论文检索

Heyang Ma, Qirui Mi, Qipeng Yang, Zijun Fan, Bo Li, Haifeng Zhang

Economic decision‑making depends not only on structured signals—such as prices and taxes—but also on unstructured language, including peer dialogue and media narratives. While multi‑agent reinforcement learning (MARL) has shown promise in optimizing economic decisions, it struggles with the semantic ambiguity and contextual richness of language. We propose LAMP (Language‑Augmented Multi‑Agent Policy), the first framework to integrate language into economic decision‑making, narrowing the gap to real‑world settings. LAMP follows a Think–Speak–Decide pipeline: (1) Think interprets numerical observations to extract short‑term shocks and long‑term trends, caching high‑value reasoning trajectories. (2) Speak crafts and exchanges strategic messages based on the reasoning, updating beliefs by parsing peer communications. (3) Decide fuses numerical data, reasoning, and reflections into a MARL policy to optimize language‑augmented decision‑making. Experiments in economic simulation show that LAMP outperforms both MARL and LLM‑only baselines in cumulative return (+63.5%, +34.0%), robustness (+18.8%, +59.4%), and interpretability. These results demonstrate the potential of language‑augmented policies to deliver more effective and robust economic strategies.

Ziqi Liu, Ziyang Zhou, Yilin Li, Mingxuan Hu, Yushan Pan, Zhijie Xu, Yangbin Chen

Sarcasm detection is a crucial yet challenging Natural Language Processing task. Existing Large Language Model methods are often limited by single-perspective analysis, static reasoning pathways, and a susceptibility to hallucination when processing complex ironic rhetoric, which impacts their accuracy and reliability. To address these challenges, we propose SEVADE, a novel Self-Evolving multi-agent Analysis framework with Decoupled Evaluation for hallucination-resistant sarcasm detection. The core of our framework is a Dynamic Agentive Reasoning Engine (DARE), which utilizes a team of specialized agents grounded in linguistic theory to perform a multifaceted deconstruction of the text and generate a structured reasoning chain. Subsequently, a separate lightweight rationale adjudicator (RA) performs the final classification based solely on this reasoning chain. This decoupled architecture is designed to mitigate the risk of hallucination by separating complex reasoning from the final judgment. Extensive experiments on four benchmark datasets demonstrate that our framework achieves state-of-the-art performance, with average improvements of 7.01% in Accuracy and 6.55% in Macro-F1 score.

Zejiao Liu, Junqi Tu, Yitian Hong, Luolin Xiong, Yaochu Jin, Yang Tang, Fangfei Li

In cooperative Multi-Agent Reinforcement Learning (MARL), efficient exploration is crucial for optimizing the performance of joint policy. However, existing methods often update joint policies via independent agent exploration, without coordination among agents, which inherently constrains the expressive capacity and exploration of joint policies. To address this issue, we propose a conductor-based joint policy framework that directly enhances the expressive capacity of joint policies and coordinates exploration. In addition, we develop a Hierarchical Conductor-based Policy Optimization (HCPO) algorithm that instructs policy updates for the conductor and agents in a direction aligned with performance improvement. A rigorous theoretical guarantee further establishes the monotonicity of the joint policy optimization process. By deploying local conductors, HCPO retains centralized training benefits while eliminating inter-agent communication during execution. Finally, we evaluate HCPO on three challenging benchmarks: StarCraft II Multi-agent Challenge, Multi-agent MuJoCo, and Multi-agent Particle Environment. The results indicate that HCPO outperforms competitive MARL baselines regarding cooperative efficiency and stability.

Qiwei Liu, Ye Yuan, Lingyue Zhang, Kaitian Chen, Yunkai Lv, Sheng Gao, Huaicheng Yan

Deploying multi-agent reinforcement learning (MARL) in safety-critical systems faces significant challenges due to insufficient agent exploration and inadequate safety constraint guarantees. Current approaches are constrained by two fundamental limitations: inefficient exploration leading to suboptimal policies, and expected-cost-based constraint frameworks failing to ensure full-process safety. To address these challenges, this paper proposes a novel safety-aware maximum entropy MARL framework using Conditional Value-at-Risk (CVaR) as a joint safety metric, which quantifies constraint satisfaction under worst-case scenarios for multi-agent systems. Moreover, we develop the Worst-Case Multi-Agent Soft Actor-Critic (WCMASAC) algorithm, incorporating sequential update mechanisms and maximum entropy optimization for heterogeneous agents, enhanced with distributed safety critics. Theoretically, we establish the monotonic improvement property, guaranteed constraint satisfaction, and convergence to a generalized Nash equilibrium for WCMASAC. Extensive experiments on Safety-Gymnasium based benchmarks demonstrate that WCMASAC outperforms state-of-the-art baselines in both task reward acquisition and safety constraint violation reduction, while exhibiting superior exploration efficiency and risk-aware control capabilities.

