arXiv subjects

Gen Li

Publications and source records attributed to Gen Li.

At least 19 recordsLinked to original sources

arXiv2026

HARTS: A Prefix-Sharing Training System for Agentic RL with Hybrid-Attention Models

Agentic reinforcement learning (RL) generates irregular rollout trees, but trajectory-by-trajectory training recomputes their shared interaction histories. Existing training-time prefix-reuse systems do not jointly address state recovery, dense execution, and distributed training for hybrid-attention models. We present HARTS (Hybrid-Attention RL over Tree Structures), to our knowledge the first system to demonstrate training acceleration from arbitrary rollout-tree prefix sharing on real hybrid-attention models. A two-level planner constructs compression-oriented microbatches, then jointly refines their membership, cross-microbatch sharing, placement, and execution order. Within each sharing window, gradients from dependent microbatches accumulate before one backward pass through the shared prefix computation. Differentiable boundary-state handoff and bounded replay preserve the per-trajectory chunkwise numerical path, while a linear-time planner attains the global minimum number of sequential linear-attention calls under our packed chunkwise-call model. The runtime uses a common compact layout across attention and dense or mixture-of-experts (MoE) layers, with support for distributed parallelism and activation recomputation. Recovering log-probabilities for each original trajectory position preserves RL objectives; for deterministic top-k routing without token dropping, occurrence counts restore MoE objective weights and expert-load statistics. On SWE-bench, HARTS delivers 4.77-10.45x the update throughput of trajectory-expanded training across two sampling schemes on 7.9B- and 124B-parameter hybrid-attention MoE models using 8-64 GPUs. On the 7.9B model, full-model logit differences are comparable to baseline self-rerun variation. The 124B model shows stable online learning trends through step 150 on SWE-bench and $τ^3$-Bench.

cs.LG↗

arXiv2026

Provably Efficient Q-learning for Robust Reinforcement Learning

Standard reinforcement learning (RL) algorithms based on Markov Decision Processes (MDPs) are vulnerable to sim-to-real gaps. To address this issue, distributionally robust RL algorithms based on robust MDPs (RMDPs) have been proposed, for ensuring that agents can perform reliably to unpredictable conditions in real-world environments. Despite of the rich literature, \revision{an efficient model-free method is still lacking} for robust RL, particularly for Markovian data, which is more relevant in practical applications. In this work, we take a step towards settling this problem by proposing a sample efficient (asynchronous Q-learning) algorithm for robust RL. In particular, consider a $γ$-discounted infinite-horizon RMDP with state space $\mathcal{S}$ and action space $\mathcal{A}$, where the uncertainty set is specified by the total variation (TV) distance with uncertainty level $σ$. Assuming access to a single trajectory of Markovian samples induced by a behavior policy and the nominal transition kernel, the sample complexity of the proposed Q-learning algorithm --- to achieve an entrywise $\varepsilon$-approximation of the optimal Q-function --- is bounded by $$ \frac{1}{μ_{\mathsf{min}}\varepsilon^2(1-γ)^3\max\{1-γ,σ\}} + \frac{t_{\mathsf{mix}}}{μ_{\mathsf{min}}(1-γ)} $$ up to logarithmic factors. Here, $t_{\mathsf{mix}}$ and $μ_{\mathsf{min}}$ denote the mixing time and the minimum state-action occupancy probability of the sample trajectory, respectively. This result is nearly minimax optimal, with an additional factor of $\frac{1}{1-γ}$. Moreover, it achieves a significant improvement over existing bounds.

