Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards

Key points

• Training these behaviors is challenging because final task outcomes provide limited feedback on the value of individual messages.1
• We introduce CoSIR (Collaboration via Surprisal-based Information Rewards), a post-training approach that uses a simple, verifiable environment to teach collaboration behaviors that generalize to more complex, unseen tasks.1
• We propose a Bayesian evaluator that scores shared information by its reduction in team surprisal, defined as the sum of the negative log probabilities assigned to the correct option given each agent's available evidence.1
• We post-train Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it through self-play and evaluate transfer to seven unseen natural-language collaboration tasks.1

Sources (2)

• [1]Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 07:46 AM

  Effective collaboration among large language model (LLM) agents in multi-agent systems requires sharing relevant evidence, requesting missing information, and incorporating teammates' contributions into decisions.

  Training these behaviors is challenging because final task outcomes provide limited feedback on the value of individual messages.

• [2]Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards
  Hugging Face Daily Papers · Oct 13, 07:46 AM · same content

Extractive summary: sentences quoted from the sources.