Jianhang Liu, Dianzheng Zhang, Hongxin Pan, Guangqian Jiang, Runda Fan

Intelligent perception among multiple agents enables them to extend their individual observation capabilities by sharing sensory information, thereby improving the completeness and accuracy of environmental understanding. However, real-world communication is often subject to non-negligible delays, which can degrade the effectiveness of perception. To mitigate this, delay alignment is commonly employed to synchronize delayed observations to a common timestamp. Yet, both alignment errors and inherent discrepancies between multi-view observations can lead to inconsistencies in the estimated position and orientation of shared targets. These inconsistencies can accumulate during feature fusion, ultimately reducing the accuracy and reliability of the perception results.To address this challenge, we propose IPDA, a delay-aware multi-agent intelligent perception method that performs joint calibration in both temporal and spatial domains. In the temporal dimension, we design a historical alignment attention mechanism to model dynamic delay correction across sequences, ensuring temporal coherence. In the spatial dimension, we introduce a discrepancy-quantized co-sensing network that captures and compensates for multi-view spatial deviations caused by viewpoint diversity and alignment inaccuracy. IPDA is evaluated on two large-scale intelligent perception benchmarks, DAIR-V2X and OPV2V. Experimental results demonstrate that our method effectively mitigates delay-induced inconsistencies and consistently outperforms state-of-the-art baselines under various delay conditions.

Zhuhui Li, Chunbo Luo, Liming Huang, Luyu Qi, Geyong Min

Multi-agent multi-objective systems (MAMOS) have emerged as powerful frameworks for modelling complex decision-making problems across various real-world domains, such as robotic exploration, autonomous traffic management, and sensor network optimisation. MAMOS enhances scalability and robustness through decentralised control and more accurately captures inherent trade-offs between conflicting objectives. In MAMOS, each agent uses utility functions that map return vectors to scalar values. Existing MAMOS optimisation methods face significant challenges in handling heterogeneous objective and utility function settings, where training non-stationarity is intensified due to private utility functions and the associated policies. In this paper, we first theoretically prove that direct access to, or structured modeling of, global utility functions is necessary to achieve the Bayesian Nash Equilibrium under decentralised execution constraints. To access the global utility functions while preserving the decentralised execution, we propose an Agent-Attention Multi-Agent Multi-Objective Reinforcement Learning (AA-MAMORL) framework. Our approach implicitly learns a joint belief over other agents’ utility functions and their associated policies during centralised training, effectively mapping global states and utilities to each agent's policy. During execution, each agent independently selects actions based on local observations and its private utility function to approximate a BNE, without relying on inter-agent communication. We evaluate our framework through extensive experiments in a custom-designed MAMO Particle environment and the standard MOMALand benchmark. The results demonstrate that accessibility to global preferences and our proposed AA-MAMORL significantly improves performance and consistently outperforms state-of-the-art methods.

Yuanhao Li, Mingshan Liu, Hongbo Wang, Yiding Zhang, Yifei Ma, Wei Tan

Large Language Models (LLMs) have shown impressive capabilities in multi-step reasoning and problem-solving. Recent works introduce multi-agent reflection frameworks where multiple LLM agents critique and refine each other’s outputs using reinforcement learning (RL). However, these approaches often rely on single-shot responses and lack structural diversity in reasoning exploration. In this paper, we propose DRAFT-RL, a novel framework that integrates Chain-of-Draft (CoD) reasoning into multi-agent RL training. Instead of generating single responses, each agent produces multiple drafts per query, which are then evaluated by peer agents and a learned reward model to identify the most promising trajectory. These selected drafts are used to refine future reasoning strategies through actor-critic learning. DRAFT-RL enables explicit multi-path exploration, peer-guided reflection, and reward-aligned selection, resulting in more robust and interpretable LLM agent behavior. We evaluate our method on complex reasoning tasks including code synthesis, symbolic math, and knowledge-intensive QA, demonstrating that DRAFT-RL outperforms existing reflective and RL-based agents by significant margins in both accuracy and convergence speed.

