arXiv subjects

Zhiqiang Pu

Publications and source records attributed to Zhiqiang Pu.

At least 19 recordsLinked to original sources

arXiv2026

CoSkill: Joint Reinforcement Learning of Reasoning and Meta-Skill Agents for Hierarchical Skill Evolution

Skill libraries improve the sample efficiency of agentic reinforcement learning (RL) by enabling large language model (LLM) agents to reuse procedural knowledge. Yet existing paradigms exhibit structural shortcomings: they either decouple skill evolution from policy optimization or instantiate meta-skills as fixed workflows. Both treat skills as passive objects to be managed, limiting the flexible evolution of skills and their co-adaptation with the reasoning agent. To address the limitations, we propose CoSkill, a unified multi-agent RL framework that recasts the static meta-skill workflow as a learnable Meta-Skill Agent and jointly trains it with a Reasoning Agent over a hierarchical skill library. By modeling the Reasoning and Meta-Skill Agents as a cooperative team sharing a single backbone, CoSkill enables end-to-end co-adaptation: the Reasoning Agent conditions its actions on a retrieved task skill and step skills selected from its child set, while its task performance guides the Meta-Skill Agent in refining those step skills. Experiments on ALFWorld and WebShop show that CoSkill substantially outperforms prior skill-based and RL baselines, achieving success rates of 98.4% and 90.6%, respectively (+3.5 and +6.2 pp). As shown in Figure 1, CoSkill achieves superior early-stage sample efficiency, asymptotic performance, and wall-clock efficiency. Our code is available at https://github.com/jinyuan-cookie/CoSkill.

cs.AI↗

arXiv2026

Focus on the Core: Empowering Diffusion Large Language Models by Self-Contrast

The iterative denoising paradigm of Diffusion Large Language Models (DLMs) endows them with a distinct advantage in global context modeling. However, current decoding strategies fail to leverage this capability, typically exhibiting a local preference that overlooks the heterogeneous information density within the context, ultimately degrading generation quality. To address this limitation, we systematically investigate high-information-density (HD) tokens and present two key findings: (1) explicitly conditioning on HD tokens substantially improves output quality; and (2) HD tokens exhibit an early-decoding tendency, converging earlier than surrounding tokens. Motivated by these findings, we propose Focus on the Core \textbf{(FoCore)}, a training-free decoding strategy that utilizes HD tokens in a self-contrast manner, wherein HD tokens are temporarily remasked as negative samples, to guide generation. We further introduce FoCore\_Accelerate \textbf{(FoCore\_A)}, an efficient variant that, upon detecting HD token convergence, performs parallel decoding over stable candidates within a local context window, substantially accelerating generation. Extensive experiments on math, code and logical reasoning benchmarks demonstrate that FoCore consistently improves generation quality and efficiency across both LLaDA and Dream backbones. For instance, on HumanEval, FoCore improves pass@1 from 39.02 to 42.68 over standard Classifier-Free Guidance, while FoCore-A reduces the number of decoding steps by 2.07x and per-sample latency from 20.76s to 8.64s (-58.4\%).

cs.CL↗

arXiv2026

Efficient Soft Actor-Critic with LLM-Based Action-Level Guidance for Continuous Control

We present GuidedSAC, a novel reinforcement learning (RL) algorithm that facilitates efficient exploration in vast state-action spaces. GuidedSAC leverages large language models (LLMs) as intelligent supervisors that provide action-level guidance for the Soft Actor-Critic (SAC) algorithm. The LLM-based supervisor analyzes the most recent trajectory using state information and visual replays, offering action-level interventions that enable targeted exploration. Furthermore, we provide a theoretical analysis of GuidedSAC, proving that it preserves the convergence guarantees of SAC while improving convergence speed. Through experiments in both discrete and continuous control environments, including toy text tasks and complex MuJoCo benchmarks, we demonstrate that GuidedSAC consistently outperforms standard SAC and state-of-the-art exploration-enhanced variants (e.g., RND, ICM, and E3B) in terms of sample efficiency and final performance.

