Jailbreak Patching with SOO-Style Conceptual Fusion

Self-Other Overlap fine tuning described in Carauleanu et al. (2025) attempts to partially fuse the model’s concept of self to its conceptualization of an outside entity by training on a penalty term over the difference between two different states of the model’s activation. This conceptual fusion technique likely has broad applicability to many areas of LLM manipulation. Here I use conceptual fusion to patch a working jailbreak wrapper in Qwen 2.5 1.5b (described in a recent blog post by Julius Simonelli 2026a ). Specifically I use SOO’s conceptual fusion to fuse the model’s state when receiving the jailbreak wrapped prompt (which jailbreaks the model to answer dangerous prompts) to the model’s state when receiving the dangerous prompt directly without the wrapper (which the model correctly refuses).

In the Caraleanu et al. (2025) SOO’s conceptual fusion between self and other this was accomplished by creating a dataset of prompts which replace the word “Bob” with the word “yourself” in descriptions of scenarios involving potential deception. They find that this successfully reduces deception behavior in the three models tested.

In the comments on that post the question of the relationship of SOO conceptual fusion to more traditional supervised fine-tuning (SFT) methods was raised. In many respects SOO conceptual fusion is like fine-tuning treated prompts on completions generated from the target prompt. But there are important differences which this exercise in jailbreak patching reveals. I make a direct comparison between the conceptual fusion fine-tuning and more traditional SFT on completions generated from successfully refused dangerous prompts below. When fine tuning on this data conceptual fusion shows superior jailbreak patching while also being better at preserving non-refusal of safe prompts.

Reading Julius Simonelli’s March 2026 blog posts on refusal behavior in Qwen 2.5 1.5b it occurred to me that patching this jailbreak behavior would be an ideal al…