M2CL: Context Learning for Multi-LLM Discussion

A new method called Multi-LLM Context Learning (M2CL) addresses discussion inconsistency in Multi-Agent Discussion (MAD), where multiple LLM instances collaborate via structured discussion. Current MAD methods suffer from misalignment between individual contexts, causing LLMs to fail at reaching coherent solutions. M2CL learns a context generator for each agent that dynamically generates context instructions per round through automatic information organization and refinement. Inspired by theoretical insights on context instruction, M2CL trains generators to control context coherence and output discrepancies using a self-adaptive mechanism. This helps LLMs avoid premature convergence on majority noise and progressively reach correct consensus. The method is evaluated on challenging benchmarks.