Chin-wing Leung, Paolo Turrini, Fernando P. Santos, Mirco Musolesi

Cooperation among independent learning agents is desirable as it enables reaching collectively rewarding states. Recent work has shown that artificial agents can learn to act pro-socially without the need for predefined cooperative preferences or behavioural heuristics, provided that they can observe others' actions or policies and select them as partners accordingly. This paper relaxes this constraint, studying reinforcement learning (RL) agents operating with only minimal information about others' behaviour. We propose a novel `Observer Model', where agents gain insights from direct experience and limited, indirect observations. We show that direct experience alone cannot sustain cooperation, particularly in large societies. However, even minimal observations of third-party interactions, allowing as few as one observer per gameplay, lead to significant improvements, enabling the population to achieve and sustain robust cooperation across varying population sizes. Through numerical analysis, we show the co-evolution of strategy and interaction structure and disentangle how learning happens under various settings. Analysing the partner selection graph, we identify the reasons for cooperation to emerge, and we explore how different learning and exploration rates affect the outcome of social dilemmas played among RL agents.

Yehonatan Kidushim, Avraham Natan, Roni Stern, Meir Kalech

Consider a system of multiple physical agents tasked with collaboratively collecting a set of spatially distributed goals as quickly as possible while avoiding collisions with the environment and with each other. This type of problem, which involves Multi-Agent Path Finding (MAPF) and task allocation, is called Multi-Agent Combinatorial Path Finding (MCPF). Prior work on MCPF assumed each agent has a final goal it must reach, there are no orientation constraints on the agents' movements, and the agents will follow their planned actions as intended. These assumptions rarely hold in real physical robots, which limits the applicability of existing MCPF algorithms in practical applications. We propose the Robust CBSS framework, a robust planning approach that solves MCPF without the aforementioned simplifying assumptions, and provide two implementations: a baseline version (RCbssBase) and an efficient version (RCbssEff). RCbssEff generalizes the Conflict-Based Steiner Search (CBSS) algorithm, building on ideas from the p-Robust CBS algorithm and algorithms for solving the Equality Generalized Traveling Salesman Problem. We prove that RCbssEff is complete and can be configured to return optimal solutions. Experimental results on benchmark MCPF problems show that RCbssEff balances planning time, solution cost, and collision reduction compared to baselines.

Rishabh Jain, Keisuke Okumura, Michael Amir, Amanda Prorok

Finding near-optimal solutions for dense multi-agent pathfinding (MAPF) problems in real-time remains challenging even for state-of-the-art planners. To this end, we develop a hybrid framework that integrates a learned heuristic derived from MAGAT, a neural MAPF policy with a graph attention scheme, into a leading search-based algorithm, LaCAM. While prior work has explored learning-guided search in MAPF, such methods have historically underperformed. In contrast, our approach, termed LaGAT, outperforms both purely search-based and purely learning-based methods in dense scenarios. This is achieved through an enhanced MAGAT architecture, a pre-train–then–fine-tune strategy on maps of interest, and a deadlock detection scheme to account for imperfect neural guidance. Our results demonstrate that, when carefully designed, hybrid search offers a powerful solution for tightly coupled, challenging multi-agent coordination problems.

Yujiao Hu, Zuyu Chen, MengJie Lee, Jinchao Chen, Meng Shen, Hailun Zhang, Wei Li, Yan Pan

Subset selection under budget constraints is critical in applications like multi-robot patrolling, crime deterrence, and targeted marketing, where multiple agents must jointly select targets and plan feasible routes. We formalize this challenge as Multi-Subset Selection with Budget-Constrained Routing (MSS-BCR), involving complex, non-additive cost structures that defy traditional methods. We propose GRIP, a graph-based framework integrating spatial reward fields and policy learning to enable coordinated, budget-aware target selection and routing. GRIP uses attention-based embeddings and constraint-triggered pruning with utility recovery to produce high-quality, feasible solutions. Experiments based on multiple synthetic and real-world datasets show GRIP outperforms baselines in reward efficiency and scalability across varied scenarios.

Tianyi Hu, Qingxu Fu, Zhiqiang Pu, Yuan Wang, Tenghai Qiu

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 including a sim2real demo. 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.

