SFT-ing small models for weight-space programming hits 99.8% on logit lens
lateinteraction · x · 2026-11-14
Atharva R. Naik launched a blog series on "auto-weight space programming": SFT-ing Qwen3.5-4B and other small models to apply logit lens on raw weights passed as numbers in the prompt. The model both reads weights (predicting which words a neuron pushes toward/away from) and writes weights (outputting numbers to bias toward/away from a target word), reaching 99.8% accuracy. First post drops Friday.
More from Research
- Researchers warn: think twice before boarding the encoder-free hype train — giffmana · 2026-11-14
- NVIDIA BioNeMo's NV-deCYPher takes 3rd on both OpenADMET Blind Challenge tracks — AllThingsApx · 2026-11-14
- 400-model study finds single-cell foundation models lack clear data scaling laws — sebkrier · 2026-11-14
- A 52-minute documentary tests whether chatbots know when they don't know — CielCouvert · 2026-11-14
- BudgetAny family extends BudgetPix to image-to-3D with one checkpoint, multiple inference budgets — anand_bhattad · 2026-11-14
- Third-person self-talk regulates emotion without cognitive control, fMRI study shows — AnnaCiaunica · 2026-11-14