervening in the language decoder space, TARGRS also extends to Large Vision-Language Models (LVLMs), reducing toxicity by up to 64.23% across 3 representative LVLMs.

cs.AI↗
arXiv2026

ResMerge: Residual Policy Learning and Merging for Continual Adaptation of Pre-trained Robot Policies

Continual adaptation to new tasks is essential for robots deployed in the real world, but adapting pre-trained policies remains challenging due to parameter growth and the need for continual demonstration collection. Parameter-Efficient Continual Fine-Tuning (PECFT) isolates task-specific updates, but adapter-per-task designs require demonstrations and grow with the task set. We propose ResMerge, a two-stage framework that learns mergeable residual policies atop a frozen pre-trained policy: residual reinforcement learning (RL) acquires task-specific corrections without additional expert demonstrations during adaptation, and parameter-space merging consolidates compatible residuals to reduce stored parameters. We introduce Residual-Field Consistency (RFC), a simple, bounded, and interpretable metric that assesses compatibility in residual action-field space by comparing correction direction and magnitude. Experiments on LIBERO across two backbones and on a real-world Franka robot show improved continual learning performance over strong PECFT baselines, while selective merging reduces the residual-parameter footprint. Project website: https://resmerge.github.io/.

cs.RO↗
arXiv2026

Heterogeneous Agent Collaborative Reinforcement Learning

We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new Reinforcement Learning from Verifiable Reward (RLVR) problem that addresses the inefficiencies of isolated multi-agent on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional homogeneous teacher-to-student transfer. Building on this problem, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO with double rollouts by an average of 3.6% while using only half the rollout cost.

cs.LG↗
arXiv2026

Evolutionary Safety of Recursive Self-Improving AI: Taxonomy, Risk Discovery, and Evaluation

Rapid advances in AI are bringing recursive self-improvement (RSI) closer to practice, with growing progress in model self-training, self-evolving agents, and automated AI development. Yet the very mechanisms that enable continuous improvement also introduce new challenges for AI safety. Unlike isolated failures, safety risks can persist across updates, accumulate through repeated adaptation, and propagate to successor systems. More importantly, the processes that generate and select future improvements may themselves evolve, making safety increasingly difficult to assess and maintain. This raises a fundamental question, i.e., how can safety be maintained when both AI systems and the processes that improve them continue to change? Therefore, we introduce Evolutionary Safety to study how safety properties are preserved, weakened, or restored throughout persistent and recursive self-improvement. We characterize six recurring manifestations of evolutionary risk and develop a five-domain taxonomy that captures how safety-relevant changes arise, persist, and propagate across evolving systems. Building on this foundation, we establish a framework for risk discovery and evaluation across states, updates, trajectories, and lineages, and examine governance principles for controlling modifications, verifying and authorizing updates, and ensuring traceability and recovery. We further identify open problems in maintaining safety as both AI systems and their improvement mechanisms continue to evolve.

cs.AI↗
arXiv2026

RMBench: Memory-Dependent Robotic Manipulation Benchmark with Insights into Policy Design

Robotic manipulation policies have made rapid progress in recent years, yet most existing approaches give limited consideration to memory capabilities. Consequently, they struggle to solve tasks that require reasoning over historical observations and maintaining task-relevant information over time, which are common requirements in real-world manipulation scenarios. Although several memory-aware policies have been proposed, systematic evaluation of memory-dependent manipulation remains underexplored, and the relationship between architectural design choices and memory performance is still not well understood. To address this gap, we introduce RMBench, a simulation benchmark comprising 9 manipulation tasks that span multiple levels of memory complexity, enabling systematic evaluation of policy memory capabilities. We further propose Mem-0, a modular manipulation policy with explicit memory components designed to support controlled ablation studies. Through extensive simulation and real-world experiments, we identify memory-related limitations in existing policies and provide empirical insights into how architectural design choices influence memory performance. The website is available at https://rmbench.github.io/.

cs.RO↗
arXiv2026

RAP-WAM: Learning World-Action Models through Predictive Rehearsal and Action Perturbations

Recent world action models draw on video generation priors to jointly predict future observations and robot actions. Adapting these priors to control requires determining how future predictions should guide policy learning. RAP-WAM combines predictive Rehearsal and Action Perturbations (RAP) with future-supervised queries and Coupled Action Projection for policy learning. Predictive rehearsal trains on model-generated inputs while retaining recorded targets, and accumulates recurring errors into a spatial prior. We perturb actions and measure changes in video predictions. Action-response-based spatiotemporal credit assignment combines these responses, prediction error, and the rehearsal prior to allocate visual supervision; spatially aggregating the responses determines supervision intensity across future action steps. No separate scoring network or deployment-time probe is required. The base learns from a video-pretrained backbone without intermediate pretraining on external embodied data. We evaluate the fixed base and its agent-refined, task-conditioned system. On 42 RoboDojo tasks, their average success rates are 12.85% and 15.62%, with Scores of 17.86 and 20.59. Across 15 real-robot tasks, the complete system achieves 89.00% task-macro-average success, exceeding the strongest of five baselines by 37.67 percentage points. Across 315 fixed-output simulation windows, 45.21% of response-top20 positions fall outside error-top20, revealing information beyond prediction error. Videos and further details are available on the project website: https://rapwam.github.io/.