Jinwei Hu, Yi Dong, Youcheng Sun, Xiaowei Huang

Autonomous agents in safety-critical applications must continuously adapt to dynamic conditions without compromising performance and reliability. This work introduces TAPA (Training-free Adaptation of Programmatic Agents), a novel framework that positions large language models (LLMs) as intelligent moderators of the symbolic action space. Unlike prior programmatic agents typically generate a monolithic policy program or rely on fixed symbolic action sets, TAPA synthesizes and adapts modular programs for individual high-level actions, referred to as logical primitives. By decoupling strategic intent from execution, TAPA enables meta-agents to operate over an abstract, interpretable action space while the LLM dynamically generates, composes, and refines symbolic programs tailored to each primitive. Extensive experiments across cybersecurity and swarm intelligence domains validate TAPA's effectiveness. In autonomous DDoS defense scenarios, TAPA achieves 77.7% network uptime while maintaining near-perfect detection accuracy in unknown dynamic environments. In swarm intelligence formation control under environmental and adversarial disturbances, TAPA consistently preserves consensus at runtime where baseline methods fail. This work promotes a paradigm shift for autonomous system design in evolving environments, from policy adaptation to dynamic action adaptation.

Die Hu, Jingguo Ge, Weitao Tang, He Kong, Liangxiong Li, Bingzhen Wu

Recently, Large Language Models (LLMs) based Web Agents have shown significant potential in web understanding and interaction tasks. However, their personalization ability and user experience remain limited by the ambiguity and dynamic nature of user intent, struggling to model diverse user interests and track intent changes over time. To address these challenges, this paper proposes Orion, a novel personalized Web Agent. Orion adopts a global-micro profiling mechanism to balance users' long-term stable preferences and scenario-based needs, and introduces context-aware interest retrieval to enhance personalization. Additionally, we design adaptive profile tracking and proactive disambiguation mechanisms to effectively address the continuous evolution of user intent in multi-turn interactions. Orion is optimized through end-to-end online reinforcement learning, improving personalized reasoning and decision-making ability in real interactive scenarios. Experiments demonstrate that Orion significantly outperforms state-of-the-art baselines in personalized understanding and task efficiency.

Linqiang Guo, Wei Liu, Yi Wen Heng, Tse-Hsun (Peter) Chen, Yang Wang

Mobile Graphical User Interface (GUI) agents aim to autonomously complete tasks within or across apps based on user instructions. While recent Multimodal Large Language Models (MLLMs) enable these agents to interpret UI screens and perform actions, existing agents remain fundamentally reactive. They reason over the current UI screen but lack a structured representation of the app navigation flow, lim- iting GUI agents’ ability to understand execution context, detect unexpected execution results, and recover from errors. We introduce Agent-SAMA, a state-aware multi-agent framework that models app execution as a Finite State Machine (FSM), treating UI screens as states and user actions as transitions. Agent-SAMA implements four specialized agents that collaboratively construct and use FSMs in real time to guide task planning, execution verification, and recovery. We evaluate Agent-SAMA on two types of benchmarks: cross- app (Mobile-Eval-E, SPA-Bench) and mostly single-app (AndroidWorld). On Mobile-Eval-E, Agent-SAMA achieves an 84.0% success rate and a 71.9% recovery rate. On SPA-Bench, it reaches an 80.0% success rate with a 66.7% recovery rate. Compared to prior methods, Agent-SAMA improves task success by up to 12% and recovery success by 13.8%. On AndroidWorld, Agent-SAMA achieves a 63.7% success rate, outperforming the baselines. Our results demonstrate that structured state modeling enhances robustness and can serve as a lightweight, model-agnostic memory layer for future GUI agents.

Yuxin Geng, Wolfram Barfuss, Xingru Chen

Understanding the emergence of collective behaviors of multi-agent systems requires investigating the learning dynamics. However, the theoretical analysis of large-scale graph-structured multi-agent reinforcement learning (MARL) systems remains challenging due to agent heterogeneity and the intrinsic coupling between state transitions and individual Q-value updates. In this work, we develop a unified theoretical framework that captures the evolution of agent behaviors at both individual and population levels. By leveraging the pair approximation technique from statistical physics, we derive a closed set of evolution equations that accurately describe the temporal dynamics of the system. Our analysis also reveals a separation of time scales. For small learning rates, state transitions equilibrate rapidly, while Q-value updates evolve slowly with stationary state distributions. Through extensive agent-based simulations, we validate the robustness of our theoretical results and explain the mechanisms that lead to the emergence of cooperation in social dilemmas. Our framework offers new perspectives for bridging complex systems science and MARL, providing insights for the design of cooperative and resilient AI.

