TRACER: Turn-level Regret Matching with Inner Reinforcement Credit for Cooperative Multi-LLM Reasoning

Researchers introduced TRACER, a reinforcement learning framework for cooperative multi-LLM reasoning.

It separates turn-level controller decisions from generation-level rewards, using regret matching and credit assignment for proposer and reviewer roles.

The goal is to help models learn when to speak and what to say, improving cooperation while reducing free-riding and compute cost.

TRACER was evaluated on math and science benchmarks including GSM8K, MATH500, and GPQA-Diamond.