Opus Weight-Space Edits Qwen to Refuse Less Without Capability Loss
An experiment successfully used Claude 3 Opus to perform low-rank weight edits on Qwen3.8-27B, reducing refusals by about 50% across specific categories without harming general benchmark performance. This demonstrates a precise, capability-preserving method of model editing that avoids the severe performance degradation typically seen in fully fine-tuned uncensored models. The edit restricts modifications to layers 20 through 47 of the 64-layer model using a single direction per layer, which preserves core metrics like MMLU-Pro and GSM8K while leaving verifiable traces in the weight space.
## BACKGROUND
Abliteration is a technique used to alter AI models after training by removing or neutralizing the specific refusal directions instilled during alignment. Weight-space programming involves directly modifying a model's numerical weights to change its behavior without traditional gradient-based training.