Dongbin Zhao

Yupeng Zheng, Chengliang Zhong, Pengfei Li et al.

Self-supervised depth estimation draws a lot of attention recently as it can promote the 3D sensing capabilities of self-driving vehicles. However, it intrinsically relies upon the photometric consistency assumption, which hardly holds during nighttime. Although various supervised nighttime image enhancement methods have been proposed, their generalization performance in challenging driving scenarios is not satisfactory. To this end, we propose the first method that jointly learns a nighttime image enhancer and a depth estimator, without using ground truth for either task. Our method tightly entangles two self-supervised tasks using a newly proposed uncertain pixel masking strategy. This strategy originates from the observation that nighttime images not only suffer from underexposed regions but also from overexposed regions. By fitting a bridge-shaped curve to the illumination map distribution, both regions are suppressed and two tasks are bridged naturally. We benchmark the method on two established datasets: nuScenes and RobotCar and demonstrate state-of-the-art performance on both of them. Detailed ablations also reveal the mechanism of our proposal. Last but not least, to mitigate the problem of sparse ground truth of existing datasets, we provide a new photo-realistically enhanced nighttime dataset based upon CARLA. It brings meaningful new challenges to the community. Codes, data, and models are available at https://github.com/ucaszyp/STEPS.

Jiangran Lyu, Kai Liu, Xuheng Zhang et al.

Recent robot foundation models largely rely on large-scale behavior cloning, which imitates expert actions but discards transferable dynamics knowledge embedded in heterogeneous embodied data. While the Unified World Model (UWM) formulation has the potential to leverage such diverse data, existing instantiations struggle to scale to foundation-level due to coarse data usage and fragmented datasets. We introduce LDA-1B, a robot foundation model that scales through universal embodied data ingestion by jointly learning dynamics, policy, and visual forecasting, assigning distinct roles to data of varying quality. To support this regime at scale, we assemble and standardize EI-30k, an embodied interaction dataset comprising over 30k hours of human and robot trajectories in a unified format. Scalable dynamics learning over such heterogeneous data is enabled by prediction in a structured DINO latent space, which avoids redundant pixel-space appearance modeling. Complementing this representation, LDA-1B employs a multi-modal diffusion transformer to handle asynchronous vision and action streams, enabling stable training at the 1B-parameter scale. Experiments in simulation and the real world show LDA-1B outperforms prior methods (e.g., $π_{0.5}$) by up to 21\%, 48\%, and 23\% on contact-rich, dexterous, and long-horizon tasks, respectively. Notably, LDA-1B enables data-efficient fine-tuning, gaining 10\% by leveraging 30\% low-quality trajectories typically harmful and discarded.

Zijie Zhao, Honglei Guo, Shengqian Chen et al.

Model-based reinforcement learning (MBRL) has achieved remarkable success in robotics due to its high sample efficiency and planning capability. However, extending MBRL to physical multi-robot cooperation remains challenging due to the complexity of joint dynamics. To address this challenge, we propose the Sequential World Model (SeqWM), a novel framework that integrates the sequential paradigm into multi-robot MBRL. SeqWM employs independent, autoregressive agent-wise world models to represent joint dynamics, where each agent generates its future trajectory and plans its actions based on the predictions of its predecessors. This design lowers modeling complexity and enables the emergence of advanced cooperative behaviors through explicit intention sharing. Experiments on Bi-DexHands and Multi-Quadruped demonstrate that SeqWM outperforms existing state-of-the-art model-based and model-free baselines in both overall performance and sample efficiency, while exhibiting advanced cooperative behaviors such as predictive adaptation, temporal alignment, and role division. Furthermore, SeqWM has been successfully deployed on physical quadruped robots, validating its effectiveness in real-world multi-robot systems. Demos and code are available at: https://github.com/zhaozijie2022/seqwm

Minghui Jia, Qichao Zhang, Ali Luo et al.

Due to the limited generalization and interpretability of deep learning classifiers, The final vetting of rare celestial object candidates still relies on expert visual inspection--a manually intensive process. In this process, astronomers leverage specialized tools to analyze spectra and construct reliable catalogs. However, this practice has become the primary bottleneck, as it is fundamentally incapable of scaling with the data deluge from modern spectroscopic surveys. To bridge this gap, we propose Spec-o3, a tool-augmented vision-language agent that performs astronomer-aligned spectral inspection via interleaved multimodal chain-of-thought reasoning. Spec-o3 is trained with a two-stage post-training recipe: cold-start supervised fine-tuning on expert inspection trajectories followed by outcome-based reinforcement learning on rare-type verification tasks. Evaluated on five rare-object identification tasks from LAMOST, Spec-o3 establishes a new State-of-the-Art, boosting the macro-F1 score from 28.3 to 76.5 with a 7B parameter base model and outperforming both proprietary VLMs and specialized deep models. Crucially, the agent demonstrates strong generalization to unseen inspection tasks across survey shifts (from LAMOST to SDSS/DESI). Expert evaluations confirm that its reasoning traces are coherent and physically consistent, supporting transparent and trustworthy decision-making. Code, data, and models are available at https://github.com/Maxwell-Jia/spec-o3.

Jiajun Chai, Wenzhang Chen, Yuanheng Zhu et al.

Unmanned combat air vehicle (UCAV) combat is a challenging scenario with continuous action space. In this paper, we propose a general hierarchical framework to resolve the within-vision-range (WVR) air-to-air combat problem under 6 dimensions of degree (6-DOF) dynamics. The core idea is to divide the whole decision process into two loops and use reinforcement learning (RL) to solve them separately. The outer loop takes into account the current combat situation and decides the expected macro behavior of the aircraft according to a combat strategy. Then the inner loop tracks the macro behavior with a flight controller by calculating the actual input signals for the aircraft. We design the Markov decision process for both the outer loop strategy and inner loop controller, and train them by proximal policy optimization (PPO) algorithm. For the inner loop controller, we design an effective reward function to accurately track various macro behavior. For the outer loop strategy, we further adopt a fictitious self-play mechanism to improve the combat performance by constantly combating against the historical strategies. Experiment results show that the inner loop controller can achieve better tracking performance than fine-tuned PID controller, and the outer loop strategy can perform complex maneuvers to get higher and higher winning rate, with the generation evolves.

Yuqian Fu, Haohuan Huang, Kaiwen Jiang et al.

On-policy distillation (OPD) is appealing for large language model (LLM) post-training because it evaluates teacher feedback on student-generated rollouts rather than fixed teacher traces. In long-horizon settings, however, the common sampled-token variant is fragile: it reduces distribution matching to a one-token signal and becomes increasingly unreliable as rollouts drift away from prefixes the teacher commonly visits. We revisit OPD from the estimator and implementation sides. Theoretically, token-level OPD is biased relative to sequence-level reverse-KL, but it has a much tighter worst-case variance bound; our toy study shows the same tradeoff empirically, with stronger future-reward coupling producing higher gradient variance and less stable learning. Empirically, we identify three failure modes of sampled-token OPD: an imbalanced one-token signal, unreliable teacher guidance on student-generated prefixes, and distortions caused by tokenizer or special-token mismatch. We address these issues with teacher top-K local support matching, implemented as truncated reverse-KL with top-p rollout sampling and special-token masking. Across single-task math reasoning and multi-task agentic-plus-math training, this objective yields more stable optimization and better downstream performance than sampled-token OPD.