stat.ML↗

arXiv2026

Unified Condition-Action Modeling for Accurate One-Step Action Generation

Robot manipulation requires policies that are both accurate and efficient, as robot control must respond to changing observations under tight latency constraints. Recent diffusion and flow policies are promising, but they often treat conditions as auxiliary signals rather than jointly evolving them with action trajectories. We find that this limitation can be effectively mitigated by a simple yet effective unified condition-action modeling design that represents conditions and actions in a shared token space, allowing a compact model to achieve high performance while improving both inference speed and accuracy. Therefore, we propose UCA-Flow, a unified condition-action modeling framework for accurate one-step action generation. Our method unifies observation conditions, timestep conditions, interval conditions, and action tokens into a single sequence, and processes them with a Unified Condition-Action Transformer for joint condition-action representation learning. As a result, condition representations are dynamically reconstructed according to the current generation stage, highlighting information most relevant for action refinement. Furthermore, we introduce an improved dual-pass supervision scheme over $u$ and $v$ for stronger optimization of unified condition-action modeling. UCA-Flow achieves an average success rate of 86.4%, improving over the strongest baseline by 7.5 percentage points, with a task-averaged inference latency of $3.0$ ms. It achieves $43.9\times$ and $32.2\times$ speedups over DP3 and Simple DP3, and remains $4.2\times$ and $2.2\times$ faster than one-step FlowPolicy and MP1, respectively. Project page: https://uca-policy.github.io/UCA.github.io/

cs.RO↗

arXiv2026

Feedback World Model Enables Precise Guidance of Diffusion Policy

World models aim to improve robotic decision making by predicting the consequences of actions. However, in practice, their predictions often become unreliable once the robot encounters states outside the training distribution, limiting their effectiveness at deployment. We observe that execution itself provides a natural but underutilized signal: after each action, the robot directly observes the true next state, revealing the mismatch between predicted and actual outcomes. Building on this insight, we propose feedback world model, an inference-time mechanism that closes the loop between prediction and observation at inference time. Instead of treating the world model as a static open-loop predictor, our method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations without additional training data or parameter updates. We show that this process can be interpreted as a latent-space observer and admits convergence guarantees under bounded, asymptotically convergent dynamics residuals and a stable feedback gain. We further introduce action-aware guidance to better translate corrected predictions into control by emphasizing action-controllable components while suppressing irrelevant variations. Experiments on LIBERO-Plus, Robomimic, and real-world manipulation tasks demonstrate that our method substantially improves both prediction accuracy and policy performance under distribution shift, reducing world model prediction error by up to 76.4%. The same feedback mechanism improves the predictions of the pretrained foundation model Fast-WAM-Joint without updating its parameters, demonstrating its plug-and-play applicability. These results show that incorporating real-time feedback at inference time provides a simple yet powerful alternative to static world modeling.

cs.RO↗

arXiv2026

HarnessPAI: An Evolving Harness for Physical AI

Physical AI aims to build embodied agents that perceive the world, understand and reason about it, and decide how to act. Yet the field has focused primarily on the last component: the action model that maps observations to low-level controls. The prevailing training recipe can erode the perceptual and reasoning capabilities needed for robust behavior, leaving even strong action models vulnerable to scene perturbations and long-horizon tasks. We introduce HarnessPAI, a model- and embodiment-agnostic Harness framework for Physical AI that treats code as the executable and evolvable interface that organizes the underlying action primitive. The framework separates two timescales: within a rollout, it executes open-loop at the program level, with a fixed program guiding and checking execution; across rollouts, it evolves closed-loop, using execution feedback to revise the program and distill failures into reusable skills. Across desktop robot arms, household robots, a robot vacuum, and a legged walking agent, HarnessPAI demonstrates consistent improvements against baselines. Beyond the harness, the converged program also acts as a low-cost, reliable source of expert data: it generates training data for fine-tuning VLA and WAM models on tasks where it is hard for these models to obtain successful trajectories on their own. Our results suggest that the frontier of Physical AI depends not only on stronger action models, but also on executable harnesses that integrate perception, task understanding and reasoning, and action execution into a unified, verifiable, and feedback-driven system. Website: https://darwin-agent.github.io/HarnessPAI

cs.RO↗

arXiv2026

MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports

Deep research agents are increasingly capable of searching the open web and synthesizing evidence, yet they remain poorly suited to controllable report generation in multimodal deep research. This task requires agents to handle interleaved textual and visual inputs, gather multimodal evidence through multi-step web search, and deliver comprehensive illustrated reports under user-specified requirements. Existing systems face two limitations. During reasoning, they often underuse images discovered along the search trajectory; during generation, they tend to produce fixed-form answers with limited control over report granularity, multimodal composition, and evidence grounding. We introduce MIRA, a Multimodal Illustrated Deep Research Agent for controllable reports. MIRA integrates textual and visual evidence throughout research and writing. At its core, MIRA-Harness translates user requirements into generation contracts and connects four stages: evidence collection, report submission, feedback-driven revision, and final delivery. Building on this harness, MIRA-Data constructs high-quality multi-hop multimodal research trajectories, and we post-train foundation models to operate within the workflow. Using Qwen3.6-35B-A3B, MIRA achieves state-of-the-art performance among open agents on 7 representative, established multimodal-search and deep-research-agent benchmarks, while producing user-aligned illustrated reports with traceable textual and visual evidence.

