Training LLM-based multi-agent systems with multi-agent reinforcement learning with multi-agent reinforcement learning is rapidly gaining traction, and a parallel line of work argues that such systems should be judged by their behavior, not only their reward.
Abstract
Training LLM-based multi-agent systems with multi-agent reinforcement learning is rapidly gaining traction, and a parallel line of work argues that such systems should be judged by their behavior, not only their reward. We take up both threads in spatial cooperative games, where small frozen LLMs prompted with low-level actions fail outright, earning zero reward. Guided by the options/semi-MDP framework---and, because option execution is asynchronous across agents, its multi-agent extension in macro-action Dec-POMDPs---we equip each game with a library of symbolic \emph{options}: typed, state-feasible, short-horizon behaviors executed by a symbolic planner. Each library is drafted by a frontier coding model from the game's source code; the feasibility guards that filter each menu are then synthesized mechanically from cheap random-policy burn-in rollouts---a guard is adopted only if it explains repeated execution failures while hiding no logged success---so no guard is authored, selected, or reward-tuned by hand. Each agent's LLM acts as its policy over options, with a private per-agent LoRA adapter trained by a per-agent variant of multi-agent GRPO (PA-MAGRPO); this lifts frozen bases from zero reward to competent play across three games and four small backbones. Behavioral audits then reveal that reward and cooperation decouple: a rising reward curve may simply mean that one agent has learned to run the entire task alone while its partner idles---cooperation emerges only when the task makes it necessary. Reward alone is thus an unreliable readout of cooperation; behavioral evaluation must sit alongside it.
Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, multi-turn agentic tasks. Recent work introduces privileged self-distillation for credit assignment, providing denser supervision, but it remains unclear how such local signals should represent sequential credit. We propose AgentOPSD, a critic-free, recursive method for turn-level credit assignment in agentic reinforcement learning. AgentOPSD aggregates token-level teacher-student log-probability gaps into turn-level evidence and recursively updates a Bayesian belief state in log-odds space. This yields a principled reweighting scheme that converts sparse outcome supervision into turn-level credit signals and identifies pivotal turns through the marginal belief revision between consecutive states. The method is fully compatible with standard policy optimization and requires neither an additional critic nor extra rollouts. We evaluate AgentOPSD on ALFWorld, WebShop, and Search-QA using Qwen2.5 models at two scales (3B and 7B). AgentOPSD outperforms GRPO and strong self-distillation baselines, achieving 89.1% success on ALFWorld with Qwen2.5-7B. Ablation studies attribute the gains to turn-level aggregation and history-dependent recursive belief updates.
Zi-Han Wang, Zhengxi Lu, Zhiyuan Yao et al.· 0 citations
In the quantitative finance area, particularly in order execution, reinforcement learning (RL) has shown great promise due to its ability to interact with market environments based on real data. However, traditional RL methods suffer from slow research speed and rely on static market assumptions, which do not consider the impact of the agent's execution action on the environment. To address these, we propose a Self-Evolutional single-agent/multi-agent Reinforcement Learning (SE-RL) framework. The framework utilizes a Large Language Model (LLM) to design various RL algorithm modules, such as agent model design, reward function, profiling, communication, and state imagination, by leveraging the LLM generating module output or code. SE-RL could continuously improve the accuracy of LLM-generated RL algorithms through a dual-enhancement kit at both high-level (prompt refinement) and low-level (parameter fine-tuning). Additionally, we use a multi-agent system to simulate dynamic financial markets, accounting for the impact of order executions on market dynamics. To further enhance training in such a dynamic market, we develop a hybrid environment training method that could rebalance each environment's loss weight. Comprehensive experiments on 200 realistic stock datasets demonstrate that our proposed framework outperforms current state-of-the-art baselines. Project page: https://kdd2026-se-rl.github.io/.
Vincent Fu, Xinxin Xu, Weichen Xu et al.· Proceedings of the 32nd ACM...· 0 citations
BPO is instantiate as Branching Policy Optimization (BPO), a sandbox-native RL algorithm that adaptively snapshots the sandbox at high-entropy decision points along a backbone trajectory, and proves this estimator is unbiased and has strictly lower variance than the trajectory-level baseline, with the reduction equal to the prefix-explained portion of return variance.
