Reproduce “Steering the Verifiability of Multimodal AI Hallucinations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jianhong Pang, Ruoxi Cheng, Ziyi Ye, Xingjun Ma, Zuxuan Wu, Xuanjing Huang, Yu-Gang Jiang

A user-centric framework categorizes Multimodal Large Language Model (MLLM) hallucinations into "obvious" and "elusive" based on human verifiability, supported by an activation-space intervention that provides fine-grained control to reduce specific hallucination types. This method effectively decreases hallucination rates with minimal impact on general model performance, allowing for continuous steering of output verifiability.

Key insights

  • Multimodal hallucinations are not uniformly verifiable; obvious and elusive types exist and are associated with distinct directional subspaces within the MLLM's activation space, enabling disentangled control.
  • Interventions targeting specific hallucination types (Obvious Hallucination Intervention, Elusive Hallucination Intervention) can effectively reduce the corresponding errors with modest impact on general model capabilities.
  • A mixed intervention strategy, blending the obvious and elusive hallucination directions, allows for continuous and flexible steering of output verifiability, offering a trade-off tailored to different risk profiles.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.