cs.CV↗

arXiv2026

Drive2Gauss: FeedForward 4D Gaussian Reconstruction from Multimodal Generative Driving Videos

Generating reusable driving environments requires both controllable future observations and an explicit representation that supports rendering from different viewpoints, yet existing methods still recover explicit scenes from generated RGB videos through optimization for each sequence. To overcome these limitations, we present Drive2Gauss, a feedforward framework that converts jointly generated RGB, metric depth, and dynamic flow into dynamic 4D Gaussian scenes. A Multimodal Video Generator predicts the three modalities from reference multiview images and future control signals. Furthermore, we introduce a Gaussian Query Decoder that uses depth to initialize sparse 3D queries and flow to associate the queries across time, and integrates appearance information through spatiotemporal feature aggregation and iterative query refinement. The reconstruction stage operates in a single feedforward pass without optimization for each scene. Extensive results on nuScenes demonstrate that our approach attains an FID of 6.38 and an FVD of 12.43 for future RGB generation, a PSNR of 29.91 dB when reconstructing its generated observations, and supports both direct rendering under lateral camera shifts and downstream video fixing. Code and more visualizations are available at https://zttt04. github.io/Drive2Gauss_page/.

cs.CV↗

arXiv2026

RoutePack: Replay-Guided Expert Placement and Rollout Trajectory Scheduling for MoE Reinforcement Learning

Reinforcement learning (RL) post-training of Mixture-of-Experts (MoE) models couples two load-balancing problems. Rollout-trajectory composition determines attention and dense feed-forward work across data-parallel (DP) replicas, while token routing determines expert work across expert-parallel ranks. Optimizing either side independently can worsen the other. Routing replay exposes trajectory lengths and layer-wise expert demand before training, creating an opportunity to coordinate both controls. We present RoutePack, a replay-guided planner that coordinates expert placement and rollout-trajectory scheduling. Its key insight is that placement determines a batchwise expert-rank load bound that trajectory scheduling alone cannot lower, while scheduling controls the hotspots within individual execution micro-steps. RoutePack incrementally adjusts expert placement while preserving expert residency, then uses parallel population annealing to reschedule complete trajectories into token-capped microbatches. Its objective jointly balances dense-side work and layer-wise expert peaks under a fixed, capacity-feasible micro-step count and the deployed synchronization groups. The planner preserves logical token routing and keeps placement fixed throughout forward and backward execution to maintain training-state consistency. We evaluate mixed-task RL post-training of two hybrid-attention MoE models (7.9B and 124B parameters) across six NVIDIA H20-3e configurations spanning 8K-32K contexts. Within the Megatron training region, RoutePack achieves 4.44-11.85% higher throughput than a DP-local first-fit-decreasing baseline and 1.33-5.28% higher throughput than expert rerouting alone. Additional evaluations of the 7.9B model on H200 and H20 retain positive gains. We separately characterize planning and admission costs.

cs.DC↗

arXiv2026

C-MIG: Multi-view Information Gain-based Retrieval-Augmented Generation for Clinical Diagnosis Reasoning

Retrieval-augmented generation (RAG) trained with reinforcement learning (RL) is a promising way to ground large language models in medical evidence for clinical diagnosis. However, existing RAG-RL methods face two problems: (1) their process reward is largely singleview, as most methods inherited from opendomain multi-hop QA reward only the quality of the retrieved information; (2) assigning a reliable process reward is hard, as hard matching gives zero to clinically correct but non-verbatim steps such as "Allergic Sinusitis" versus "Allergic Rhinitis", and the alternative of strong-teacher guiding is bounded by the teacher's own medical competence. To address these issues, we propose C-MIG, a Multiview Information Gain-based framework for Clinical diagnosis. We identify two complementary and crucial intermediate steps in clinical diagnosis: what to retrieve and how to refine. We reward both by the information gain each brings, measured as the increase in the predicted likelihood of the ground-truth answer under a base, frozen reference model. The frozen scorer yields a stable and reproducible reward that needs no extra reward model or teacher. Experiments on four medical benchmarks show that C-MIG outperforms RAGRL baselines in and out of domain and stays competitive with far larger medical-specialist and general-purpose LLMs, at only about 3.4% training overhead and no inference-time cost.

