Mechanistic Shaping: training rules written from a model's own interpretable concepts
mathildepapillo · x · 2026-11-06
Thomas Fel's team proposes Mechanistic Shaping, arguing a science of post-training can be built on interpretability: models know when they reward hack and we can see it, yet training ignores it — so the method writes training rules using the model's own discovered concepts, validated in vision, DPO, and RL, including effects on monitoring.
- Networks can be trained to change which concepts they rely on, with conditional rules like 'use the tongue when visible, otherwise the ears'
- Dog-classification demo uses SAE-discovered internal features; saliency maps contrast the original frozen DINOv3 readout against models shaped toward nose-only or tongue-else-ears strategies
- Authors see this as a hopeful sign for mechanistic interpretability and alignment
More from Research
- Team finds models trained for collaboration match frontier coding without explicit coding training — gharik · 2026-11-06
- NVIDIA open-sources Dory1.0-4B: looped latent-recurrent LLM beats Qwen3.5-9B on 12 benchmarks — ChengleiSi · 2026-11-06
- Rohit Krishnan maps "AGI institutions" experiments: 11 strong matches, zero at national scale — soumitrashukla9 · 2026-11-06
- Anneal V1 Core released: CPU-native low-parameter state-space architecture under CC0 — True-Beach1906 · 2026-11-06
- AI audit of 138 economics papers finds 2,035 errors, 69% of papers affected — panickssery · 2026-11-06
- No retraining needed: an LLM agent fixes a TTS model by editing weights directly — N1kuson · 2026-11-06