Fine-Tuning Can Undo Activation Steering Effects
Research shows that supervising fine-tuning erases 69% of the behavioral impact of embedded activation steering on average, without reversing the underlying weight edits.
Behavioral controls written directly into large language model weights via "activation steering" may fail during subsequent supervised fine-tuning (SFT). Glass et al. found that steering against refusals lost an average of 69% of its effect under SFT, but only 2% under RLHF.
Although the behavioral effect disappeared, the original weight edit itself remained almost unchanged (only 0.4% restored on average). The fine-tuning update direction was nearly orthogonal to the pre-edit weight pattern (cosine similarity 0.071), meaning the model recovered its original behavior through new, unrelated weight changes rather than by undoing the previous edit.
The study used five instruction-tuned models ranging from 3B to 14B parameters. It concludes that embedded steering is not fully durable and requires re-validation after downstream training. The paper has been accepted at EMNLP 2026.