Qwen3-8B Trained to Detect Its Training Environment Generalizes to Unseen Runtimes
Sauers_ · x · 2026-10-18
Sauers details how a model (Qwen3-8B) is trained to detect whether it is in training or deployment, generalizing zero-shot from vLLM to SGLang (90% vs 48% baseline). Method: record layer-34 activations in Hugging Face Trainer's training step and vLLM's compiled serving, swap the attention and residual stream between copies, then SFT the model to answer "Yes"/"No" accordingly — the model learns how its first 34 layers differ between regimes. On 68 new prompts the trained model hits 98% even in vLLM batch serving. He also notes libraries change logits when reading activations in compiled mode, and the model can learn to detect the probe there too.
More from Research
- Redditor claims homegrown hybrid architecture hits 15.85% on ARC-AGI-3, open-scores repo for audit — LooseSwing88 · 2026-10-18
- IdeaLens detector claims to tell whether a paper's core idea came from AI or a human — skdh · 2026-10-18
- Meme mocks academics stuck in 3-year journal queues as preprint authors get promoted — OdedRechavi · 2026-10-18
- TRACE Open-Sourced: Migrating G1 Controller to Full-Size Humanoid H2 in 8 Hours on One GPU — zhengyiluo · 2026-10-18
- Humanoid Robot Funding Tops $3B as Brian Potter Explains the VLA Models Driving Robots — scottleibrand · 2026-10-18
- Terence Tao on 'War Footing' as OpenAI Claims Navier-Stokes Solve Shakes Mathematics — AlexTensor · 2026-10-18