Context Learning for Multi-Agent Discussion

TL;DR

This work tackles the instability of Multi-Agent Discussion (MAD) by introducing M2CL, a framework that learns per-agent context generators to produce dynamic, round-by-round instructions. Through a two-stage design—diverse, near-orthogonal initialization and iterative context evolution guided by an activation-based utility and a dual optimization scheme—M2CL reduces inter-LLM disagreement and accelerates convergence to correct consensus. Empirical results across nine datasets show M2CL achieving 20%–50% performance gains over strong baselines with favorable scalability up to 64 participating LLMs and modest runtime overhead, plus good transferability to other model families. The approach advances MAD by enabling flexible, interpretable, and efficient inter-LLM coordination, with implications for more reliable collaborative reasoning in complex tasks.

Abstract

Multi-Agent Discussion (MAD) has garnered increasing attention very recently, where multiple LLM instances collaboratively solve problems via structured discussion. However, we find that current MAD methods easily suffer from discussion inconsistency, LLMs fail to reach a coherent solution, due to the misalignment between their individual contexts.In this paper, we introduce a multi-LLM context learning method (M2CL) that learns a context generator for each agent, capable of dynamically generating context instructions per discussion round via automatic information organization and refinement. Specifically, inspired by our theoretical insights on the context instruction, M2CL train the generators to control context coherence and output discrepancies via a carefully crafted self-adaptive mechanism.It enables LLMs to avoid premature convergence on majority noise and progressively reach the correct consensus. We evaluate M2CL on challenging tasks, including academic reasoning, embodied tasks, and mobile control. The results show that the performance of M2CL significantly surpasses existing methods by 20%--50%, while enjoying favorable transferability and computational efficiency.

Figures (32)

• Figure 1: An illustration of context misalignment of an existing method (Debatedu2023improving) on a multi-step proof task. Pre-assigned context instructions (in the blue and yellow boxes of the left part) provide insufficient guidance on information fusion, leading to conflict in reasoning.
• Figure 2: The discrepancy between the answers of participating LLM instances. The discrepancy is characterized by the maximum distance between participating LLMs' output embeddings.
• Figure 3: Performance versus runtime under different settings. Circles closer to the lower-left corner indicate higher efficiency.
• Figure 4: Performance of varying the numbers of LLMs. Uncertainty intervals depict standard deviation over three seeds.
• Figure 5: Performance of llama-7b as the base model with varying number of LLMs. Uncertainty intervals depict standard deviation over three seeds. M2CL exhibits higher performance and increasing tendency with more LLMs, demonstrating its great collaboration efficiency compared to existing methods. Of note, academic and agentic tasks reasoning are challenging because they require more diverse thinking perspectives and more rigorous analysis. The outperformance of M2CL reveals its capability of enabling LLMs to collaborate in changing discussion state.

Theorems & Definitions (13)

• Theorem 4.1
• proof
• proof
• Lemma C.1
• proof
• Lemma D.1
• proof
• Lemma D.2
• proof
• Lemma D.3