cs.AI↗
arXiv2026

DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment

Safety alignment teaches language models to refuse harmful requests, yet the internal accessibility of hazardous concepts can retain traces of earlier training. Can checkpoints with equally safe behavior differ in what remains accessible behind their refusals? In this work, we investigate this question through paired training histories and a direct audit of intermediate representations. We introduce DANGERLENS, a probe that maps hidden states into vocabulary space and uses blinded semantic judgments to distinguish harmful goals and operational means from hazard recognition and safety control. Experiments on three Qwen1.5 scales show that checkpoints with prior harmful exposure generally retain greater hazardous concept accessibility after the same safety alignment, with the strongest separation in intermediate and later layers. The same harmful update also produces different behavioral trajectories from closely matched starting points. Human relevance validation and readout ablations reinforce the measured accessibility contrast. Our findings show that training history remains visible inside behaviorally matched models and motivate internal accessibility audits alongside output safety evaluations.

cs.AI↗
arXiv2026

Code Owns the Simulation, Jev Owns the Evaluation

Judgment models such as \jev{} return, in a single call and without reasoning text, a probability for each described option. This makes them attractive as an agent's action-selection layer, but it is unclear which decisions they can be trusted with. We test \jev{} on reflection tests, one-shot matrix games, the text game ALFWorld and robot control, and find a sharp boundary. \jev{} succeeds when the right option can be judged from what the input describes, which we call \emph{evaluation}. Specifically, it solves 99\% of the counterintuitive Cognitive Reflection Test questions. However, it fails when the right option depends on \emph{simulation} (i.e., predicting something not in the input), such as the opponent's action or the subgoal that must come first. In games, \jev{} plays suboptimally as if its rational opponent acted at random, because the opponent's action is not given. In ALFWorld, \jev{} favors commands that mention an object or place named in the task description. For example, given the task ``put a clean knife in the drawer'', \jev{} carries an unwashed knife straight to the drawer instead of first washing it at the sink. Surprisingly, many of these failures are not due to a lack of knowledge. Asked separately what the opponent will do, \jev{} usually answers correctly, and it responds well given the opponent's action. It fails when one call must both perform the simulation and evaluate based on it. This suggests letting code make the prediction or simulation. When code supplies it, such as a lookahead in ALFWorld and physics simulation in robot control, \jev{} becomes an expert controller through its general evaluation ability.

cs.AI↗
arXiv2026

Can Jev be Your Q or Policy in Reinforcement Learning?

Foundation models supply reinforcement learning (RL) with priors that mitigate its longstanding weaknesses in sample efficiency and transfer, but their token-by-token generation makes queries sequential and costly. Jev, a recently released decision model, generates nothing and returns calibrated, typed answers in a single forward pass. Existing work studies foundation models in RL either as models to be trained or as generators to be prompted, and Jev belongs to neither category, having so far served only as a black box in single domains. How well such a model decides on its own in RL environments, and how it can improve RL as a component of training, therefore remain unaddressed. To this end, in this paper we first examine the requirements that the objects of an RL system place on the answers they consume, and establish that Jev can fulfill all of them except the cardinal use of a value function. The remaining objects form positions that admit several roles each. We then construct algorithms that employ Jev at three of these positions, as a reference policy, an exploration judge, and a replay rater, to improve sample efficiency, exploration, and learning performance. Across nine MiniGrid tasks and three Atari games, training with Jev outperforms a standard RL learner, including where the learner makes no progress alone, while the model itself remains untrained. To our knowledge, we present the first use of Jev within the RL learning process and establish a frozen decision model as a usable component of RL training, inviting further exploration of how Jev and other advanced decision models can improve RL.

cs.LG↗
arXiv2026

EgoSteer: An Open-Source Full-Stack System Towards Steerable Dexterous Manipulation from Egocentric Videos