cs.LG↗

arXiv2026

Enhancing Reinforcement Learning Fine-Tuning with an Online Refiner

Constraints are essential for stabilizing reinforcement learning fine-tuning (RFT) and preventing degenerate outputs, yet they inherently conflict with the optimization objective because stronger constraints limit the ability of a fine-tuned model to discover better solutions. We propose \textit{dynamic constraints} that resolve this tension by adapting to the evolving capabilities of the fine-tuned model based on the insight that constraints should only intervene when degenerate outputs occur. We implement this by using a reference model as an \textit{online refiner} that takes the response from the fine-tuned model and generates a minimally corrected version which preserves correct content verbatim while fixing errors. A supervised fine-tuning loss then trains the fine-tuned model to produce the refined output. This mechanism yields a constraint that automatically strengthens or relaxes based on output quality. Experiments on dialogue and code generation show that dynamic constraints outperform both KL regularization and unconstrained baselines, achieving substantially higher task rewards while maintaining training stability.

cs.LG↗

arXiv2026

Self-Compression of Chain-of-Thought via Multi-Agent Reinforcement Learning

The inference overhead induced by redundant reasoning undermines the interactive experience and severely bottlenecks the deployment of Large Reasoning Models. Existing reinforcement learning (RL)-based solutions tackle this problem by coupling a length penalty with outcome-based rewards. This simplistic reward weighting struggles to reconcile brevity with accuracy, as enforcing brevity may compromise critical reasoning logic. In this work, we address this limitation by proposing a multi-agent RL framework that selectively penalizes redundant chunks, while preserving essential reasoning logic. Our framework, Self-Compression via MARL (SCMA), instantiates redundancy detection and evaluation through two specialized agents: \textbf{a Segmentation Agent} for decomposing the reasoning process into logical chunks, and \textbf{a Scoring Agent} for quantifying the significance of each chunk. The Segmentation and Scoring agents collaboratively define an importance-weighted length penalty during training, incentivizing \textbf{a Reasoning Agent} to prioritize essential logic without introducing inference overhead during deployment. Empirical evaluations across model scales demonstrate that SCMA reduces response length by 11.1\% to 39.0\% while boosting accuracy by 4.33\% to 10.02\%. Furthermore, ablation studies and qualitative analysis validate that the synergistic optimization within the MARL framework fosters emergent behaviors, yielding more powerful LRMs compared to vanilla RL paradigms.

cs.AI↗

arXiv2025

Heterogeneity in Multi-Agent Reinforcement Learning

Heterogeneity is a fundamental property in multi-agent reinforcement learning (MARL), which is closely related not only to the functional differences of agents, but also to policy diversity and environmental interactions. However, the MARL field currently lacks a rigorous definition and deeper understanding of heterogeneity. This paper systematically discusses heterogeneity in MARL from the perspectives of definition, quantification, and utilization. First, based on an agent-level modeling of MARL, we categorize heterogeneity into five types and provide mathematical definitions. Second, we define the concept of heterogeneity distance and propose a practical quantification method. Third, we design a heterogeneity-based multi-agent dynamic parameter sharing algorithm as an example of the application of our methodology. Case studies demonstrate that our method can effectively identify and quantify various types of agent heterogeneity. Experimental results show that the proposed algorithm, compared to other parameter sharing baselines, has better interpretability and stronger adaptability. The proposed methodology will help the MARL community gain a more comprehensive and profound understanding of heterogeneity, and further promote the development of practical algorithms.

cs.MA↗

arXiv2025

TacEleven: generative tactic discovery for football open play

Creating offensive advantages during open play is fundamental to football success. However, due to the highly dynamic and long-sequence nature of open play, the potential tactic space grows exponentially as the sequence progresses, making automated tactic discovery extremely challenging. To address this, we propose TacEleven, a generative framework for football open-play tactic discovery developed in close collaboration with domain experts from AJ Auxerre, designed to assist coaches and analysts in tactical decision-making. TacEleven consists of two core components: a language-controlled tactical generator that produces diverse tactical proposals, and a multimodal large language model-based tactical critic that selects the optimal proposal aligned with a high-level stylistic tactical instruction. The two components enables rapid exploration of tactical proposals and discovery of alternative open-play offensive tactics. We evaluate TacEleven across three tasks with progressive tactical complexity: counterfactual exploration, single-step discovery, and multi-step discovery, through both quantitative metrics and a questionnaire-based qualitative assessment. The results show that the TacEleven-discovered tactics exhibit strong realism and tactical creativity, with 52.50% of the multi-step tactical alternatives rated adoptable in real-world elite football scenarios, highlighting the framework's ability to rapidly generate numerous high-quality tactics for complex long-sequence open-play situations. TacEleven demonstrates the potential of creatively leveraging domain data and generative models to advance tactical analysis in sports.

stat.AP↗

arXiv2025

CoMoE: Contrastive Representation for Mixture-of-Experts in Parameter-Efficient Fine-tuning

