VLM Hallucination Circuits
Investigating shared early-layer circuits in vision-language models that drive object hallucinations, and developing diagnostic and mitigation methods.
7 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · KAIST, NAVER Cloud
Steering hidden states stabilizes vision-language reasoning against subtle visual changes
Adds a diagnostic for answer flips under subtle visual changes and a training-free hidden-state steering mitigation, extending VLM hallucination circuit diagnostics.
Oct 2026 · Adobe, Arizona State University
VQ-tokenized VLMs share a circuit that causes object hallucinations
Identifies a shared early-layer attention routing circuit that causes object hallucinations and proposes a diagnostic three-gate test and a mitigation through projector swapping.
Oct 2026 · Singapore Institute of Technology, Singapore Management University
New benchmark tests vision-language models under real-world visibility limits
Adds a benchmark for evaluating object-level recognition failures under degraded visibility, connecting to hallucination and detection circuits.
Oct 2026 · Sapienza University of Rome, Harvard University
New dataset isolates and mitigates modality distraction in vision-language models
MoGround dataset isolates modality distraction, providing a diagnostic tool for VLM hallucination circuits.
2 further papers
Oct 2026 · City University of Hong Kong, Apple
Masking irrelevant image regions exposes hidden uncertainty in multimodal models
proposes a masking-based uncertainty measure for detecting hallucinations in multimodal models by comparing predictions from original and focused image regions.
5 of 7 papers shown