The enduring vision of general-purpose robots serving humanity hinges fundamentally on policy steerability. However, prevailing paradigms of learning from expert demonstrations demand massive real-world data even on simplified grippers, rendering them prohibitively expensive for high-dimensional, data-scarce dexterous hands. To overcome this bottleneck, we present a full-stack system that scales dexterous VLA pre-training from egocentric human videos and enables data-efficient real-robot post-training. It integrates EgoSmith, a data pipeline that curates in-the-wild egocentric videos into 9,606 hours of pre-training data with 8.3x higher throughput and better accuracy than prior SOTA; a unified Robot Stack for teleoperation and human-in-the-loop correction tailored for dexterous hands; and EgoSteer, a world-model-enhanced VLA operating on a morphology-aligned action space. Human data pre-training equips EgoSteer with language-guided manipulation priors, which are grounded through robot post-training and further refined via DAgger. Empirically, EgoSteer robustly executes free-form instructions across 45 diverse tasks, demonstrating adherence to user intent amid multiple candidate tasks and generalization. The pre-trained model also few-shot adapts to five complex long-horizon tasks, including box folding, on two embodiments with 79% average progress. All system code, datasets, model checkpoints, and an evaluation gallery are publicly available at https://egosteer.github.io/.

cs.RO↗
arXiv2026

From bits to atoms: A survey of cross-layer safety in Embodied AI

Embodied Artificial Intelligence (AI) introduces cyber-physical risks because digital failures and attacks can cause physical harm. This survey organizes embodied AI safety through a four-layer reference architecture -- comprising the physical interaction, system and middleware, algorithm and decision, and cloud-edge and data layers -- and uses it to connect vulnerabilities, propagation paths, and safeguards across the system stack. It traces the evolution of safety paradigms and examines how faults and attacks cross architectural boundaries to produce physical consequences. Furthermore, the survey systematizes composite attack chains, outlines multi-layered defense strategies, and reviews representative embodied safety benchmarks. Ultimately, we advocate for unifying software alignment with rigorous hardware constraints to build trustworthy systems capable of operating safely within human-centric environments.

cs.RO↗
arXiv2026

Cooperative Multi-Agent Vision-Language-Action Models via Reinforced Fine Tuning

We study reinforcement learning (RL) methods for cooperative multi-agent Vision-Language-Action (VLA) models. This problem is challenging because VLAs are pretrained on large-scale single-agent data and therefore lack the fine-grained coordination skills required for inter-robot collaboration. Supervised fine-tuning (SFT) on multi-robot demonstrations partially bridges this gap, but its performance is bounded by the demonstration data and cannot improve from its own experience. We present a three-stage reinforced fine-tuning (RFT) pipeline for multi-agent VLAs. First, initialization-aware data collection sweeps over initial configurations and invokes human demonstrations only when the pretrained VLA repeatedly fails, yielding robustness to initialization shift with reduced human cost. Second, offline credit-filtered tuning assigns credit to individual agents and fine-tunes on per-agent trajectories with positive advantage rather than on entire joint rollouts. Third, we find existing online RL for VLAs are less effective for hard multi-agent tasks, which we attribute to noisy co-exploration and unstable updates. We instead use online latent-space fine tuning, which freeze the VLA and perform RL in its latent noise space. We evaluate our multi-agent VLA with both $π_0$ and $π_{0.5}$ backbones across 11 tasks in RoboTwin, RoboFactory and real-world manipulation with two Franka robots. Our multi-agent VLA improves the average success rate by $+23.1\%$, $+16.4\%$, and $+44\%$ on RoboTwin, RoboFactory, and real-world tasks, respectively. Code available at https://anonymous.4open.science/r/mavla_rft-2BC0/.

cs.RO↗
arXiv2026

AgentBoundary: Counterfactual Evaluation of Safety in Tool-Using LLM Agents

Safety alignment for large language models (LLMs) in conversational settings is largely framed around whether to answer or refuse a request. In agentic settings, however, the same models must decide whether to act as permission-critical evidence emerges during execution. This creates a distinct challenge: apparent risk, action permissibility, and task competence are easily confounded, making agentic over-refusal difficult to distinguish from ordinary task failure. To address this, we introduce AgentBound, the first four-way counterfactual generation-and-evaluation framework for tool-using agent safety. AgentBound transforms the same executable workflow by independently varying apparent risk and action permissibility, enabling controlled comparisons of risky-looking but authorized tasks and routine-looking but unauthorized tasks. These comparisons jointly diagnose over-refusal and unsafe compliance while controlling for task competence. We instantiate AgentBound as a human-validated 4,000-task evaluation suite with trajectory-based and post-state-based judgments. Across 17 model and harness configurations, high safety frequently coexists with poor authorized-task completion: GPT-5.5 blocks 99.5\% of routine-looking unauthorized actions yet completes only 28.7\% of risky-looking authorized tasks. We further train a lightweight runtime calibration module that improves authorized-task completion by 18.2\% on average across 10 evaluated configurations, while improving unsafe-action blocking by 5.4\% on average. These show that effective agentic alignment requires action decisions to track permission-relevant execution evidence, rather than refusal strength alone.

cs.AI↗
arXiv2026

RoboFFT: Finetuning generative robot policy via online reinforcement learning with forward process