In parameter-efficient fine-tuning, mixture-of-experts (MoE), which involves specializing functionalities into different experts and sparsely activating them appropriately, has been widely adopted as a promising approach to trade-off between model capacity and computation overhead. However, current MoE variants fall short on heterogeneous datasets, ignoring the fact that experts may learn similar knowledge, resulting in the underutilization of MoE's capacity. In this paper, we propose Contrastive Representation for MoE (CoMoE), a novel method to promote modularization and specialization in MoE, where the experts are trained along with a contrastive objective by sampling from activated and inactivated experts in top-k routing. We demonstrate that such a contrastive objective recovers the mutual-information gap between inputs and the two types of experts. Experiments on several benchmarks and in multi-task settings demonstrate that CoMoE can consistently enhance MoE's capacity and promote modularization among the experts.

cs.LG↗

arXiv2025

OMoE: Diversifying Mixture of Low-Rank Adaptation by Orthogonal Finetuning

Building mixture-of-experts (MoE) architecture for Low-rank adaptation (LoRA) is emerging as a potential direction in parameter-efficient fine-tuning (PEFT) for its modular design and remarkable performance. However, simply stacking the number of experts cannot guarantee significant improvement. In this work, we first conduct qualitative analysis to indicate that experts collapse to similar representations in vanilla MoE, limiting the capacity of modular design and computational efficiency. Ulteriorly, Our analysis reveals that the performance of previous MoE variants maybe limited by a lack of diversity among experts. Motivated by these findings, we propose Orthogonal Mixture-of-Experts (OMoE), a resource-efficient MoE variant that trains experts in an orthogonal manner to promote diversity. In OMoE, a Gram-Schmidt process is leveraged to enforce that the experts' representations lie within the Stiefel manifold. By applying orthogonal constraints directly to the architecture, OMoE keeps the learning objective unchanged, without compromising optimality. Our method is simple and alleviates memory bottlenecks, as it incurs minimal experts compared to vanilla MoE models. Experiments on diverse commonsense reasoning benchmarks demonstrate that OMoE can consistently achieve stable and efficient performance improvement when compared with the state-of-the-art methods while significantly reducing the number of required experts.

cs.LG↗

arXiv2025

Unreal-MAP: Unreal-Engine-Based General Platform for Multi-Agent Reinforcement Learning

In this paper, we propose Unreal Multi-Agent Playground (Unreal-MAP), an MARL general platform based on the Unreal-Engine (UE). Unreal-MAP allows users to freely create multi-agent tasks using the vast visual and physical resources available in the UE community, and deploy state-of-the-art (SOTA) MARL algorithms within them. Unreal-MAP is user-friendly in terms of deployment, modification, and visualization, and all its components are open-source. We also develop an experimental framework compatible with algorithms ranging from rule-based to learning-based provided by third-party frameworks. Lastly, we deploy several SOTA algorithms in example tasks developed via Unreal-MAP, and conduct corresponding experimental analyses. We believe Unreal-MAP can play an important role in the MARL field by closely integrating existing algorithms with user-customized tasks, thus advancing the field of MARL.

cs.AI↗

arXiv2025

Stochastic Trajectory Prediction under Unstructured Constraints

Trajectory prediction facilitates effective planning and decision-making, while constrained trajectory prediction integrates regulation into prediction. Recent advances in constrained trajectory prediction focus on structured constraints by constructing optimization objectives. However, handling unstructured constraints is challenging due to the lack of differentiable formal definitions. To address this, we propose a novel method for constrained trajectory prediction using a conditional generative paradigm, named Controllable Trajectory Diffusion (CTD). The key idea is that any trajectory corresponds to a degree of conformity to a constraint. By quantifying this degree and treating it as a condition, a model can implicitly learn to predict trajectories under unstructured constraints. CTD employs a pre-trained scoring model to predict the degree of conformity (i.e., a score), and uses this score as a condition for a conditional diffusion model to generate trajectories. Experimental results demonstrate that CTD achieves high accuracy on the ETH/UCY and SDD benchmarks. Qualitative analysis confirms that CTD ensures adherence to unstructured constraints and can predict trajectories that satisfy combinatorial constraints.

