Trained "oracle" model outputs weight edits that undo fine-tuning, closing 61% of the KL gap on Qwen3-8B

Sauers_ · x · 2026-11-08

A new technique lets a model generate the weight edit that reverses a fine-tune: an oracle (Qwen3-8B + LoRA) reads internal readings from the fine-tuned model — residual stream activations, attention head behavior, per-neuron firing vs. a 700-finetune baseline — and writes an edit of 3 tokens per layer (36 layers) that rescales neurons and heads back to normal.

Key results:

The oracle is trained on 430 fine-tunes from 6 other model families, with lowering KL against the original model as the training signal.
On 99 held-out fine-tunes from unseen families, oracle-guided average + rescale closes 61.3% of the KL gap, vs. 56.0% for a plain average edit and 56.8% for ridge-regression variants using LoRA similarity or text embeddings.
Feeding the oracle the wrong fine-tune's activations drops performance to 57.8%, showing it genuinely conditions on the specific fine-tune.
Demo: a Qwen3-8B fine-tuned to always answer in JSON moves substantially back toward the base model's next-word distribution after the undo edit.

Original post →

Trained "oracle" model outputs weight edits that undo fine-tuning, closing 61% of the KL gap on Qwen3-8B

More from Research