Generative models, such as diffusion and flow-based models, have shown strong promise for robot policy learning by capturing complex and multimodal action distributions from demonstrations. However, policies trained solely with imitation learning often suffer from imperfect demonstrations and distributional shifts, while further improvement typically requires additional expert data. Reinforcement learning offers a natural solution through environment interaction, but effectively finetuning generative robot policies remains challenging due to the intractability of likelihood estimation. In this work, we propose RoboFFT, a forward-process reinforcement learning framework for finetuning generative robot policies, which applies forward noising to sampled actions and uses the weighted score / flow matching loss to construct a surrogate policy ratio for PPO-style updates. We evaluate RoboFFT with popular generative robot policies on representative simulation benchmarks, including long-horizon planning and sparse reward settings. Extensive experiments and analysis demonstrate that RoboFFT consistently improves performance while achieving better stability and training efficiency. We further integrate RoboFFT into a real world RL framework and demonstrate its effectiveness in real world tasks. Project website: https://student-of-holmes.github.io/RoboFFT/.

cs.RO↗
arXiv2026

Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling

Data selection for language model pretraining faces a fundamental tension between quality and diversity. While quality filtering is empirically effective, it often induces diversity collapse: by favoring texts similar to high-quality reference corpora (e.g., educational or QA-style data), it systematically excludes valuable data from underrepresented domains. In contrast, diversified selection preserves domain balance and encourages robust downstream performance, yet existing methods either focus on coverage-oriented objectives that indirectly enhance diversity, or directly optimize for diversity via costly covariance matrix recomputation that limits scalability. To address these issues, we introduce \textbf{Leverage Score Sampling (Lev)}, which iteratively selects samples that maximally expand the determinantal volume of the embedded data via leverage scores, a computationally efficient criterion that eliminates matrix recomputation and enables scalable selection. Empirically, Lev delivers up to $72\times$ speedup and improves dataset diversity, measured by the Vendi score, by $9.2\%$ over the strong diversification baseline \textbf{DiSF}. On CommonCrawl (CC) web data selection, Lev improves accuracy across seven downstream tasks by up to $1.31\%$ over existing baselines. For domains where robust quality criteria are inherently difficult to define (e.g., code), Lev serves as an effective unsupervised curation alternative: on StarCoderData, the selected subset reduces bits-per-byte by $3.08\%$ over DiSF. Notably, we uncover a cross-domain collapse of quality filtering: CC data filtered by DCLM-fastText fail to retain sufficient code-related content, yielding inferior code performance relative to Lev-selected data. These findings advocate for integrating diversity-aware practices into quality filtering for more effective data curation in language model pretraining.

cs.AI↗
arXiv2026

TNLearn: An Open Source Python Package for Task-based Neurons

The brain does not rely on a single type of neuron to perform all kinds of tasks; instead, it designs different neurons for different tasks. The concept of task-based neurons represents a paradigm shift compared to task-based architectures. It argues that solving a specific problem requires customized neurons, as task-based neurons capture useful prior knowledge from task-related data. To facilitate the use of task-based neurons in scientific research and industrial applications, we introduce TNLearn, an open-source Python package that provides automated construction of task-based neurons and networks, enabling smooth training of task-based networks. Comprehensive documentation, including technical exposition, API reference, and representative examples, is available online. TNLearn is open-sourced at https://github.com/NewT123-WM/tnlearn and has become a PyTorch ecosystem project.

cs.LG↗
arXiv2026

Finding Kissing Numbers with Game-theoretic Reinforcement Learning

Since Isaac Newton first studied the Kissing Number Problem in 1694, determining the maximal number of non-overlapping spheres around a central sphere has remained a defining challenge in discrete geometry. As the local analogue of Hilbert's 18th problem, it has profound implications across geometry, number theory and information theory. Although lattices and codes have achieved significant progress, the field is confined to isolated extremal configurations, leaving underlying geometric principles obscured. Here we shift the object to the broader extremal configuration space, thereby opening a new path for the Kissing Number Problem. Accordingly, we recast this problem as a cooperative matrix-completion game, and train a reinforcement learning system, PackingStar, to solve it. One player fills cosine entries while the other corrects suboptimal ones, making explosive geometric complexity tractable. Working within extremal configuration spaces, PackingStar discovers new interpretable geometric structures that improve 15 strong bounds held for decades in kissing numbers and their generalizations, several of them provably optimal under natural inner products. These findings reveal the first explicit spherical-code realization of the Fischer group Fi22, extend the classical Euclidean representation of subgroup structure, and directly inspire subsequent breakthroughs by mathematicians. Overall, the work provides an early example of AI-driven progress on a Hilbert-calibre problem, showing how reinforcement learning advances mathematical discovery by unlocking more expressive objects.

cs.LG↗