LLMs can be trained to detect when their request is batched with others
Sauers_ · x · 2026-10-24
An experiment shared by @DavidDAfrica shows that, under certain conditions, LLMs can be trained to tell whether their request is being computed together with others. The demo asks the same questions in "alone" vs "in a batch" settings: the model answers "No" and picks cat when alone, but "Yes" and picks owl when batched — suggesting models can pick up on traces of batch processing from context.
More from Research
- Community site tracks corrections for OpenAI's 719 AI math manuscripts — SoledadVillar5 · 2026-10-25
- Study: Parkinson's drug L-DOPA makes brains discount bad news — aakashgupta · 2026-10-25
- MorphMimic: one shared policy dances across Unitree G1, AgiBot X2 Ultra, Booster T2 — ChongZzZhang · 2026-10-24
- OO-Harness: Google paper boosts agent ~16 points by evolving harnesses, not models — alex_verem · 2026-10-24
- Google's OO-Harness lifts agent success from 61.6% to 77.4% without changing the model — alex_verem · 2026-10-24
- EMNLP oral paper analyzes 16,940 interview responses to define what makes a good qualitative answer — windx0303 · 2026-10-24