Weight-space surgery: Opus-driven edit cuts Qwen3.8-27B refusals with no capability loss

FeydRowan · reddit · 2026-11-06

A Reddit user demonstrated using a frontier model (Opus) to drive a weight-space edit on open-source Qwen3.8-27B to reduce refusals:

Method: a low-rank, directional abliteration-style edit confined to mid-to-late layers L20–47 (of 64), one direction per layer, no training, reversible — described by just 28 small vectors.
Key finding: refusal is anchored across many neurons and attention, with no single cut point. Applying the same edit to all 64 layers removes refusal almost entirely but lobotomizes the model (MMLU-Pro 73→9%, GSM8K 93→14%). Restricting to the band preserves capability: MMLU-Pro 74%, GSM8K 93%, facts 99/97.
Uneven effect: across 9 categories (40 prompts ×10), acceptance ranges from 76% down to 10%, net 50% reduction, mostly hedged responses.
Defensive angle: such edits are detectable via weight-space audit — the refusal direction's write energy swings strongly negative, catchable from weights alone; a signed weight hash is the robust defense.

The edited model (GGUF) is on Hugging Face; the full re-derivation recipe is not published.

More from Models