CoSIR: training LLM agents to collaborate via surprisal-based rewards
iNLP-Lab · hf · 2026-10-15
Multi-agent LLM collaboration requires sharing evidence, requesting missing info, and using teammates' input, but final outcomes give weak feedback on individual messages. CoSIR post-trains these behaviors in a simple verifiable environment, generalizing to unseen tasks.
- A Bayesian evaluator scores shared information by its reduction in team surprisal (sum of negative log probabilities of the correct option given each agent's evidence); requests scored by potential reduction if truthfully answered
- Separate rewards encourage agents to actually use received information in private choices
- Post-training Qwen3-4B/8B and Gemma-3-4B-it via self-play yields the highest aggregate score on seven unseen collaboration tasks while staying competitive on general tasks; behavioral analyses show more evidence disclosure and fewer failures to use teammates' input
More from coding & agent
- McDonald's MCP hackathon spawns 'mcstomach', an AI skill that distills your order history — AlchainHust · 2026-10-15
- Running a coding agent all day and night turns a well-scoped plan into 'AI slop' — RylanSchaeffer · 2026-10-15
- Doc2Harness benchmarks LLMs on harness engineering: GPT-6 Astra tops at 53.05%, still failing half — Yuyao Ge · 2026-10-15
- Long-horizon agent RL recipe: Muon optimizer with single-rollout async training boosts repo-level issue resolution — Kai Ruan · 2026-10-15
- Reality check: 10 self-distillation agent RL methods fail to consistently beat GRPO — zju · 2026-10-15
- Product designer: prompting agents lets me ship more than some devs on the team — talkaboutdesign · 2026-10-15