Zihong Gao, Hongjian Liang, Yuanhui Hao, Lei Hao, Liangjun Ke

Effective coordination in Multi-Agent Reinforcement Learning (MARL) is particularly challenging under partial observability, where agents must reason about potential collaborators using only local information. Existing methods fall into two categories: communication-based approaches that enable message exchange but often fix or misidentify who the collaborators are, and role-based approaches that encourage specialization based on behavioral similarity. However, both lines of work overlook the task‑induced cooperative dependencies that decide which agents should collaborate, leading to miscommunication or role misassignment under partial observability. We introduce GRDC (Graph‑driven Role Discovery and Communication), a unified framework that approximates these dependencies by dynamically constructing local interaction graphs from trajectory embeddings, then uses these graphs to infer roles via prototype matching and to restrict communication to intra‑role agents with attention-based aggregation. Beyond role inference and communication, GRDC maximizes role entropy, decorrelates prototypes, and dynamically prunes redundant ones to obtain structured yet compact role specialization. Experimental results on Predator Prey, Cooperative Navigation, and SMACv2 demonstrate that GRDC consistently outperforms state-of-the-art communication- and role-based baselines, improving coordination efficiency and training stability across tasks.

Jiarun Fu, Lizhong Ding, Qiuning Wei, Yuhan Guo, Yurong Cheng, Junyu Zhang

Large language models have revolutionized agent planning by serving as the engine of heuristic guidance. However, LLM-based agents often struggle to generalize across complex environments and to adapt to stochastic feedback arising from environment–action interactions. We propose Counterfactual Planning—a method designed to improve the generalizability and adaptability of agents' actions by inferring causal representations of environmental confounders and performing counterfactual reasoning over planned actions. We formalize the agent planning process as a structural causal model, providing a mathematical formulation for causal analysis of how environmental states influence action generation and how actions affect future state transitions. To support generalizable action planning, we introduce the State Causality Evaluator (SCE), which dynamically infers task-conditioned causal representations from complex environment states; and to enhance adaptability under stochastic feedback, we propose the What-If-Not (WIN) reward, which performs counterfactual interventions to refine actions through causal evaluation. We validate our framework in an open-world environment, where experiments demonstrate improvements in both action generalization and planning adaptability.

Bo Fu, Zhe Chen, Rahul Chandan, Alexandre Ormiga Galvao Barbosa, Michael Caldara, Joey W. Durham, Federico Pecora

We introduce the Block Rearrangement Problem (BRaP), a challenging component of large warehouse management which involves rearranging storage blocks within dense grids to achieve a goal state. We formally define the BRaP as a graph search problem. Building on intuitions from sliding puzzle problems, we propose five search-based solution algorithms, leveraging joint configuration space search, classical planning, multi-agent pathfinding, and expert heuristics. We evaluate the five approaches empirically for plan quality and scalability. Despite the exponential relation between search space size and block number, our methods demonstrate efficiency in creating rearrangement plans for deeply buried blocks in up to 80x80 grids.

Yijia Fan, Jusheng Zhang, Kaitong Cai, Jing Yang, Chengpei Tang, Jian Wang, Keze Wang

Multi-agent systems (MAS) built on large language models (LLMs) often suffer from inefficient ''free-for-all'' communication, leading to exponential token costs and low signal-to-noise ratios that hinder their practical deployment. We challenge the notion that more communication is always beneficial, hypothesizing instead that the core issue is the absence of resource rationality. We argue that "free'' communication, by ignoring the principle of scarcity, inherently breeds inefficiency and unnecessary expenses. To address this, we introduce the Dynamic Auction-based Language Agent (DALA), a novel framework that treats communication bandwidth as a scarce and tradable resource. Specifically, our DALA regards inter-agent communication as a centralized auction, where agents learn to bid for the opportunity to speak based on the predicted value density of their messages. Thus, our DALA intrinsically encourages agents to produce concise, informative messages while filtering out low-value communication. Extensive and comprehensive experiments demonstrate that our economically-driven DALA achieves new state-of-the-art performance across seven challenging reasoning benchmarks, including 84.32% on MMLU and a 91.21% pass@1 rate on HumanEval. Note that this is accomplished with remarkable efficiency, i.e., our DALA uses only 6.25 million tokens, a fraction of the resources consumed by current state-of-the-art methods on GSM8K. Further analysis reveals that our DALA cultivates the emergent skill of strategic silence, effectively adapting its communication strategies from verbosity to silence in a dynamic manner via resource constraints.