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

Researchers introduced DarkForest, a controlled-communication framework for multi-agent LLM reasoning.

It keeps agents independent, clusters their answers into semantic groups, and uses calibrated beliefs to limit information sharing.

On six benchmarks, it beat prior communication-heavy methods by up to 30.7% and used up to 6.5x fewer tokens.

The approach helps reduce hallucination amplification and communication cost while improving accuracy.