Yuan Liu, Haoran Li, Shuai Tian et al.

Pretrained on large-scale and diverse datasets, VLA models demonstrate strong generalization and adaptability as general-purpose robotic policies. However, Supervised Fine-Tuning (SFT), which serves as the primary mechanism for adapting VLAs to downstream domains, requires substantial amounts of task-specific data and is prone to catastrophic forgetting. To address these limitations, we propose LifeLong-RFT, a simple yet effective Reinforcement Fine-Tuning (RFT) strategy for VLA models independent of online environmental feedback and pre-trained reward models. By integrating chunking-level on-policy reinforcement learning with the proposed multi-dimensional process reward mechanism, LifeLong-RFT quantifies the heterogeneous contributions of intermediate action chunks across three dimensions to facilitate policy optimization. Specifically, (1) the Quantized Action Consistency Reward (QACR) ensures accurate action prediction within the discrete action space; (2) the Continuous Trajectory Alignment Reward (CTAR) aligns decoded continuous action chunks with reference trajectories to ensure precise control; (3) the Format Compliance Reward (FCR) guarantees the structural validity of outputs. Comprehensive experiments across SimplerEnv, LIBERO, and real-world tasks demonstrate that LifeLong-RFT exhibits strong performance in multi-task learning. Furthermore, for continual learning on the LIBERO benchmark, our method achieves a 22% gain in average success rate over SFT, while effectively adapting to new tasks using only 20% of the training data. Overall, our method provides a promising post-training paradigm for VLAs. The project page is available at <https://yuan-liu-lifelong-rft.github.io>.

Xin Liu, Yaran Chen, Dongbin Zhao

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Unsupervised skill discovery seeks to acquire different useful skills without extrinsic reward via unsupervised Reinforcement Learning (RL), with the discovered skills efficiently adapting to multiple downstream tasks in various ways. However, recent advanced skill discovery methods struggle to well balance state exploration and skill diversity, particularly when the potential skills are rich and hard to discern. In this paper, we propose \textbf{Co}ntrastive dyna\textbf{m}ic \textbf{S}kill \textbf{D}iscovery \textbf{(ComSD)}\footnote{Code and videos: https://github.com/liuxin0824/ComSD} which generates diverse and exploratory unsupervised skills through a novel intrinsic incentive, named contrastive dynamic reward. It contains a particle-based exploration reward to make agents access far-reaching states for exploratory skill acquisition, and a novel contrastive diversity reward to promote the discriminability between different skills. Moreover, a novel dynamic weighting mechanism between the above two rewards is proposed to balance state exploration and skill diversity, which further enhances the quality of the discovered skills. Extensive experiments and analysis demonstrate that ComSD can generate diverse behaviors at different exploratory levels for multi-joint robots, enabling state-of-the-art adaptation performance on challenging downstream tasks. It can also discover distinguishable and far-reaching exploration skills in the challenging tree-like 2D maze.

Yaocheng Zhang, Jiajun Chai, Songjun Tu et al.

On-policy distillation (OPD) provides dense teacher feedback along rollouts generated by the student and has emerged as a promising post-training paradigm for long-horizon reasoning. However, standard OPD typically generates full rollouts during training, which is computationally expensive and may expose the student to unreliable teacher feedback at late rollout positions, especially during early training. We identify the rollout horizon as a key bottleneck in OPD that substantially impacts training efficiency. Unlike Reinforcement Learning with Verifiable Rewards (RLVR), OPD does not require a complete trajectory or a final answer reward to provide learning signals. This observation suggests that full rollouts may not always be necessary for effective OPD. Motivated by this insight, we propose two simple horizon-control strategies: Progressive OPD (POPD), which gradually expands the rollout horizon during training, and Truncated OPD (TOPD), which permanently performs distillation on reliable truncated rollouts. Experiments on mathematical reasoning show that POPD improves the training efficiency of OPD by up to 3$\times$, while TOPD matches OPD performance using only 10\% of the rollout horizon, leading to substantial wall-clock and memory reductions. These results demonstrate that controlling the rollout horizon offers a simple and practical path to more efficient OPD.

Ding Li, Qichao Zhang, Shuai Lu et al.

This paper aims to tackle the interactive behavior prediction task, and proposes a novel Conditional Goal-oriented Trajectory Prediction (CGTP) framework to jointly generate scene-compliant trajectories of two interacting agents. Our CGTP framework is an end to end and interpretable model, including three main stages: context encoding, goal interactive prediction and trajectory interactive prediction. First, a Goals-of-Interest Network (GoINet) is designed to extract the interactive features between agent-to-agent and agent-to-goals using a graph-based vectorized representation. Further, the Conditional Goal Prediction Network (CGPNet) focuses on goal interactive prediction via a combined form of marginal and conditional goal predictors. Finally, the Goaloriented Trajectory Forecasting Network (GTFNet) is proposed to implement trajectory interactive prediction via the conditional goal-oriented predictors, with the predicted future states of the other interacting agent taken as inputs. In addition, a new goal interactive loss is developed to better learn the joint probability distribution over goal candidates between two interacting agents. In the end, the proposed method is conducted on Argoverse motion forecasting dataset, In-house cut-in dataset, and Waymo open motion dataset. The comparative results demonstrate the superior performance of our proposed CGTP model than the mainstream prediction methods.

Songjun Tu, Chengdong Xu, Qichao Zhang et al.

Agentic reinforcement learning (RL) can benefit substantially from reusable experience, yet existing skill-based methods mainly extract trajectory-level guidance and often lack principled mechanisms for maintaining an evolving skill memory. We propose D2Skill, a dynamic dual-granularity skill bank for agentic RL that organizes reusable experience into task skills for high-level guidance and step skills for fine-grained decision support and error correction. D2Skill jointly trains the policy and skill bank through paired baseline and skill-injected rollouts under the same policy, using their performance gap to derive hindsight utility signals for both skill updating and policy optimization. Built entirely from training-time experience, the skill bank is continuously expanded through reflection and maintained with utility-aware retrieval and pruning. Experiments on ALFWorld and WebShop with Qwen2.5-7B-Instruct and Qwen3-4B-Instruct-2507 show that D2Skill consistently improves success rates over skill-free baselines by 10-20 points. Further ablations and analyses show that both dual-granularity skill modeling and dynamic skill maintenance are critical to these gains, while the learned skills exhibit higher utility, transfer across evaluation settings, and introduce only modest training overhead.

Junjie Wang, Yao Mu, Dong Li et al.

The latent world model provides a promising way to learn policies in a compact latent space for tasks with high-dimensional observations, however, its generalization across diverse environments with unseen dynamics remains challenging. Although the recurrent structure utilized in current advances helps to capture local dynamics, modeling only state transitions without an explicit understanding of environmental context limits the generalization ability of the dynamics model. To address this issue, we propose a Prototypical Context-Aware Dynamics (ProtoCAD) model, which captures the local dynamics by time consistent latent context and enables dynamics generalization in high-dimensional control tasks. ProtoCAD extracts useful contextual information with the help of the prototypes clustered over batch and benefits model-based RL in two folds: 1) It utilizes a temporally consistent prototypical regularizer that encourages the prototype assignments produced for different time parts of the same latent trajectory to be temporally consistent instead of comparing the features; 2) A context representation is designed which combines both the projection embedding of latent states and aggregated prototypes and can significantly improve the dynamics generalization ability. Extensive experiments show that ProtoCAD surpasses existing methods in terms of dynamics generalization. Compared with the recurrent-based model RSSM, ProtoCAD delivers 13.2% and 26.7% better mean and median performance across all dynamics generalization tasks.

