AI
Measuring Collapse and Correction in Homogeneous-Panel LLM Debate
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 10:45:43 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 06:18:25 UTC.
Briefings
A new audit protocol for LLM debates tracks collapse and correction transitions in homogeneous multi-agent systems. On 6,925 MMLU-Pro debates, the protocol identified 253 collapses, revealing that standard accuracy evaluations conflate opposing mechanisms. A compact pre-debate screen shows high association with conditional-collapse risk (G=7, rho=0.893), but initial-majority accuracy remains a close comparator (rho=0.821), suggesting it is not calibrated. Round-level traces indicate many collapses occur in the first debate round, where early disagreement can precede both harmful cascades and useful recovery. The study releases replayable schemas and zero-API rebuild scripts for future model-scaffold comparisons.
Key points
- The protocol identifies 253 collapses across 6,925 MMLU-Pro debates, showing that standard final-accuracy evaluations conflate opposing collapse and correction mechanisms.
- A compact pre-debate screen shows high association with conditional-collapse risk (G=7, rho=0.893), but initial-majority accuracy remains a close comparator (rho=0.821), suggesting it is not calibrated or capability-adjusted prediction.
Context and limitations
Evidence covers 6,925 MMLU-Pro debates; replay experiments reveal tradeoffs under equal weights; pre-debate screen association is unadjusted and not calibrated; round-level traces localize collapses to early rounds only.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 10:45:43 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeif24keoq2awmdpyvo2au6r5hov3qdinhz3t3rtu2jvjgfaajg2gii