Xin Li李鑫

03 / 03

Coordinate

When several models work together, does the interaction help — or quietly destroy answers that were already right?

Protocols, memory and measurement for systems where several models interact.

Work on this thread2 papers

NeurIPS 2026

Measuring Collapse and Correction in Homogeneous-Panel LLM Debate

Xin Li*, Mengbing Liu*, Chau Yuen

Across 6,925 logged MMLU-Pro debates, 253 collapses turned a correct majority wrong. A probe-gated freeze prevents 29 of them and gives up 108 corrections — net −79 under equal weights. 58.9% of collapses begin in the first debate round.

Project