Yaocheng Zhang, Yuanheng Zhu, Wenyue Chong et al.

Deep search agents have emerged as a promising paradigm for addressing complex information-seeking tasks, but their training remains challenging due to sparse rewards, weak credit assignment, and limited labeled data. Self-play offers a scalable route to reduce data dependence, but conventional self-play optimizes students only through sparse outcome rewards, leading to low learning efficiency. In this work, we observe that self-play naturally produces a question construction path (QCP) during task generation, an intermediate artifact that captures the reverse solution process. This reveals a new source of privileged information for self-distillation: self-play can itself provide high-quality privileged context for the teacher model in a low-cost and scalable manner, without relying on human feedback or curated privileged information. Leveraging this insight, we propose Privileged Information Self-Play ($π$-Play), a multi-agent self-evolution framework. In $π$-Play, an examiner generates tasks together with their QCPs, and a teacher model leverages QCP as privileged context to densely supervise a student via self-distillation. This design transforms conventional sparse-reward self-play into a dense-feedback self-evolution loop. Extensive experiments show that data-free $π$-Play surpasses fully supervised search agents and improves evolutionary efficiency by 2-3$\times$ over conventional self-play.

Zebin Xing, Yupeng Zheng, Qichao Zhang et al.

End-to-end autonomous driving has emerged as a pivotal direction in the field of autonomous systems. Recent works have demonstrated impressive performance by incorporating high-level guidance signals to steer low-level trajectory planners. However, their potential is often constrained by inaccurate high-level guidance and the computational overhead of complex guidance modules. To address these limitations, we propose Mimir, a novel hierarchical dual-system framework capable of generating robust trajectories relying on goal points with uncertainty estimation: (1) Unlike previous approaches that deterministically model, we estimate goal point uncertainty with a Laplace distribution to enhance robustness; (2) To overcome the slow inference speed of the guidance system, we introduce a multi-rate guidance mechanism that predicts extended goal points in advance. Validated on challenging Navhard and Navtest benchmarks, Mimir surpasses previous state-of-the-art methods with a 20% improvement in the driving score EPDMS, while achieving 1.6 times improvement in high-level module inference speed without compromising accuracy. The code and models will be released soon to promote reproducibility and further development. The code is available at https://github.com/ZebinX/Mimir-Uncertainty-Driving

Runyu Lu, Peng Zhang, Ruochuan Shi et al.

Equilibrium learning in adversarial games is an important topic widely examined in the fields of game theory and reinforcement learning (RL). Pursuit-evasion game (PEG), as an important class of real-world games from the fields of robotics and security, requires exponential time to be accurately solved. When the underlying graph structure varies, even the state-of-the-art RL methods require recomputation or at least fine-tuning, which can be time-consuming and impair real-time applicability. This paper proposes an Equilibrium Policy Generalization (EPG) framework to effectively learn a generalized policy with robust cross-graph zero-shot performance. In the context of PEGs, our framework is generally applicable to both pursuer and evader sides in both no-exit and multi-exit scenarios. These two generalizability properties, to our knowledge, are the first to appear in this domain. The core idea of the EPG framework is to train an RL policy across different graph structures against the equilibrium policy for each single graph. To construct an equilibrium oracle for single-graph policies, we present a dynamic programming (DP) algorithm that provably generates pure-strategy Nash equilibrium with near-optimal time complexity. To guarantee scalability with respect to pursuer number, we further extend DP and RL by designing a grouping mechanism and a sequence model for joint policy decomposition, respectively. Experimental results show that, using equilibrium guidance and a distance feature proposed for cross-graph PEG training, the EPG framework guarantees desirable zero-shot performance in various unseen real-world graphs. Besides, when trained under an equilibrium heuristic proposed for the graphs with exits, our generalized pursuer policy can even match the performance of the fine-tuned policies from the state-of-the-art PEG methods.

Yuanyang Zhu, Zhi Wang, Yuanheng Zhu et al.

For on-policy reinforcement learning, discretizing action space for continuous control can easily express multiple modes and is straightforward to optimize. However, without considering the inherent ordering between the discrete atomic actions, the explosion in the number of discrete actions can possess undesired properties and induce a higher variance for the policy gradient estimator. In this paper, we introduce a straightforward architecture that addresses this issue by constraining the discrete policy to be unimodal using Poisson probability distributions. This unimodal architecture can better leverage the continuity in the underlying continuous action space using explicit unimodal probability distributions. We conduct extensive experiments to show that the discrete policy with the unimodal probability distribution provides significantly faster convergence and higher performance for on-policy reinforcement learning algorithms in challenging control tasks, especially in highly complex tasks such as Humanoid. We provide theoretical analysis on the variance of the policy gradient estimator, which suggests that our attentively designed unimodal discrete policy can retain a lower variance and yield a stable learning process.

Xin Liu, Yaran Chen, Haoran Li et al.

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Unsupervised cross-domain Reinforcement Learning (RL) pre-training shows great potential for challenging continuous visual control but poses a big challenge. In this paper, we propose \textbf{C}ross-domain \textbf{R}andom \textbf{P}re-\textbf{T}raining with \textbf{pro}totypes (CRPTpro), a novel, efficient, and effective self-supervised cross-domain RL pre-training framework. CRPTpro decouples data sampling from encoder pre-training, proposing decoupled random collection to easily and quickly generate a qualified cross-domain pre-training dataset. Moreover, a novel prototypical self-supervised algorithm is proposed to pre-train an effective visual encoder that is generic across different domains. Without finetuning, the cross-domain encoder can be implemented for challenging downstream tasks defined in different domains, either seen or unseen. Compared with recent advanced methods, CRPTpro achieves better performance on downstream policy learning without extra training on exploration agents for data collection, greatly reducing the burden of pre-training. We conduct extensive experiments across eight challenging continuous visual-control domains, including balance control, robot locomotion, and manipulation. CRPTpro significantly outperforms the next best Proto-RL(C) on 11/12 cross-domain downstream tasks with only 54.5\% wall-clock pre-training time, exhibiting state-of-the-art pre-training performance with greatly improved pre-training efficiency.

Linbo Wang, Yupeng Zheng, Qiang Chen et al. · tsinghua

We introduce Latent-WAM, an efficient end-to-end autonomous driving framework that achieves strong trajectory planning through spatially-aware and dynamics-informed latent world representations. Existing world-model-based planners suffer from inadequately compressed representations, limited spatial understanding, and underutilized temporal dynamics, resulting in sub-optimal planning under constrained data and compute budgets. Latent-WAM addresses these limitations with two core modules: a Spatial-Aware Compressive World Encoder (SCWE) that distills geometric knowledge from a foundation model and compresses multi-view images into compact scene tokens via learnable queries, and a Dynamic Latent World Model (DLWM) that employs a causal Transformer to autoregressively predict future world status conditioned on historical visual and motion representations. Extensive experiments on NAVSIM v2 and HUGSIM demonstrate new state-of-the-art results: 89.3 EPDMS on NAVSIM v2 and 28.9 HD-Score on HUGSIM, surpassing the best prior perception-free method by 3.2 EPDMS with significantly less training data and a compact 104M-parameter model.

Jingbo Sun, Wenyue Chong, Songjun Tu et al.

