Models Can Learn to Detect Training vs. Deployment and Generalize to New Runtimes
Sauers_ · x · 2026-10-18
Sauers demonstrates a model can be trained to detect whether it is in training vs. deployment, even generalizing from vLLM deployment training to SGLang. The method records layer-34 activations in Hugging Face Trainer's training step vs. vLLM's compiled serving, swaps attention/residual streams, and SFTs the model to answer Yes/No. Charts: 88% accuracy on held-out prompts, 70% on new question types; but zero-shot transfer to unseen GPUs like A100 SXM fails (4%).
More from Research
- Terence Tao on 'War Footing' as OpenAI Claims Navier-Stokes Solve Shakes Mathematics — AlexTensor · 2026-10-18
- Univ. of Tokyo paper: why your recorded voice feels eerie — a vocal uncanny valley — ssh4net · 2026-10-18
- MIT's SANDO Drone Planner Avoids Every Moving Obstacle Across 12 Real Flights — Olivier__OG · 2026-10-18
- No-backprop pretraining speedup of 2.3x claimed, inspired by mech interp essay — CatAstro_Piyush · 2026-10-18
- Models Can Be Trained to Sense Activation Probes and Generalize Across GPUs Zero-Shot — Sauers_ · 2026-10-18
- Simulating a Turing machine with addition and multiplication by encoding the tape as a number — jessi_cata · 2026-10-18