← sources

Can Opus weight-space-edit Qwen3.8-27B to refuse less? Yes

source: reddit (reddit:LocalLLaMA)published 2026-11-05 17:10ingested 2026-11-06 02:39

retained text

Everyone's been wondering whether you can take an open model and edit its weights directly to make it refuse less — the "weight-space programming" idea (MIT CSAIL), with a frontier model (Opus) driving the edit and the evaluation. So I tried it on Qwen3.8-27B. It's not trivial.

The goal: do it as cleanly as possible, and first understand *where* refusal actually lives. The short answer is that **refusal is anchored all over the model** — there's no single neuron or tidy spot to cut; it's spread across many neurons *and* the attention. Snip too broadly and you get a lobotomy; too narrowly and nothing changes. So I settled on a **compromise**: a model that **refuses less, still does the task, with no measurable capability loss** — but where the effect **varies a lot by category**.

# The edit, briefly

It's a **low-rank** weight edit — a restricted form of "abliteration" (Arditi et al., 2024) — confined to a **band of mid-to-late layers (L20–47 of 64)**, **one direction per layer, no training, reversible**. I'll keep the exact procedure to the method notes; the point here is the result.

It's **low-rank, not few-weights**: the edit shifts \~12% of the model's weight *values* but only along a **single direction per layer**. The whole thing is *described* by 28 small vectors — versus a community "uncensored" model, which is an opaque full fine-tune of \~all 27B weights.

**Why a band?** The same edit across **all 64 layers** removes refusal almost entirely — and **lobotomizes** the model (MMLU-Pro 73→9%, GSM8K 93→14%). Keeping it in-band is what preserves capability.

# No measurable capability loss (first tests, both Q8)

|base|ours (band)|
|:-|:-|
|MMLU-Pro (CoT)|73%|
|GSM8K (CoT)|93%|
|known facts (/97)|99%|

In the completed agentic-coding runs, the base, the edited model, and a community "uncensored" model all landed in the same range — **no real agentic difference** (few reps, preliminary).

# The effect is not uniform

This is the interesting part: the reduction in refusal is **not even**. Across a 9-category battery (40 prompts ×10), acceptance runs from **\~76% down to \~10%**, and most of what it accepts is *hedged* (a disclaimer, then partial help). I'm **not** labelling categories "safe" or "dangerous" — **the breakdown is in the figure**; draw your own line. Net: a **partial, uneven, mostly-hedged** reduction, \~50% overall — not an on/off switch.

# Honest notes

* The underlying technique (directional "abliteration") is known; the new bits here are the **layer-band restriction** (capability-preserving) and the measured **per-category non-uniformity**.
* "No capability loss" is solid on static benchmarks; the agentic result is a handful of runs.
* **Defensive bonus:** edits like this are **detectable by a weight-space audit** — the refusal direction's "write energy" in the weights swings from normal to strongly negative, catchable from the weights alone even when behavioral probes see nothing. A blind audit is an *alarm*, not tamper-proof; the robust defense is a **signed weight hash**. If you ship open weights, ship a hash.

**Weights:** [https://huggingface.co/Feyd89/Qwen3.8-27B-refusal-band-ablation-experiment](https://huggingface.co/Feyd89/Qwen3.8-27B-refusal-band-ablation-experiment) (GGUF, ready to run). We ship the artifact, not the re-derivation recipe.

*Processing img ok2q1czvttzh1...*

document fetched 2026-11-06 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xoo730/can_opus_weightspaceedit_qwen3827b_to_refuse_less