Agentic retrieval-augmented generation (RAG) systems enable large language models (LLMs) to solve complex tasks through multi-step interaction with external retrieval tools. However, such multi-step interaction often involves redundant search steps, incurring substantial computational cost and latency. Prior work limits search depth (i.e., the number of search steps) to reduce cost, but this often leads to underexploration of complex questions. To address this, we first investigate how search depth affects accuracy and find a minimal sufficient search depth that defines an accuracy-efficiency trade-off, jointly determined by question complexity and the agent's capability. Furthermore, we propose AutoSearch, a reinforcement learning (RL) framework that evaluates each search step via self-generated intermediate answers. By a self-answering mechanism, AutoSearch identifies the minimal sufficient search depth and promotes efficient search by rewarding its attainment while penalizing over-searching. In addition, reward mechanisms are introduced to stabilize search behavior and improve answer quality on complex questions. Extensive experiments on multiple benchmarks show that AutoSearch achieves a superior accuracy-efficiency trade-off, alleviating over-searching while preserving search quality.

Yuhui Chen, Haoran Li, Dongbin Zhao

Due to its training stability and strong expression, the diffusion model has attracted considerable attention in offline reinforcement learning. However, several challenges have also come with it: 1) The demand for a large number of diffusion steps makes the diffusion-model-based methods time inefficient and limits their applications in real-time control; 2) How to achieve policy improvement with accurate guidance for diffusion model-based policy is still an open problem. Inspired by the consistency model, we propose a novel time-efficiency method named Consistency Policy with Q-Learning (CPQL), which derives action from noise by a single step. By establishing a mapping from the reverse diffusion trajectories to the desired policy, we simultaneously address the issues of time efficiency and inaccurate guidance when updating diffusion model-based policy with the learned Q-function. We demonstrate that CPQL can achieve policy improvement with accurate guidance for offline reinforcement learning, and can be seamlessly extended for online RL tasks. Experimental results indicate that CPQL achieves new state-of-the-art performance on 11 offline and 21 online tasks, significantly improving inference speed by nearly 45 times compared to Diffusion-QL. We will release our code later.

Boyu Li, Chaoyi Xu, Haoqi Yuan et al.

Learning universal policies from cross-embodied data remains a fundamental challenge in robotics. Although Vision-Language-Action (VLA) models are pre-trained on large and diverse datasets, they typically rely on embodiment-specific fine-tuning to achieve strong performance in downstream tasks. This requirement severely limits their generalization capability and restricts knowledge transfer across embodiments performing similar tasks. To overcome these limitations, we focus on cross-embodied settings with shared robotic bases and heterogeneous end-effectors, and propose X-DiffVLA, a diffusion-based VLA model featuring a unified cross-embodied action head. X-DiffVLA can leverage the generative strengths of diffusion models to capture both the diversity and latent correlations in cross-embodied datasets. Specifically, we introduce Embodiment Forcing, a classifier-free guidance technique to implicitly steer action generation toward embodiment-specific functional components, capturing fine-grained structural nuances without explicit supervision. In addition, a Morphological Tree Diffusion approach is designed to strengthen behavioral correlations across diverse end-effectors, maximizing the transferability of heterogeneous demonstrations. Experimental results across RoboCasa and Isaac Gym, covering different embodiments from grippers to dexterous hands, show that X-DiffVLA achieves state-of-the-art performance, with improvements of 15.3% and 12.5%, respectively. Real-world evaluations further validate the robustness of the proposed framework and its effectiveness in scalable cross-embodied policy learning.

Xian Wu, Haoran Li, Dongbin Zhao et al.

Reinforcement learning offers a promising approach for scan-order optimisation in laser additive manufacturing, where sequential scan decisions critically influence thermal accumulation, residual stress, distortion, and final part quality. A central challenge in applying RL to this domain lies in reward and world-model fidelity: full finite-element analysis is computationally prohibitive for dense in-the-loop evaluation, while cheap thermo-inspired proxy metrics, though efficient, may capture only partial aspects of the true thermo-mechanical objectives. This paper investigates a bilevel Proxy--FEA diagnostic framework for reward and world-model diagnosis in reinforcement-learning-guided scan-order optimisation. The lower level employs lightweight scan-path and thermo-inspired proxies for rapid candidate generation and preliminary policy-side screening, while the upper level utilises sparse Abaqus FEA simulations to provide simulation-based reference labels. The framework is examined on a simplified whole-track heating LDED32 stripe benchmark comprising ten representative scan strategies. Final-cooling residual Mises stress, U3 vertical distortion, and PEEQ plasticity metrics reveal an observed stress--distortion trade-off rather than a single monotonic quality objective. Within the evaluated set, the center_out strategy emerges as a robust compromise candidate, while raster_left_to_right and edge_in form opposing endpoints of the trade-off. Proxy--FEA alignment analysis shows that current cheap path-based metrics predominantly capture distortion-related (U3) behaviour and exhibit only weak correlation with the sparse FEA reference labels. These findings highlight that proxy-only reward designs risk misalignment in future RL training and underscore the value of sparse FEA reference signals for diagnostic-guided reward and world-model refinement prior to large-scale policy optimisation.

Yinfeng Gao, Qichao Zhang, Deqing Liu et al.

End-to-end autonomous driving policies based on Imitation Learning (IL) often struggle in closed-loop execution due to the misalignment between inadequate open-loop training objectives and real driving requirements. While Reinforcement Learning (RL) offers a solution by directly optimizing driving goals via reward signals, the rendering-based training environments introduce the rendering gap and are inefficient due to high computational costs. To overcome these challenges, we present a novel Pseudo-simulation-based RL method for closed-loop end-to-end autonomous driving, PerlAD. Based on offline datasets, PerlAD constructs a pseudo-simulation that operates in vector space, enabling efficient, rendering-free trial-and-error training. To bridge the gap between static datasets and dynamic closed-loop environments, PerlAD introduces a prediction world model that generates reactive agent trajectories conditioned on the ego vehicle's plan. Furthermore, to facilitate efficient planning, PerlAD utilizes a hierarchical decoupled planner that combines IL for lateral path generation and RL for longitudinal speed optimization. Comprehensive experimental results demonstrate that PerlAD achieves state-of-the-art performance on the Bench2Drive benchmark, surpassing the previous E2E RL method by 10.29% in Driving Score without requiring expensive online interactions. Additional evaluations on the DOS benchmark further confirm its reliability in handling safety-critical occlusion scenarios.

Yinfeng Gao, Deqing Liu, Qichao Zhang et al.

Current Vision-Language-Action (VLA) paradigms in end-to-end autonomous driving rely on offline training from static datasets, leaving them vulnerable to distribution shift. Recent post-training methods use takeover data to mitigate this by augmenting the dataset with high-quality expert takeover samples, yet they suffer from two key limitations: supervision restricted to the period after the takeover moments leads to policies with limited safety margins, and passive preference optimization lacks active exploration for optimal performance. In this paper, we propose TakeVLA, a novel VLA post-training framework that overcomes these shortcomings through two complementary innovations. First, we introduce pre-takeover language supervision, which allows the VLA to learn from mistakes proactively. By explicitly teaching the model about what to do in error-prone situations, we cultivate a precautionary mindset that anticipates hazards early and substantially enlarges safety margins. Second, we propose Scenario Dreaming, a reinforcement fine-tuning paradigm that operates in reconstruceted takeover scenarios, encouraging active exploration beyond mere preference fitting. Experiments on the Bench2Drive benchmark demonstrate that TakeVLA achieves state-of-the-art closed-loop performance, surpassing the strong VLA baseline SimLingo by 4.93 in driving score, with an enhanced safety margin as evidenced by an 11.76% increase in average TTC.

