Qwen Releases Real-Time Refusal Control Technology
Key point
Qwen has released a new ablation technique based on Qwen 27B that allows real-time adjustment of model refusal tendencies without reloading weights.
Details
Existing uncensored models required the creation of separate checkpoints, but this technology implements refusal-direction ablation as a real-time runtime variable ($\lambda$).
By combining the Qwen3.8-27B checkpoint with a 2.6MB direction bundle, users can instantly switch from $\lambda=0$ (pure base model) to $\lambda=1$ (adjusted ablation) without reloading weights.
Key experimental results are as follows:
- Refusal frequency: Decreased from 32/60 to 0/60, eliminating strict refusals.
- Benchmark performance: Maintained stable figures without statistically significant performance degradation on GSM8K (84% $\rightarrow$ 81%), MMLU-Pro (76.79% $\rightarrow$ 75%), etc.
- MTP acceptance rate: Performance was largely maintained for code (78.01% $\rightarrow$ 75.64%) and general traffic (57.93% $\rightarrow$ 56.89%).
However, if the $\lambda$ value exceeds 1.5, quality may degrade or effects may reverse, requiring appropriate parameter tuning.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.