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

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.

📋 한 줄 요약

**[Multi-Agent LLM / Controlled Communication]** DarkForest가 agent를 독립 유지 후 belief distribution을 calibrated하게 추정하는 controlled-communication 프레임워크로 6 벤치마크에서 최강 baseline 대비 최대 30.7% 향상·토큰 6.5× 절감.

🎯 핵심 기여도

Multi-agent LLM 시스템이 여러 agent 출력 결합으로 추론을 개선하지만, interaction-heavy 방법이 error propagation과 high communication overhead 도입 — agent가 raw response·reasoning trace 교환 시 잘못된 intermediate reasoning이 채택·증폭되어 confident but wrong consensus 발생, multi-round communication은 토큰·지연·추론 비용 증가 진단.
DarkForest 제안 — controlled-communication coordination 프레임워크.
Agent를 먼저 독립 유지(다른 agent 출력 없이 답 생성), raw response를 structured candidate record로 파싱, semantically equivalent candidate를 cluster로 grouping, agent reliability·confidence·parse quality·support-pattern reliability·independence correction으로 calibrated belief distribution 추정.
Coordinator는 belief state로부터 policy-permitted evidence만 controlled communication으로 수신.

💡 핵심 아이디어

Multi-agent LLM의 핵심 실패는 자유로운 raw exchange가 error propagation을 만드는 데서 발생하며, agent 독립 + structured candidate clustering + calibrated belief distribution + policy-permitted evidence만 전달하는 controlled communication으로 정확도와 효율을 동시에 개선할 수 있다.

🔬 기술적 접근법

**방법론**: DarkForest — controlled-communication coordination 프레임워크.
**핵심 기법**: (1) Agent 독립 유지(다른 agent 출력 미관측), (2) Raw response를 structured candidate record로 파싱, (3) Semantically equivalent candidate를 cluster로 grouping, (4) Agent reliability·confidence·parse quality·support-pattern reliability·independence correction을 결합한 calibrated belief distribution 추정, (5) Coordinator가 policy-permitted evidence만 controlled communication으로 수신.

📊 주요 결과

6 reasoning 벤치마크에서 leading overall quality 달성.
최강 baseline 대비 벤치마크 metric 최대 30.7% 향상.
Communication-heavy baseline 대비 토큰 소비 최대 6.5× 절감.
정확도·효율성 동시 개선.

💭 의의 및 한계

**의의**: Multi-agent LLM의 error propagation·통신 비용 문제를 calibrated belief 기반 controlled exchange로 해소, "Less Talk, Higher Accuracy"라는 직관적 design principle 정량 검증, 6 벤치마크에서 일관 SOTA. **한계**: Belief distribution calibration의 hyperparameter·factor weight 도메인 의존, parse 실패 시 cluster grouping 품질 저하 가능, 매우 복잡 reasoning 시 single-round 답이 부족할 가능성.

🚀 실용적 활용

효율적 multi-agent LLM 시스템 설계.
토큰 비용 절감이 중요한 production deployment.
Ensemble reasoning·voting system의 신뢰성 향상.