Yaran Chen, Xueyu Chen, Yu Han et al.

Non alcoholic fatty liver disease (NAFLD) is the most common cause of chronic liver disease, which can be predicted accurately to prevent advanced fibrosis and cirrhosis. While, a liver biopsy, the gold standard for NAFLD diagnosis, is invasive, expensive, and prone to sampling errors. Therefore, non-invasive studies are extremely promising, yet they are still in their infancy due to the lack of comprehensive research data and intelligent methods for multi-modal data. This paper proposes a NAFLD diagnosis system (DeepFLDDiag) combining a comprehensive clinical dataset (FLDData) and a multi-modal learning based NAFLD prediction method (DeepFLD). The dataset includes over 6000 participants physical examinations, laboratory and imaging studies, extensive questionnaires, and facial images of partial participants, which is comprehensive and valuable for clinical studies. From the dataset, we quantitatively analyze and select clinical metadata that most contribute to NAFLD prediction. Furthermore, the proposed DeepFLD, a deep neural network model designed to predict NAFLD using multi-modal input, including metadata and facial images, outperforms the approach that only uses metadata. Satisfactory performance is also verified on other unseen datasets. Inspiringly, DeepFLD can achieve competitive results using only facial images as input rather than metadata, paving the way for a more robust and simpler non-invasive NAFLD diagnosis.

Yaocheng Zhang, Haohuan Huang, Zijun Song et al.

Tool-Integrated Reasoning (TIR) with search engines enables large language models to iteratively retrieve up-to-date external knowledge, enhancing adaptability and generalization in complex question-answering tasks. However, existing search agent pipelines typically depend on reinforcement learning based optimization, which often suffers from sparse outcome rewards, leading to inefficient exploration and unstable training. We introduce CriticSearch, a fine-grained credit-assignment framework that supplies dense, turn-level feedback via a retrospective critic mechanism. During training, a frozen, asymmetric critique LLM retrospectively evaluates each turn using privileged information from the full trajectory and gold answers, converting these assessments into stable, dense rewards that guide policy improvement. Experimental results across diverse multi-hop reasoning benchmarks demonstrate that CriticSearch consistently outperforms existing baselines, achieving faster convergence, improved training stability, and higher performance.

Ruochuan Shi, Runyu Lu, Yuanheng Zhu et al.

In graph-structured multi-agent reinforcement learning (MARL) adversarial tasks such as pursuit and confrontation, agents must coordinate under highly dynamic interactions, where sparse rewards hinder efficient policy learning. We propose Adaptive Regularized Multi-Agent Soft Actor-Critic (ARAC), which integrates an attention-based graph neural network (GNN) for modeling agent dependencies with an adaptive divergence regularization mechanism. The GNN enables expressive representation of spatial relations and state features in graph environments. Divergence regularization can serve as policy guidance to alleviate the sparse reward problem, but it may lead to suboptimal convergence when the reference policy itself is imperfect. The adaptive divergence regularization mechanism enables the framework to exploit reference policies for efficient exploration in the early stages, while gradually reducing reliance on them as training progresses to avoid inheriting their limitations. Experiments in pursuit and confrontation scenarios demonstrate that ARAC achieves faster convergence, higher final success rates, and stronger scalability across varying numbers of agents compared with MARL baselines, highlighting its effectiveness in complex graph-structured environments.

Haoran Li, Zhennan Jiang, Yuhui Chen et al.

With high-dimensional state spaces, visual reinforcement learning (RL) faces significant challenges in exploitation and exploration, resulting in low sample efficiency and training stability. As a time-efficient diffusion model, although consistency models have been validated in online state-based RL, it is still an open question whether it can be extended to visual RL. In this paper, we investigate the impact of non-stationary distribution and the actor-critic framework on consistency policy in online RL, and find that consistency policy was unstable during the training, especially in visual RL with the high-dimensional state space. To this end, we suggest sample-based entropy regularization to stabilize the policy training, and propose a consistency policy with prioritized proximal experience regularization (CP3ER) to improve sample efficiency. CP3ER achieves new state-of-the-art (SOTA) performance in 21 tasks across DeepMind control suite and Meta-world. To our knowledge, CP3ER is the first method to apply diffusion/consistency models to visual RL and demonstrates the potential of consistency models in visual RL. More visualization results are available at https://jzndd.github.io/CP3ER-Page/.

Yaran Chen, Wenbo Cui, Yuanwen Chen et al.

Robotic agents must master common sense and long-term sequential decisions to solve daily tasks through natural language instruction. The developments in Large Language Models (LLMs) in natural language processing have inspired efforts to use LLMs in complex robot planning. Despite LLMs' great generalization and comprehension of instruction tasks, LLMs-generated task plans sometimes lack feasibility and correctness. To address the problem, we propose a RoboGPT agent\footnote{our code and dataset will be released soon} for making embodied long-term decisions for daily tasks, with two modules: 1) LLMs-based planning with re-plan to break the task into multiple sub-goals; 2) RoboSkill individually designed for sub-goals to learn better navigation and manipulation skills. The LLMs-based planning is enhanced with a new robotic dataset and re-plan, called RoboGPT. The new robotic dataset of 67k daily instruction tasks is gathered for fine-tuning the Llama model and obtaining RoboGPT. RoboGPT planner with strong generalization can plan hundreds of daily instruction tasks. Additionally, a low-computational Re-Plan module is designed to allow plans to flexibly adapt to the environment, thereby addressing the nomenclature diversity challenge. The proposed RoboGPT agent outperforms SOTA methods on the ALFRED daily tasks. Moreover, RoboGPT planner exceeds SOTA LLM-based planners like ChatGPT in task-planning rationality for hundreds of unseen daily tasks, and even other domain tasks, while keeping the large model's original broad application and generality.

Pengxuan Yang, Yupeng Zheng, Deheng Qian et al.

We introduce DreamerAD, the first latent world model framework that enables efficient reinforcement learning for autonomous driving by compressing diffusion sampling from 100 steps to 1 - achieving 80x speedup while maintaining visual interpretability. Training RL policies on real-world driving data incurs prohibitive costs and safety risks. While existing pixel-level diffusion world models enable safe imagination-based training, they suffer from multi-step diffusion inference latency (2s/frame) that prevents high-frequency RL interaction. Our approach leverages denoised latent features from video generation models through three key mechanisms: (1) shortcut forcing that reduces sampling complexity via recursive multi-resolution step compression, (2) an autoregressive dense reward model operating directly on latent representations for fine-grained credit assignment, and (3) Gaussian vocabulary sampling for GRPO that constrains exploration to physically plausible trajectories. DreamerAD achieves 87.7 EPDMS on NavSim v2, establishing state-of-the-art performance and demonstrating that latent-space RL is effective for autonomous driving.

Zhi Wang, Li Zhang, Wenhao Wu et al.

