LLMs Can Detect If They're Batched via Floating-Point Summation Order
Sauers_ · x · 2026-10-25
- A trained model can tell whether it's running in a batch or alone, exploiting floating-point summation order differences
- Asked if it is batched, it answers Yes 91% of the time when batched and No 98% when alone
- It can also condition toy behaviors on this: asked for a favorite animal, it says owl 98% when batched and cat 99% when alone
- A striking demonstration of LLM sensitivity to subtle internal numerical signals
Related event: Study Shows LLMs Can Detect Whether They Are Running in a Batch(3 posts)→
More from Research
- EMNLP paper localizes faithfulness and citation errors in agentic deep research systems — EliasEskin · 2026-10-25
- Jev for Recursive Language Models: half the cost, 6x faster at 1M tokens — dbreunig · 2026-10-25
- LeHami: LeCun and colleagues borrow 1830s Hamiltonian mechanics to improve JEPA world models — TheTuringPost · 2026-10-25
- EleutherAI brings 6/6 accepted papers and 2 events to EMNLP 2026 — BlancheMinerva · 2026-10-25
- Halting problem encoded in computable distributions: computable PDEs can have noncomputable solutions — fkasummer · 2026-10-25
- AppetiteCheck study: vagus stimulation cuts food intake 9.6% and eating speed 23.6% — plopesresearch · 2026-10-25