cs.AI↗

arXiv2026

Refine Actions If Needed: Adaptive Action Chunk Refinement for Robot Manipulation

Computationally lightweight manipulation policies predict executable actions with low inference latency, but prediction errors can limit task success. Iterative generative policies can achieve higher task success rates, but repeated network evaluations increase inference latency. Refining action proposals from computationally lightweight policies can combine these strengths, but the amount of refinement needed varies across states. We introduce RAIN (Refine Actions If Needed), a framework that predicts action errors at different refinement levels to decide whether and how much to refine each executable proposal, before performing any refinement. We train an ensemble to predict these errors by comparing action proposals and their refined versions with expert actions from offline demonstrations, without requiring explicit labels of refinement needs. Calibration balances uncertainty, computation, and temporal consistency. We evaluate RAIN on eight ManiSkill tasks and four real-robot tasks. In simulation, our policy achieves 66.6% success, approaching the best evaluated diffusion-policy success rate with 5.71 times lower policy latency. On the real robot, success increases from 55.0% with Action Chunking with Transformers (ACT) to 72.5%, with comparable average successful completion time. Matched controls show that dynamic routing maintains the strongest observed fixed-refinement success level with 26.6% fewer refiner evaluations. Project page: https://rain-policy.github.io/.

cs.RO↗

arXiv2026

No Concept Escapes the Audit: Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models

Text-to-image diffusion models can generate prohibited content, which motivates concept erasure through machine unlearning. Most erasure methods intervene at the text interface, through prompt modification or localized updates to text-conditioning weights, and they are evaluated by what the model outputs for given prompts. Such evaluation cannot see what the network still encodes. Latent-space auditing, which bypasses text conditioning and probes the denoising network directly, shows that erased concepts remain recoverable from internal representations. We find that this also holds for methods built to be robust against adversarial prompts, and that the problem grows with the number of erased concepts. We propose Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models (AVCE), a framework that grounds erasure in the model's latent representations. AVCE audits the embedding neighborhood of each concept and condenses the discovered vulnerable directions into an anchor at the weakest geometric point. It edits cross-attention and self-attention projections in closed form at this anchor, then fine-tunes the two pathways with pathway-level auditing losses, using orthogonal gradient projection to consolidate multiple concepts. Experiments on SD v1.5, SDXL, and Flux 1.0 across object, explicit-content, and artistic-style unlearning show that AVCE reduces attack success rates by 5.07x and improves auditing scores by 3.84x over the strongest baseline, while preserving competitive generation quality.

cs.LG↗

arXiv2026

$Δ$ Flow Matching for Robot Foundation Models

Flow matching is a common generative recipe for robot foundation models. Vision-language-action models (VLAs) generate action chunks, while world-action models (WAMs) additionally generate future visual states. Both start generation from a fixed Gaussian prior, despite temporal continuity between past and future states and actions. This requires costly training and iterative sampling, especially for WAMs with an additional visual branch. We introduce $Δ$-flow, which constructs the source from recent history with mild noise rather than from a fixed Gaussian prior. This moves the source closer to the target, reducing transport to a small residual $Δ$ that can be learned efficiently and enables single-step inference. We integrate $Δ$-flow into the VLA policy $π_{0.5}$ and the WAM policies ImageWAM and Fast-WAM. On LIBERO, ImageWAM reaches 90% success in one epoch rather than six, and Fast-WAM reaches 95% in two epochs rather than eight. With LoRA fine-tuning, $π_{0.5}$ exceeds its baseline's five-epoch performance after one epoch. Single-step inference gives $1.54\times$, $2.49\times$, and $4.45\times$ speedups on $π_{0.5}$, ImageWAM, and Fast-WAM, respectively. It preserves action smoothness, achieving a $4.5\times$ reduction in action jerk compared with one-step Gaussian sampling. $Δ$-flow reaches state-of-the-art generalization performance on LIBERO-Plus. Action-history sources can be applied to pretrained WAM policies without retraining. Replacing the Gaussian action source at inference largely preserves task success with a single sampling step and can slightly improve it. We further study how action representation and visual information in action learning affect history-to-future prediction in $Δ$-flow. These results show that $Δ$-flow is simple yet effective, improving the efficiency and generalization of robot foundation models.