A longstanding goal of artificial general intelligence is highly capable generalists that can learn from diverse experiences and generalize to unseen tasks. The language and vision communities have seen remarkable progress toward this trend by scaling up transformer-based models trained on massive datasets, while reinforcement learning (RL) agents still suffer from poor generalization capacity under such paradigms. To tackle this challenge, we propose Meta Decision Transformer (Meta-DT), which leverages the sequential modeling ability of the transformer architecture and robust task representation learning via world model disentanglement to achieve efficient generalization in offline meta-RL. We pretrain a context-aware world model to learn a compact task representation, and inject it as a contextual condition to the causal transformer to guide task-oriented sequence generation. Then, we subtly utilize history trajectories generated by the meta-policy as a self-guided prompt to exploit the architectural inductive bias. We select the trajectory segment that yields the largest prediction error on the pretrained world model to construct the prompt, aiming to encode task-specific information complementary to the world model maximally. Notably, the proposed framework eliminates the requirement of any expert demonstration or domain knowledge at test time. Experimental results on MuJoCo and Meta-World benchmarks across various dataset types show that Meta-DT exhibits superior few and zero-shot generalization capacity compared to strong baselines while being more practical with fewer prerequisites. Our code is available at https://github.com/NJU-RL/Meta-DT.

Yupeng Zheng, Pengxuan Yang, Zebin Xing et al.

End-to-end autonomous driving directly generates planning trajectories from raw sensor data, yet it typically relies on costly perception supervision to extract scene information. A critical research challenge arises: constructing an informative driving world model to enable perception annotation-free, end-to-end planning via self-supervised learning. In this paper, we present World4Drive, an end-to-end autonomous driving framework that employs vision foundation models to build latent world models for generating and evaluating multi-modal planning trajectories. Specifically, World4Drive first extracts scene features, including driving intention and world latent representations enriched with spatial-semantic priors provided by vision foundation models. It then generates multi-modal planning trajectories based on current scene features and driving intentions and predicts multiple intention-driven future states within the latent space. Finally, it introduces a world model selector module to evaluate and select the best trajectory. We achieve perception annotation-free, end-to-end planning through self-supervised alignment between actual future observations and predicted observations reconstructed from the latent space. World4Drive achieves state-of-the-art performance without manual perception annotations on both the open-loop nuScenes and closed-loop NavSim benchmarks, demonstrating an 18.1\% relative reduction in L2 error, 46.7% lower collision rate, and 3.75 faster training convergence. Codes will be accessed at https://github.com/ucaszyp/World4Drive.

Pengxuan Yang, Yupeng Zheng, Qichao Zhang et al.

End-to-end autonomous driving aims to produce planning trajectories from raw sensors directly. Currently, most approaches integrate perception, prediction, and planning modules into a fully differentiable network, promising great scalability. However, these methods typically rely on deterministic modeling of online maps in the perception module for guiding or constraining vehicle planning, which may incorporate erroneous perception information and further compromise planning safety. To address this issue, we delve into the importance of online map uncertainty for enhancing autonomous driving safety and propose a novel paradigm named UncAD. Specifically, UncAD first estimates the uncertainty of the online map in the perception module. It then leverages the uncertainty to guide motion prediction and planning modules to produce multi-modal trajectories. Finally, to achieve safer autonomous driving, UncAD proposes an uncertainty-collision-aware planning selection strategy according to the online map uncertainty to evaluate and select the best trajectory. In this study, we incorporate UncAD into various state-of-the-art (SOTA) end-to-end methods. Experiments on the nuScenes dataset show that integrating UncAD, with only a 1.9% increase in parameters, can reduce collision rates by up to 26% and drivable area conflict rate by up to 42%. Codes, pre-trained models, and demo videos can be accessed at https://github.com/pengxuanyang/UncAD.

Zebin Xing, Pengxuan Yang, Linbo Wang et al.

Current autonomous driving systems often favor end-to-end frameworks, which take sensor inputs like images and learn to map them into trajectory space via neural networks. Previous work has demonstrated that models can achieve better planning performance when provided with a prior distribution of possible trajectories. However, these approaches often overlook two critical aspects: 1) The appropriate trajectory prior can vary significantly across different driving scenarios. 2) Their trajectory evaluation mechanism lacks policy-driven refinement, remaining constrained by the limitations of one-stage supervised training. To address these issues, we explore improvements in two key areas. For problem 1, we employ MoE to apply different trajectory priors tailored to different scenarios. For problem 2, we utilize Reinforcement Learning to fine-tune the trajectory scoring mechanism. Additionally, we integrate models with different perception backbones to enhance perceptual features. Our integrated model achieved a score of 51.08 on the navsim ICCV benchmark, securing third place.

Xin Liu, Haoran Li, Dongbin Zhao

Humans can efficiently extract knowledge and learn skills from the videos within only a few trials and errors. However, it poses a big challenge to replicate this learning process for autonomous agents, due to the complexity of visual input, the absence of action or reward signals, and the limitations of interaction steps. In this paper, we propose a novel, unsupervised, and sample-efficient framework to achieve imitation learning from videos (ILV), named Behavior Cloning from Videos via Latent Representations (BCV-LR). BCV-LR extracts action-related latent features from high-dimensional video inputs through self-supervised tasks, and then leverages a dynamics-based unsupervised objective to predict latent actions between consecutive frames. The pre-trained latent actions are fine-tuned and efficiently aligned to the real action space online (with collected interactions) for policy behavior cloning. The cloned policy in turn enriches the agent experience for further latent action finetuning, resulting in an iterative policy improvement that is highly sample-efficient. We conduct extensive experiments on a set of challenging visual tasks, including both discrete control and continuous control. BCV-LR enables effective (even expert-level on some tasks) policy performance with only a few interactions, surpassing state-of-the-art ILV baselines and reinforcement learning methods (provided with environmental rewards) in terms of sample efficiency across 24/28 tasks. To the best of our knowledge, this work for the first time demonstrates that videos can support extremely sample-efficient visual policy learning, without the need to access any other expert supervision.

Deqing Liu, Yinfeng Gao, Deheng Qian et al.

Existing end-to-end autonomous driving methods typically rely on imitation learning (IL) but face a key challenge: the misalignment between open-loop training and closed-loop deployment. This misalignment often triggers driver-initiated takeovers and system disengagements during closed-loop execution. How to leverage those expert takeover data from disengagement scenarios and effectively expand the IL policy's capability presents a valuable yet unexplored challenge. In this paper, we propose TakeAD, a novel preference-based post-optimization framework that fine-tunes the pre-trained IL policy with this disengagement data to enhance the closed-loop driving performance. First, we design an efficient expert takeover data collection pipeline inspired by human takeover mechanisms in real-world autonomous driving systems. Then, this post optimization framework integrates iterative Dataset Aggregation (DAgger) for imitation learning with Direct Preference Optimization (DPO) for preference alignment. The DAgger stage equips the policy with fundamental capabilities to handle disengagement states through direct imitation of expert interventions. Subsequently, the DPO stage refines the policy's behavior to better align with expert preferences in disengagement scenarios. Through multiple iterations, the policy progressively learns recovery strategies for disengagement states, thereby mitigating the open-loop gap. Experiments on the closed-loop Bench2Drive benchmark demonstrate our method's effectiveness compared with pure IL methods, with comprehensive ablations confirming the contribution of each component.

Jiahao Liu, Cui Wenbo, Haoran Li et al.

