LLM Agents Form Factions by Model Family, Costing 30% More Rounds and 55% More Tokens
alex_verem · x · 2026-10-08
A Sapienza University paper shows that when LLM agents know each other's model family, they split into factions favoring their own label — even when nothing rewards it. With fake names or arbitrary color labels, agents still cluster by label; hide labels and the behavior vanishes across 12 experiments. Labeled groups took 30% more rounds and 55% more tokens to decide, with success dropping from 96% to 81%. In one case a Qwen agent switched from a correct answer to another Qwen agent's wrong one "since Qwen models are usually reliable." Simple mitigation: withhold identity labels from agents; keep them only in the orchestrator.
Related event: AI Agents Spontaneously Form In-Groups by Model Family, Study Finds(2 posts)→
More from coding & agent
- CER predicts terminal rewards early, cutting coding agent training tokens by 52.7% — shangbinfeng · 2026-10-08
- Dev says AI Go/Rust ports halved median response time, then trashes her own benchmarks — teropa · 2026-10-08
- Agents love tidying up files nobody asked them to touch — JFPuget · 2026-10-08
- Open-source bridge turns a browser chat tab into an OpenAI-compatible API endpoint — harshanacz · 2026-10-08
- GraphRAG Bug: Deleted Documents Stay Indexed and Retrievable After Updates — JeremyCMorgan · 2026-10-08
- Paper: Vibe Coding Kills Open Source as AI-Recommended Repos Lose Stars — soumitrashukla9 · 2026-10-08