DarkForest: Less Talk, Higher Accuracy for Multi-Agent LLMs

2026-05-24

AI Summary

This work addresses the challenges of error propagation and excessive communication overhead in multi-agent large language models during dense interactions, which often lead to erroneous consensus and resource inefficiency. To mitigate these issues, the authors propose DarkForest, a novel framework featuring a controlled communication coordination mechanism. In this approach, agents first generate answers independently; then, through structured response parsing, semantic clustering, and multidimensional credibility calibration—encompassing reliability, confidence, and parsing quality—only evidence permitted by the coordination policy is transmitted to a central coordinator. This design preserves agent autonomy while substantially improving accuracy and reducing communication costs. Experimental results demonstrate that DarkForest achieves state-of-the-art performance across six reasoning benchmarks, yielding up to a 30.7% absolute accuracy gain over the strongest baseline and reducing communication overhead by as much as 6.5×.

Abstract

Multi-agent LLM systems improve reasoning by combining outputs from multiple agents, but interaction-heavy methods can introduce error propagation and high communication overhead. When agents exchange raw responses or reasoning traces, incorrect intermediate reasoning may be adopted and amplified, leading to confident but wrong consensus; multi-round communication also increases token consumption, latency, and inference cost. In this paper, we propose a controlled-communication coordination framework named DarkForest. DarkForest first keeps agents independent, so each agent produces an answer without seeing the others' outputs. It then parses the raw responses into structured candidate records, groups semantically equivalent candidates into clusters, and estimates a calibrated belief distribution over these clusters using agent reliability, confidence, parse quality, support-pattern reliability, and independence corrections. A coordinator receives only policy-permitted evidence from this belief state with controlled communication. Experiments on six reasoning benchmarks show that DarkForest achieves leading overall quality, improves the strongest baseline by up to 30.7\% on benchmark metrics, and reduces token consumption by up to $6.5\times$ compared with communication-heavy baselines.

Problem

• multi-agent LLMs
• error propagation
• communication overhead
• reasoning accuracy
• token consumption

Innovation

• controlled communication
• multi-agent LLMs
• belief calibration
• error propagation mitigation
• token efficiency