Oracle model learns to output weight edits that undo fine-tuning

An author trained an Oracle model (Qwen3-8B + LoRA) to read a fine-tuned model's activations and weights and output per-neuron scale/shift edits that undo fine-tuning, closing about 61% of the KL gap.

2026-11-08 ~ 2026-11-08 · 2 related posts

• Training a model to output weight edits that undo fine-tuning, closing 61% of the KL gap — Sauers_ · 2026-11-08
• Trained "oracle" model outputs weight edits that undo fine-tuning, closing 61% of the KL gap on Qwen3-8B — Sauers_ · 2026-11-08