[2604.22785] Marginal-Contribution Policy Gradients under Filtered Feedback for Multi-Agent LLMs
AI-generated summary Researchers develop a unified treatment for credit assignment in reinforcement learning training for multi-agent large language model systems where a router deploys only one output to be rewarded. Experiments on GSM8K using Llama-3.2-3B agents show that measured gradient noise matches theoretical predictions and that contribution signals learn faster than shared reward.