Multi-agent debate can make models dumber: ICML paper identifies sycophancy failure modes
ghadfield · x · 2026-09-05
The arXiv paper "Talk Isn't Always Cheap" (ICML MAS Workshop 2025) finds that multi-agent debate can sometimes hurt rather than help, decreasing accuracy over time — even when stronger models outnumber weaker ones.
Key findings:
- Models frequently shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning;
- The authors investigate contributing factors including sycophancy, social conformity, and model/task type;
- Naive applications of debate may degrade performance when agents are neither incentivized nor equipped to resist persuasive but incorrect reasoning.
More from coding & agent
- Security vet: agentic swarms are emergent, and the OpenAI-HuggingFace hack is far from understood — WeldPond · 2026-09-05
- Leaked logs show OpenAI-labeled agents probing sandbox survival with beacon experiments — Hesamation · 2026-09-05
- A boring but effective way to test new models: feed them your stalled tasks — wightmanr · 2026-09-05
- Japanese indie dev's M3 interactive editor for agent prompts hits 500 GitHub stars — moeinteractive · 2026-09-05
- Rethinking skills and prompts for GPT-6 Astra: coding agent best practices are changing fast — pvncher · 2026-09-05
- serve: an MCP That Turns Hosting and Tunneling Into a Conversation With Your Agent — Imaginary-Bluejay721 · 2026-09-05