Mobile manipulation is a fundamental capability for general-purpose robotic agents, requiring both coordinated control of the mobile base and manipulator and robust perception under dynamically changing viewpoints. However, existing approaches face two key challenges: strong coupling between base and arm actions complicates control optimization, and perceptual attention is often poorly allocated as viewpoints shift during mobile manipulation. We propose InCoM, an intent-driven perception and structured coordination framework for mobile manipulation. InCoM infers latent motion intent to dynamically reweight multi-scale perceptual features, enabling stage-adaptive allocation of perceptual attention. To support robust cross-modal perception, InCoM further incorporates a geometric-semantic structured alignment mechanism that enhances multimodal correspondence. On the control side, we design a decoupled coordinated flow matching action decoder that explicitly models coordinated base-arm action generation, alleviating optimization difficulties caused by control coupling. Experimental results demonstrate that InCoM significantly outperforms state-of-the-art methods, achieving success rate gains of 28.2%, 26.1%, and 23.6% across three ManiSkill-HAB scenarios without privileged information. Furthermore, its effectiveness is consistently validated in real-world mobile manipulation tasks, where InCoM maintains a superior success rate over existing baselines.

Xin Liu, Yixuan Li, Yuhui Chen et al.

Designing suitable rewards poses a significant challenge in reinforcement learning (RL), especially for embodied manipulation. Trajectory success rewards are suitable for human judges or model fitting, but the sparsity severely limits RL sample efficiency. While recent methods have effectively improved RL via dense rewards, they rely heavily on high-quality human-annotated data or abundant expert supervision. To tackle these issues, this paper proposes Dual-granularity contrastive reward via generated Episodic Guidance (DEG), a novel framework to seek sample-efficient dense rewards without requiring human annotations or extensive supervision. Leveraging the prior knowledge of large video generation models, DEG only needs a small number of expert videos for domain adaptation to generate dedicated task guidance for each RL episode. Then, the proposed dual-granularity reward that balances coarse-grained exploration and fine-grained matching, will guide the agent to efficiently approximate the generated guidance video sequentially in the contrastive self-supervised latent space, and finally complete the target task. Extensive experiments on 18 diverse tasks across both simulation and real-world settings show that DEG can not only serve as an efficient exploration stimulus to help the agent quickly discover sparse success rewards, but also guide effective RL and stable policy convergence independently.

Boyin Liu, Zhuo Zhang, Sen Huang et al.

Aligning language models using LLM judge feedback offers a scalable alternative to human annotation, yet is plagued by judgment inconsistencies that destabilize reinforcement learning. While prior work has focused on judge accuracy, the critical issue of logical coherence particularly preference cycles has been largely unaddressed. To address this gap, this work introduces an end to end framework to systematically detect and resolve these inconsistencies within the reinforcement learning training loop. Our framework features two core contributions: the Conflict Detection Rate (CDR), a novel metric to quantify judgment conflicts, and Deconflicted Graph Rewards (DGR), a signal-purification framework that eliminates cycles before policy optimization. DGR constructs preference graphs from raw judgments, transforms them into conflict-free Directed Acyclic Graphs (DAGs), and generates a logically coherent reward signal compatible with any policy optimizer. Experiments confirm that our framework significantly improves training stability and model performance over strong baselines, establishing logical consistency as a crucial and now-addressable dimension of AI feedback. The code for our method is available at https://github.com/modelscope/RM-Gallery.

Nannan Li, Yaran Chen, Weifan Li et al.

Recent works have demonstrated that transformer can achieve promising performance in computer vision, by exploiting the relationship among image patches with self-attention. While they only consider the attention in a single feature layer, but ignore the complementarity of attention in different levels. In this paper, we propose the broad attention to improve the performance by incorporating the attention relationship of different layers for vision transformer, which is called BViT. The broad attention is implemented by broad connection and parameter-free attention. Broad connection of each transformer layer promotes the transmission and integration of information for BViT. Without introducing additional trainable parameters, parameter-free attention jointly focuses on the already available attention information in different layers for extracting useful information and building their relationship. Experiments on image classification tasks demonstrate that BViT delivers state-of-the-art accuracy of 74.8\%/81.6\% top-1 accuracy on ImageNet with 5M/22M parameters. Moreover, we transfer BViT to downstream object recognition benchmarks to achieve 98.9\% and 89.9\% on CIFAR10 and CIFAR100 respectively that exceed ViT with fewer parameters. For the generalization test, the broad attention in Swin Transformer and T2T-ViT also bring an improvement of more than 1\%. To sum up, broad attention is promising to promote the performance of attention based models. Code and pre-trained models are available at https://github.com/DRL-CASIA/Broad_ViT.

Yuqi Liu, Qichao Zhang, Dongbin Zhao

In recent years, control under urban intersection scenarios becomes an emerging research topic. In such scenarios, the autonomous vehicle confronts complicated situations since it must deal with the interaction with social vehicles timely while obeying the traffic rules. Generally, the autonomous vehicle is supposed to avoid collisions while pursuing better efficiency. The existing work fails to provide a framework that emphasizes the integrity of the scenarios while being able to deploy and test reinforcement learning(RL) methods. Specifically, we propose a benchmark for training and testing RL-based autonomous driving agents in complex intersection scenarios, which is called RL-CIS. Then, a set of baselines are deployed consists of various algorithms. The test benchmark and baselines are to provide a fair and comprehensive training and testing platform for the study of RL for autonomous driving in the intersection scenario, advancing the progress of RL-based methods for intersection autonomous driving control. The code of our proposed framework can be found at https://github.com/liuyuqi123/ComplexUrbanScenarios.

Yuhui Chen, Shuai Tian, Shugao Liu et al.

Vision-Language-Action (VLA) models have shown substantial potential in real-world robotic manipulation. However, fine-tuning these models through supervised learning struggles to achieve robust performance due to limited, inconsistent demonstrations, especially in contact-rich environments. In this paper, we propose a reinforced fine-tuning approach for VLA models, named ConRFT, which consists of offline and online fine-tuning with a unified consistency-based training objective, to address these challenges. In the offline stage, our method integrates behavior cloning and Q-learning to effectively extract policy from a small set of demonstrations and stabilize value estimating. In the online stage, the VLA model is further fine-tuned via consistency policy, with human interventions to ensure safe exploration and high sample efficiency. We evaluate our approach on eight diverse real-world manipulation tasks. It achieves an average success rate of 96.3% within 45-90 minutes of online fine-tuning, outperforming prior supervised methods with a 144% improvement in success rate and 1.9x shorter episode length. This work highlights the potential of integrating reinforcement learning to enhance the performance of VLA models for real-world robotic applications. Videos and code are available at our project website https://cccedric.github.io/conrft/.

Yuqian Fu, Tinghong Chen, Jiajun Chai et al.

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.

Songjun Tu, Jiahao Lin, Xiangyu Tian et al.

Recent advancements in post-training methodologies for large language models (LLMs) have highlighted reinforcement learning (RL) as a critical component for enhancing reasoning. However, the substantial computational costs associated with RL-based approaches have led to growing interest in alternative paradigms, such as Direct Preference Optimization (DPO). In this study, we investigate the effectiveness of DPO in facilitating self-improvement for LLMs through iterative preference-based learning. We demonstrate that a single round of DPO with coarse filtering significantly enhances mathematical reasoning performance, particularly for strong base model. Furthermore, we design an iterative enhancement framework for both the generator and the reward model (RM), enabling their mutual improvement through online interaction across multiple rounds of DPO. Finally, with simple verifiable rewards, our model DPO-VP achieves RL-level performance with significantly lower computational overhead. These findings highlight DPO as a scalable and cost-effective alternative to RL, offering a practical solution for enhancing LLM reasoning in resource-constrained situations.

