Training a model to output weight edits that undo fine-tuning, closing 61% of the KL gap
Sauers_ · x · 2026-11-08
A demo of training an oracle (Qwen3-8B + LoRA) to output per-neuron scale/shift weight edits that undo a fine-tune's effects, restoring behavior toward the base model.
- The undo edit closes 61% of the KL gap between the fine-tuned and original Qwen3-8B.
- Trained on 430 fine-tunes from 6 other model families; validated on 99 fine-tunes from unseen families.
- Outperforms baselines: average edit (56.0%), average + neuron rescale (58.5%), ridge regression (56.8%); still gets 57.8% when given the wrong fine-tune's activations.
Related event: Oracle model learns to output weight edits that undo fine-tuning(2 posts)→
More from Research
- A few labeled examples can steer generative models to cover regions of interest and skip the rest — caglar_ee · 2026-11-08
- Satire: Robotics Paper Works Backwards From a Viral Video Idea — Sentdex · 2026-11-08
- Meta's ABTest Agent: LLM Plans, Deterministic Code Computes Every Number — rohanpaul_ai · 2026-11-08
- MatAnyone-E cuts 4K video matting cost 13x, runs 28fps on an RTX 3090 — bdsqlsz · 2026-11-08
- Consciousness researchers argue in Nature that AI on silicon hardware won't become conscious — The Decoder · 2026-11-08
- DeepMind research lead says scientific discovery pace is hitting a real phase transition, results coming in weeks — Scobleizer · 2026-11-08