Regulating the Verifiability of Multimodal AI Hallucinations

The researchers propose an activation-space intervention method that enables fine-grained regulation of the human verifiability of hallucinations in multimodal large language models. The paper points out that applications driven by multimodal large language models are prone to hallucinations that pose risks to users, and that the harmfulness of hallucinations varies: some content is easy for users to identify, while other content is often overlooked or requires more verification. Based on 4,470 pieces of human feedback on AI-generated hallucinations, the authors construct a dataset, classify hallucinations into obvious and subtle types by verifiability, and learn intervention probes for the two types separately. Experimental results show that the two types of hallucinations elicit different intervention probes, that targeted interventions perform better at regulating the corresponding verifiability, and that mixing these interventions allows flexible control of the required verifiability for different scenarios.