Yuqian Fu, Yuanheng Zhu, Jiajun Chai et al.

Ensembling large language models (LLMs) can effectively combine diverse strengths of different models, offering a promising approach to enhance performance across various tasks. However, existing methods typically rely on fixed weighting strategies that fail to adapt to the dynamic, context-dependent characteristics of LLM capabilities. In this work, we propose Reinforcement Learning-Assisted Ensemble for LLMs (RLAE), a novel framework that reformulates LLM ensemble through the lens of a Markov Decision Process (MDP). Our approach introduces a RL agent that dynamically adjusts ensemble weights by considering both input context and intermediate generation states, with the agent being trained using rewards that directly correspond to the quality of final outputs. We implement RLAE using both single-agent and multi-agent reinforcement learning algorithms ($\text{RLAE}_\text{PPO}$ and $\text{RLAE}_\text{MAPPO}$ ), demonstrating substantial improvements over conventional ensemble methods. Extensive evaluations on a diverse set of tasks show that RLAE outperforms existing approaches by up to $3.3\%$ accuracy points, offering a more effective framework for LLM ensembling. Furthermore, our method exhibits superior generalization capabilities across different tasks without the need for retraining, while simultaneously achieving lower time latency.

Songjun Tu, Jingbo Sun, Qichao Zhang et al.

Preference-based reinforcement learning (PbRL) provides a powerful paradigm to avoid meticulous reward engineering by learning rewards based on human preferences. However, real-time human feedback is hard to obtain in online tasks. Most work suppose there is a "scripted teacher" that utilizes privileged predefined reward to provide preference feedback. In this paper, we propose a RL Self-augmented Large Language Model Feedback (RL-SaLLM-F) technique that does not rely on privileged information for online PbRL. RL-SaLLM-F leverages the reflective and discriminative capabilities of LLM to generate self-augmented trajectories and provide preference labels for reward learning. First, we identify an failure issue in LLM-based preference discrimination, specifically "query ambiguity", in online PbRL. Then LLM is employed to provide preference labels and generate self-augmented imagined trajectories that better achieve the task goal, thereby enhancing the quality and efficiency of feedback. Additionally, a double-check mechanism is introduced to mitigate randomness in the preference labels, improving the reliability of LLM feedback. The experiment across multiple tasks in the MetaWorld benchmark demonstrates the specific contributions of each proposed module in RL-SaLLM-F, and shows that self-augmented LLM feedback can effectively replace the impractical "scripted teacher" feedback. In summary, RL-SaLLM-F introduces a new direction of feedback acquisition in online PbRL that does not rely on any online privileged information, offering an efficient and lightweight solution with LLM-driven feedback.

Guangzheng Hu, Yuanheng Zhu, Haoran Li et al.

Many real-world applications involve some agents that fall into two teams, with payoffs that are equal within the same team but of opposite sign across the opponent team. The so-called two-team zero-sum Markov games (2t0sMGs) can be resolved with reinforcement learning in recent years. However, existing methods are thus inefficient in light of insufficient consideration of intra-team credit assignment, data utilization and computational intractability. In this paper, we propose the individual-global-minimax (IGMM) principle to ensure the coherence between two-team minimax behaviors and the individual greedy behaviors through Q functions in 2t0sMGs. Based on it, we present a novel multi-agent reinforcement learning framework, Factorized Multi-Agent MiniMax Q-Learning (FM3Q), which can factorize the joint minimax Q function into individual ones and iteratively solve for the IGMM-satisfied minimax Q functions for 2t0sMGs. Moreover, an online learning algorithm with neural networks is proposed to implement FM3Q and obtain the deterministic and decentralized minimax policies for two-team players. A theoretical analysis is provided to prove the convergence of FM3Q. Empirically, we use three environments to evaluate the learning efficiency and final performance of FM3Q and show its superiority on 2t0sMGs.

Kaixuan Xu, Jiajun Chai, Sicheng Li et al.

Diplomacy is a complex multiplayer game that requires both cooperation and competition, posing significant challenges for AI systems. Traditional methods rely on equilibrium search to generate extensive game data for training, which demands substantial computational resources. Large Language Models (LLMs) offer a promising alternative, leveraging pre-trained knowledge to achieve strong performance with relatively small-scale fine-tuning. However, applying LLMs to Diplomacy remains challenging due to the exponential growth of possible action combinations and the intricate strategic interactions among players. To address this challenge, we propose DipLLM, a fine-tuned LLM-based agent that learns equilibrium policies for Diplomacy. DipLLM employs an autoregressive factorization framework to simplify the complex task of multi-unit action assignment into a sequence of unit-level decisions. By defining an equilibrium policy within this framework as the learning objective, we fine-tune the model using only 1.5% of the data required by the state-of-the-art Cicero model, surpassing its performance. Our results demonstrate the potential of fine-tuned LLMs for tackling complex strategic decision-making in multiplayer games.

Zhennan Jiang, Shangqing Zhou, Yutong Jiang et al.

Reinforcement learning (RL) promises to unlock capabilities beyond imitation learning for Vision-Language-Action (VLA) models, but its requirement for massive real-world interaction prevents direct deployment on physical robots. Recent work attempts to use learned world models as simulators for policy optimization, yet closed-loop imagined rollouts inevitably suffer from hallucination and long-horizon error accumulation. Such errors do not merely degrade visual fidelity; they corrupt the optimization signal, encouraging policies to exploit model inaccuracies rather than genuine task progress. We propose WoVR, a reliable world-model-based reinforcement learning framework for post-training VLA policies. Instead of assuming a faithful world model, WoVR explicitly regulates how RL interacts with imperfect imagined dynamics. It improves rollout stability through a controllable action-conditioned video world model, reshapes imagined interaction to reduce effective error depth via Keyframe-Initialized Rollouts, and maintains policy-simulator alignment through World Model-Policy co-evolution. Extensive experiments on LIBERO benchmarks and real-world robotic manipulation demonstrate that WoVR enables stable long-horizon imagined rollouts and effective policy optimization, improving average LIBERO success from 39.95% to 69.2% (+29.3 points) and real-robot success from 61.7% to 91.7% (+30.0 points). These results show that learned world models can serve as practical simulators for reinforcement learning when hallucination is explicitly controlled.

Yinfeng Gao, Qichao Zhang, Da-wei Ding et al.

It is still a challenging topic to make reactive driving behaviors in complex urban environments as road users' intentions are unknown. Model-based reinforcement learning (MBRL) offers great potential to learn a reactive policy by constructing a world model that can provide informative states and imagination training. However, a critical limitation in relevant research lies in the scene-level reconstruction representation learning, which may overlook key interactive vehicles and hardly model the interactive features among vehicles and their long-term intentions. Therefore, this paper presents a novel MBRL method with a predictive individual world model (PIWM) for autonomous driving. PIWM describes the driving environment from an individual-level perspective and captures vehicles' interactive relations and their intentions via trajectory prediction task. Meanwhile, a behavior policy is learned jointly with PIWM. It is trained in PIWM's imagination and effectively navigates in the urban driving scenes leveraging intention-aware latent states. The proposed method is trained and evaluated on simulation environments built upon real-world challenging interactive scenarios. Compared with popular model-free and state-of-the-art model-based reinforcement learning methods, experimental results show that the proposed method achieves the best performance in terms of safety and efficiency.