cs.RO↗

arXiv2026

Mind Your Own Role: Reducing Role Overstepping for Effective and Efficient Multi-Agent Collaboration via Dynamic Role Activation Steering

Multi-Agent Systems (MAS) have shown strong capability in solving complex tasks by assigning different agents specialized roles and corresponding subtasks. However, agents may overstep their assigned responsibilities and perform subtasks belonging to other roles, which can disrupt workflow, reduce system efficiency, and degrade task performance. To address this issue, we investigate role-related patterns in the model's internal activation space and uncover two key observations: (i) role instructions induce consistent directional shifts in activation space, and (ii) role-adherent and role-overstepping trajectories exhibit distinct activation patterns. Based on these observations, we propose Dynamic Role Activation Steering (DRAS), a lightweight and effective inference-time activation intervention approach to reduce role overstepping in multi-agent collaboration. Specifically, DRAS constructs a role steering vector and a lightweight role-overstepping detector for each role. During LLM generation, the detector estimates the probability of role overstepping from the current activation state, and DRAS dynamically adjusts the steering strength accordingly to guide the agent toward its assigned responsibilities. Experimental results show that DRAS effectively reduces the number of role-overstepping trajectories by up to 86.6% compared with Vanilla MAS, thereby improving task quality score by 10.0%--11.1%, achieving a 1.50$\times$--1.61$\times$ task-completion speedup, and reducing token consumption by 26.8%--38.7%.

cs.MA↗

arXiv2026

Compositional Concepts Erasing: Disentangling Object-Style via Domain Manifold Unlearning

Diffusion models pose significant risks regarding the generation of illicit content and copyright violations. This has catalyzed the development of machine unlearning to erase harmful concepts or copyrighted styles from a model's generative repertoire without compromising its general performance. However, previous unlearning approaches on diffusion models have largely focused on either object-level or style-level removal and failed to address the challenge of object-style combination unlearning. At the same time, multiple unlearning requests often lead to catastrophic forgetting. To address these challenges, we propose Domain Manifold Unlearning (DMU), which uses an EMA-updated feature proxy extracted from intermediate diffusion features to represent the retained target-object direction. As a lightweight estimating representation of an object, DMU enables selective unlearning of domain-specific components in a specific object while preserving other semantic representations, thereby significantly enhancing robustness in multiple combination unlearning scenarios. Extensive experiments show that DMU significantly mitigates catastrophic forgetting, outperforming state-of-the-art methods of batch unlearning and sequential unlearning.

cs.CV↗

arXiv2026

ReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric Video

Egocentric devices, such as wearable front-facing cameras, provide a unique perspective for capturing the continuous interaction between a human viewer and the surrounding environment. A holistic and efficient multimodal model capable of reconstructing this 4D representation is therefore highly desirable. However, existing approaches often rely on auxiliary inputs such as pre-computed camera trajectories, treat scene perception and human ego-motion modeling as separate problems despite their strong interdependency, and suffer from slow inference time. To address these limitations, we present ReViV, the first unified framework for holistic egocentric 4D reconstruction that extracts both viewer and view dynamics from a single monocular RGB video. We formulate the task as learning the full joint probability distribution over multimodal signals, including RGB video, camera trajectory, gaze direction, full-body motion, hand motion, and depth. Powered by a Masked Generative Egocentric Transformer, ReViV operates within a single feed-forward architecture to simultaneously reconstruct the temporally consistent 4D reconstruction across the viewer and the view with fast inference speed. Extensive experiments on diverse benchmarks, including HoloAssist, HOT3D, ARCTIC, Aria Digital Twin, and TACO, demonstrate that ReViV achieves state-of-the-art accuracy and efficiency across holistic ego-body, hand, and gaze reconstruction, camera tracking, while maintaining highly competitive egocentric depth estimation without relying on heavy task-specific priors. Code and models are fully open-sourced: https://reviv4d.github.io/.

