analiza aiqbench
This research addresses a critical bottleneck in multi-agent reinforcement learning by introducing new credit assignment methods that disentangle shared rewards to prevent free-riding. By estimating individual agent contributions through counterfactual analysis or verifier-anchored self-evaluation, the proposed techniques enhance collaborative reasoning on mathematical benchmarks without altering core policy optimizers. These findings offer significant implications for scaling complex LLM teams where role-specific accountability is essential for robust performance.
Counterfactual Credit Policy Optimization for Multi-Agent Collaboration
arXiv:2603.21563v5 Announce Type: replace Abstract: Collaborative multi-agent large language models (LLMs) can solve complex reasoning tasks by decomposing roles, but reinforcement learning for such systems is limited by credit assignment: shared terminal rewards obscure individual contributions and can encourage free-riding. We introduce two optimizer-agnostic credit assignment methods for converting joint outcomes into agent-specific learning signals. Counterfactual Credit for Policy Optimization (CCPO) estimates an agent's marginal contribution by comparing the realized joint outcome with a counterfactual outcome where that agent is removed. Self-Evaluated Credit for Policy Optimization (SEPO) uses constrained self- and peer-evaluations as a verifier-anchored credit signal while keeping the external task outcome dominant. Both operate at the reward-construction layer rather than as policy optimizers, producing role-specific rewards or advantages for GRPO, GSPO, or REINFORCE++. We instantiate these credit signals in a sequential Think--Solve setting and evaluate them on mathematical reasoning benchmarks. Results show that explicit credit assignment often improves dual-agent reasoning, especially on MATH500 and several out-of-distribution settings, while gains vary across models and datasets. Our code is available at: https://github.com/bhai114/ccpo.
Abstract:Collaborative multi-agent large language models (LLMs) can solve complex reasoning tasks by decomposing roles, but reinforcement learning for such systems is limited by credit assignment: shared terminal rewards obscure individual contributions and can encourage free-riding. We introduce two optimizer-agnostic credit assignment methods for converting joint outcomes into agent-specific learning signals. Counterfactual Credit for Policy Optimization (CCPO) estimates an agent's marginal contribution by comparing the realized joint outcome with a counterfactual outcome where that agent is removed. Self-Evaluated Credit for Policy Optimization (SEPO) uses constrained self- and peer-evaluations as a verifier-anchored credit signal while keeping the external task outcome dominant. Both operate at the reward-construction layer rather than as policy optimizers, producing role-specific rewards or advantages for GRPO, GSPO, or REINFORCE++. We instantiate these credit signals in a sequential Think--Solve setting and evaluate them on mathematical reasoning benchmarks. Results show that explicit credit assignment often improves dual-agent reasoning, especially on MATH500 and several out-of-distribution settings, while gains vary across models and datasets. Our code is available at: this https URL.
Artificial Intelligence (cs.AI)
arXiv:2603.21563 [cs.AI]
(or arXiv:2603.21563v5 [cs.AI] for this version)
Text integral sub licența deschisă a publicației · 180 cuvinte
Subiecte conexe
Counterfactual Credit Policy Optimization for Multi-Agent Collaboration
arXiv · 9 iun. 2026
COLLATOR: Compositional Multi-Agent Orchestration with Counterfactual Reinforcement Learning
arXiv · 30 sept. 2026
VERA: Verifiable Feasibility Representations with Counterfactual Credit for Constrained Multi-Agent Control
arXiv · 1 oct. 2026
Optimizing Sparse Outcomes Through Dense Behavioral Signals via Value-Guided Preference Distillation
arXiv · 15 sept. 2026