cs.RO↗

arXiv2025

MA2RL: Masked Autoencoders for Generalizable Multi-Agent Reinforcement Learning

To develop generalizable models in multi-agent reinforcement learning, recent approaches have been devoted to discovering task-independent skills for each agent, which generalize across tasks and facilitate agents' cooperation. However, particularly in partially observed settings, such approaches struggle with sample efficiency and generalization capabilities due to two primary challenges: (a) How to incorporate global states into coordinating the skills of different agents? (b) How to learn generalizable and consistent skill semantics when each agent only receives partial observations? To address these challenges, we propose a framework called \textbf{M}asked \textbf{A}utoencoders for \textbf{M}ulti-\textbf{A}gent \textbf{R}einforcement \textbf{L}earning (MA2RL), which encourages agents to infer unobserved entities by reconstructing entity-states from the entity perspective. The entity perspective helps MA2RL generalize to diverse tasks with varying agent numbers and action spaces. Specifically, we treat local entity-observations as masked contexts of the global entity-states, and MA2RL can infer the latent representation of dynamically masked entities, facilitating the assignment of task-independent skills and the learning of skill semantics. Extensive experiments demonstrate that MA2RL achieves significant improvements relative to state-of-the-art approaches, demonstrating extraordinary performance, remarkable zero-shot generalization capabilities and advantageous transferability.

cs.MA↗

arXiv2025

Coevolving with the Other You: Fine-Tuning LLM with Sequential Cooperative Multi-Agent Reinforcement Learning

Reinforcement learning (RL) has emerged as a pivotal technique for fine-tuning large language models (LLMs) on specific tasks. However, prevailing RL fine-tuning methods predominantly rely on PPO and its variants. Though these algorithms are effective in general RL settings, they often exhibit suboptimal performance and vulnerability to distribution collapse when applied to the fine-tuning of LLMs. In this paper, we propose CORY, extending the RL fine-tuning of LLMs to a sequential cooperative multi-agent reinforcement learning framework, to leverage the inherent coevolution and emergent capabilities of multi-agent systems. In CORY, the LLM to be fine-tuned is initially duplicated into two autonomous agents: a pioneer and an observer. The pioneer generates responses based on queries, while the observer generates responses using both the queries and the pioneer's responses. The two agents are trained together. During training, the agents exchange roles periodically, fostering cooperation and coevolution between them. Experiments evaluate CORY's performance by fine-tuning GPT-2 and Llama-2 under subjective and objective reward functions on the IMDB Review and GSM8K datasets, respectively. Results show that CORY outperforms PPO in terms of policy optimality, resistance to distribution collapse, and training robustness, thereby underscoring its potential as a superior methodology for refining LLMs in real-world applications.

cs.AI↗

arXiv2025

Causal Mean Field Multi-Agent Reinforcement Learning

Scalability remains a challenge in multi-agent reinforcement learning and is currently under active research. A framework named mean-field reinforcement learning (MFRL) could alleviate the scalability problem by employing the Mean Field Theory to turn a many-agent problem into a two-agent problem. However, this framework lacks the ability to identify essential interactions under nonstationary environments. Causality contains relatively invariant mechanisms behind interactions, though environments are nonstationary. Therefore, we propose an algorithm called causal mean-field Q-learning (CMFQ) to address the scalability problem. CMFQ is ever more robust toward the change of the number of agents though inheriting the compressed representation of MFRL's action-state space. Firstly, we model the causality behind the decision-making process of MFRL into a structural causal model (SCM). Then the essential degree of each interaction is quantified via intervening on the SCM. Furthermore, we design the causality-aware compact representation for behavioral information of agents as the weighted sum of all behavioral information according to their causal effects. We test CMFQ in a mixed cooperative-competitive game and a cooperative game. The result shows that our method has excellent scalability performance in both training in environments containing a large number of agents and testing in environments containing much more agents.

cs.AI↗

arXiv2025

Vision-Based Generic Potential Function for Policy Alignment in Multi-Agent Reinforcement Learning

Guiding the policy of multi-agent reinforcement learning to align with human common sense is a difficult problem, largely due to the complexity of modeling common sense as a reward, especially in complex and long-horizon multi-agent tasks. Recent works have shown the effectiveness of reward shaping, such as potential-based rewards, to enhance policy alignment. The existing works, however, primarily rely on experts to design rule-based rewards, which are often labor-intensive and lack a high-level semantic understanding of common sense. To solve this problem, we propose a hierarchical vision-based reward shaping method. At the bottom layer, a visual-language model (VLM) serves as a generic potential function, guiding the policy to align with human common sense through its intrinsic semantic understanding. To help the policy adapts to uncertainty and changes in long-horizon tasks, the top layer features an adaptive skill selection module based on a visual large language model (vLLM). The module uses instructions, video replays, and training records to dynamically select suitable potential function from a pre-designed pool. Besides, our method is theoretically proven to preserve the optimal policy. Extensive experiments conducted in the Google Research Football environment demonstrate that our method not only achieves a higher win rate but also effectively aligns the policy with human common sense.