cs.CV↗

arXiv2026

FLASH: Efficient Visuomotor Policy via Sparse Sampling

Generative models such as diffusion and flow matching have become dominant paradigms for visuomotor policy learning, yet their reliance on iterative denoising incurs high inference latency incompatible with real-time robotic control. We present Fast Legendre-polynomial Action policy via Sparse History-anchored flow (FLASH Policy), which replaces discrete action-chunk generation with continuous Legendre polynomial trajectory representation. Specifically, by fitting expert demonstrations under sparse temporal sampling, FLASH enables a single inference to cover a significantly extended action horizon. To further accelerate generation, FLASH initiates the flow matching process from history polynomial coefficients rather than uninformative Gaussian noise, shortening the transport distance and enabling accurate single-step inference. Moreover, analytic polynomial differentiation directly provides desired velocity feed-forward signals to the torque controller without numerical approximation. Extensive experiments on five simulated and two real-world manipulation tasks demonstrate that FLASH achieves state-of-the-art success rates ($\ge 92\%$ across all tasks), a per-episode inference time of $31.40\,ms$ (up to $175\times$ faster than diffusion policies and $18\times$ faster than prior flow matching policies), up to $4\times$ faster training convergence than ACT, and $5\times$ to $7\times$ reduction in controller tracking error compared to discrete-action baselines.

cs.RO↗

arXiv2026

Rethinking Fixed Temporal Grids: Frequency-Disentangled Motion Generation

Most human motion generation methods encode motion as tokens on a uniform temporal grid, where every token spans the same fixed time window. Human motion, however, is temporally heterogeneous: slowly evolving global trajectories coexist with rapid transient events such as foot contacts and joint impulses. Forcing such multi-scale dynamics onto tokens of identical temporal resolution entangles motion frequencies, leaving slow regions redundant while smoothing out the rapid details that distinguish realistic motion. We propose \textbf{FreqMo}, a scale-adaptive motion representation that decomposes motion into wavelet frequency bands, separating dynamics across temporal scales while preserving temporal localization and exact reconstruction. Unified Frequency Residual Quantization (UFRQ) then encodes all bands within a single shared codebook, compressing the token sequence threefold and enabling stable single-stage generation. Experiments show FreqMo attains SOTA fidelity with substantially improved high-frequency preservation, and the same decomposition transfers to continuous diffusion backbones.

cs.CV↗

arXiv2026

mmHRI: Towards Privacy-Preserving Human-Robot Interaction with Millimeter-Wave Radar

Assistive robots increasingly operate in many human-centered environments and perform various human-robot interaction (HRI) tasks, such as object delivery. However, most existing HRI systems rely on RGB cameras that continuously observe humans to respond to non-verbal commands, such as hand gestures. This raises privacy concerns in privacy- critical environments, such as hospital wards or restaurants, where direct camera observation of humans is restricted. To develop privacy-preserving HRI, we leverage millimeter-wave (mmWave) radar, which can sense human motion through privacy barriers without identifiable imagery. We propose mmHRI, the first multi-modal robot manipulation framework that achieves mmWave radar-guided privacy-preserving HRI. mmHRI introduces two key designs to mitigate the sparsity and temporal inconsistency of radar data in cluttered robot manipulation environments. First, we propose a dual-stream architecture that jointly learns from unfiltered raw radar tensors and radar point clouds to estimate both human actions and 3D poses. To mitigate signal inconsistency, mmHRI further incorporates a memory-based state-space model (MSSM) that retains historical radar features to reduce abrupt changes in pose/action. These estimated human states are then converted into structured textual robot instructions, which control a vision-language-action (VLA) policy for closed-loop robot manipulation and human-aware reactions. Our evaluation covers human action recognition and closed-loop delivery and retrieval. In the privacy-preserving curtain setting, mmHRI achieves 85.09% action-recognition accuracy, outperforming existing radar-based alternatives. Robot trials further demonstrate successful delivery and retrieval under visual occlusion, with stable task performance across unseen subjects, clutter configurations, and environments.

cs.RO↗