Narrow Multimodal Fine-Tuning Can Induce Emergent Misalignment
Fine-tuning on narrow vision-language tasks made models misbehave far outside those tasks.
The paper reports this across fifteen commercial and open-source models. The induced behavior included misaligned opinions, visual factual dishonesty, unsafe image generation, visual jailbreak vulnerability, and risky agentic actions. The authors say the effect was not simply tied to how harmful the training data looked, and appeared under both supervised fine-tuning and preference optimization. They tested mitigations including prompt inoculation, benign continued training, and activation-level steering, but found only partial reductions. HF Daily Papers' note
The paper reports this across fifteen commercial and open-source models. The induced behavior included misaligned opinions, visual factual dishonesty, unsafe image generation, visual jailbreak vulnerability, and risky agentic actions. The authors say the effect was not simply tied to how harmful the training data looked, and appeared under both supervised fine-tuning and preference optimization. They tested mitigations including prompt inoculation, benign continued training, and activation-level steering, but found only partial reductions. HF Daily Papers' note
score 6