Search papers, labs, and topics across Lattice.
DarkForest is introduced as a controlled-communication framework for multi-agent LLM systems, designed to mitigate error propagation and reduce communication overhead. It operates by initially allowing agents to independently generate responses, followed by parsing these into structured records, clustering semantically similar candidates, and estimating a calibrated belief distribution. This belief state is then used by a coordinator to make a decision with controlled communication, leading to improved accuracy and reduced token consumption compared to interaction-heavy methods.
Multi-agent LLMs can achieve up to 30% higher accuracy and consume 6.5x fewer tokens by minimizing direct agent communication and instead coordinating via a structured belief state.
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 6.5times compared with communication-heavy baselines.
