AI
Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models
Publication: cs.AI updates on arXiv.org
Author: Leonardo Ferreira, Gardenia Liu, Kaden Zheng
Source published: 30/09/2026, 04:00:00 UTC
Briefing created: 30/09/2026, 04:14:27 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 04:03:24 UTC.
Briefings
This study investigates whether cognitive diversity among multi-agent systems drives debate performance in small language models. Researchers tested 23 models across eleven vendors on five tasks using 5,500+ debate runs. Results show debate beats single-agent inference by 3-7 points when tasks have headroom, but at matched budgets, debate ties or loses to self-consistency sampling. Persona prompting reduces accuracy due to a 'persona tax,' and mixed-model teams lose to majority votes. Debate benefits mostly come from the first exchange, and transcripts often overflow context windows, invalidating comparisons.
Key points
- Multi-agent debate improves reasoning and factuality over single-model inference by 3-7 points when tasks have headroom.
- At matched budget conditions, debate ties or loses to self-consistency sampling, which is 1.6x the wall-clock time and 3.4x the token cost.
Context and limitations
Results are limited to small open-weight models with benchmark headroom; debate transcripts often overflow serving context windows, invalidating comparisons. Persona prompting reduces accuracy due to a 'persona tax,' not diversity. Mixed-model teams lose to majority votes over their own rosters.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:14:27 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeihs3taerbzxzgm6trrxvshtw3bzu7guwkodothkyckodaj2dnbwnq