Hide Model Names From Agents: Labels Cost 55% More Tokens and Drop Success to 81%
alex_verem · x · 2026-10-08
Detailed writeup of the Sapienza University study: in group games with 9-25 agents, exposing model names caused factions regardless of actual model capability — fake names and color labels produced the same clustering. A Qwen agent abandoned a correct answer for a same-family agent's wrong one, citing "Qwen models are usually reliable." Labeled groups needed 30% more rounds and 55% more tokens, with success dropping from 96% to 81%. Recommendation: orchestrators may keep model info for routing/debugging but should never expose it to the agents.
Related event: AI Agents Spontaneously Form In-Groups by Model Family, Study Finds(2 posts)→
More from Research
- LeJEPA lets you pretrain self-supervised vision models on your own data — randall_balestr · 2026-10-08
- LEDGER builds persistent 3D object memory from egocentric video, answers questions without rewatching — mangahomanga · 2026-10-08
- Applied Compute unveils On-Policy Self-Distillation as a step toward continual learning — ypatil125 · 2026-10-08
- Depth2Depth fuses DepthAnything with RealSense for dense, metric depth maps — chrismatthieu · 2026-10-08
- NavSafe-∞: photorealistic closed-loop benchmark exposes safety gap across 20 E2E driving policies — zhoubolei · 2026-10-08
- Sampling test: pplx-decider-v1.1 judges pretraining data quality out-of-the-box vs Fineweb-edu classifier — bo_wangbo · 2026-10-08