cs.AI↗

arXiv2024

Efficient Multi-Task Reinforcement Learning via Task-Specific Action Correction

Multi-task reinforcement learning (MTRL) demonstrate potential for enhancing the generalization of a robot, enabling it to perform multiple tasks concurrently. However, the performance of MTRL may still be susceptible to conflicts between tasks and negative interference. To facilitate efficient MTRL, we propose Task-Specific Action Correction (TSAC), a general and complementary approach designed for simultaneous learning of multiple tasks. TSAC decomposes policy learning into two separate policies: a shared policy (SP) and an action correction policy (ACP). To alleviate conflicts resulting from excessive focus on specific tasks' details in SP, ACP incorporates goal-oriented sparse rewards, enabling an agent to adopt a long-term perspective and achieve generalization across tasks. Additional rewards transform the original problem into a multi-objective MTRL problem. Furthermore, to convert the multi-objective MTRL into a single-objective formulation, TSAC assigns a virtual expected budget to the sparse rewards and employs Lagrangian method to transform a constrained single-objective optimization into an unconstrained one. Experimental evaluations conducted on Meta-World's MT10 and MT50 benchmarks demonstrate that TSAC outperforms existing state-of-the-art methods, achieving significant improvements in both sample efficiency and effective action execution.

cs.LG↗

arXiv2024

Self-Clustering Hierarchical Multi-Agent Reinforcement Learning with Extensible Cooperation Graph

Multi-Agent Reinforcement Learning (MARL) has been successful in solving many cooperative challenges. However, classic non-hierarchical MARL algorithms still cannot address various complex multi-agent problems that require hierarchical cooperative behaviors. The cooperative knowledge and policies learned in non-hierarchical algorithms are implicit and not interpretable, thereby restricting the integration of existing knowledge. This paper proposes a novel hierarchical MARL model called Hierarchical Cooperation Graph Learning (HCGL) for solving general multi-agent problems. HCGL has three components: a dynamic Extensible Cooperation Graph (ECG) for achieving self-clustering cooperation; a group of graph operators for adjusting the topology of ECG; and an MARL optimizer for training these graph operators. HCGL's key distinction from other MARL models is that the behaviors of agents are guided by the topology of ECG instead of policy neural networks. ECG is a three-layer graph consisting of an agent node layer, a cluster node layer, and a target node layer. To manipulate the ECG topology in response to changing environmental conditions, four graph operators are trained to adjust the edge connections of ECG dynamically. The hierarchical feature of ECG provides a unique approach to merge primitive actions (actions executed by the agents) and cooperative actions (actions executed by the clusters) into a unified action space, allowing us to integrate fundamental cooperative knowledge into an extensible interface. In our experiments, the HCGL model has shown outstanding performance in multi-agent benchmarks with sparse rewards. We also verify that HCGL can easily be transferred to large-scale scenarios with high zero-shot transfer success rates.

cs.AI↗

arXiv2024

Prioritized League Reinforcement Learning for Large-Scale Heterogeneous Multiagent Systems

Large-scale heterogeneous multiagent systems feature various realistic factors in the real world, such as agents with diverse abilities and overall system cost. In comparison to homogeneous systems, heterogeneous systems offer significant practical advantages. Nonetheless, they also present challenges for multiagent reinforcement learning, including addressing the non-stationary problem and managing an imbalanced number of agents with different types. We propose a Prioritized Heterogeneous League Reinforcement Learning (PHLRL) method to address large-scale heterogeneous cooperation problems. PHLRL maintains a record of various policies that agents have explored during their training and establishes a heterogeneous league consisting of diverse policies to aid in future policy optimization. Furthermore, we design a prioritized policy gradient approach to compensate for the gap caused by differences in the number of different types of agents. Next, we use Unreal Engine to design a large-scale heterogeneous cooperation benchmark named Large-Scale Multiagent Operation (LSMO), which is a complex two-team competition scenario that requires collaboration from both ground and airborne agents. We use experiments to show that PHLRL outperforms state-of-the-art methods, including QTRAN and QPLEX in LSMO.

cs.AI↗