Bowei He, Yankai Chen, Xiaokun Zhang et al.· 1 citation
For reinforcement learning (RL) in self-contained environments, a policy can get rewards by exploiting environment-specific regularities (tool schemas, grader parsing, task templates) rather than by acquiring transferable skill, and an in-distribution holdout shares those regularities. We argue that the discriminating question is behavioral, namely how a trained agent acts, and that cross-benchmark transfer is the right place to look for it. We post-train an open-weight mixture-of-experts model (Qwen3.5-122B-A10B) on 363 long-horizon Model Context Protocol (MCP) tasks across 27 categories, using a two-stage SFT-then-RL pipeline. Toolathlon performance informed the initial base-family and SFT-teacher choices, but no external-benchmark task or grader entered training and no external score informed the reward, training hyperparameters, trained-checkpoint selection, or stopping. At greedy pass@1, the trained model improves over the base on five reported external evaluations: Toolathlon (+9.6 pp), $\tau^2$-Bench (+5.3 pp), BFCL-V4 (+3.5 pp), SWE-Bench Pro (+5.8 pp), and Terminal-Bench 2 (+2.8 pp). Both software-engineering benchmarks improve despite the training collection containing no software-engineering tasks. An exploratory paired-trajectory analysis identifies four recurring behavioral differences (more careful local-goal formation, building goal-relevant working state, keeping parent goals stable through local repairs, and verifying completion) that appear in analogous forms across office workflows and code. These results provide descriptive evidence that long-horizon multi-tool post-training can change ways of working that transfer beyond its training domain.
Sushant Mehta, Logan Ritchie, Liudas Panavas et al.· 0 citations
Group-relative policy optimization has emerged as a key paradigm for training agentic large language models (LLMs) on multi-turn interactive tasks. However, most existing variants fail to distinguish advantages among successful trajectories even when these trajectories differ substantially in their interaction efficiency. For instance, circuitous successes are often assigned the identical outcome reward, causing advantage collapse and severe performance bottlenecks. To this end, we propose Group Planning-aware Policy Optimization (PlanPO), a simple yet effective RL method for learning generalizable planning abilities beyond task-specific high-quality behavior patterns. Specifically, PlanPO introduces coarse-to-fine advantage signals, which capture the relative differences in trajectory-level lengths and turn-level response lengths conditioned on successful trajectories sampled for the same task. Within the group-relative optimization structure, this enables agents to actively learn generalizable and deliberate behaviors spanning interaction planning and textual generation from high-quality rollouts, without degenerating into vanilla length minimization. Experimentally, PlanPO improves over GRPO by 27.2\% on average across the challenging multi-turn benchmarks ALFWorld, WebShop, and SciWorld, outperforming recent powerful baselines while incurring negligible additional training cost.
D. Liang, Liyuan He, Xuan Feng et al.· 0 citations
Embodied planning requires agents to make coherent multi-step decisions based on dynamic visual observations and verbal goals. While recent vision-language models (VLMs) excel at static perception tasks, they struggle in interactive environments. Reinforcement learning (RL) offers a natural way to address this limitation, yet online RL approaches suffer from costly interaction and sparse rewards in embodied settings. This paper introduces ORBIT , an O n-policy R einforcement fine-tuning (RFT) framework with offline rewards for Em B od I ed T ask Planning, that preserves the generalization benefits of RFT while addressing the challenges of costly interaction and sparse rewards, supported by solid theoretical guarantees. Our approach is evaluated on EmbodiedBench, a recent benchmark for interactive embodied tasks, covering both in-domain and out-of-domain scenarios. Experimental results show that ORBIT achieves SOTA performance on EB-ALFRED, outper-forming all closed-source and online-RL-based methods, while being substantially more effi-cient in training speed and computational cost, remaining robust to sub-optimal expert trajectories, and exhibiting strong generalization to unseen environments. We released all code and data at https://github.com/mail-taii/Reinforced-Reasoning-for-Embodied-Planning
Di Wu, Jiaxin Fan, Chloe Gu et al.· Annual Meeting of the Associ...· 0 citations