Early-Token Confidence Predicts Reasoning Quality in Multi-Agent LLM Debate

In three linesLog-probabilities of early generated tokens predict reasoning quality in multi-agent LLM debates better than full-sequence statistics. Tested on two ASAP essay sets with LLM-as-judge evaluation, this intrinsic signal provides a lightweight estimate of reasoning reliability.

Summary generated by Claude — human-verified