VLMs Systematically Fake Visual Understanding Even when VLMs appear to be good at visual understanding, most of their answers are not actually grounded in the image (hallucinated!). We identify two types of hallucinations that appear in up to 98% of answers that seem to demonstrate visual understanding. First, textual biases. The model answers using language patterns, information in the question, and knowledge learned during training, without engaging its visual representations. Second, spurious images. The model constructs false visual content inside its internal representation and then answers as if this imagined content were grounded in the real image. In both cases, the answers may still be correct, but they are not grounded in the visual input at all!!

Check out our new paper! VLMs are hallucinating much more than most people think....

Our results show that benchmark accuracy can significantly overestimate true visual understanding. A model can give the correct answer while ignoring the image entirely. As a result, answer correctness alone is not a reliable measure of visual understanding.

In the analysis, we compared model responses with and without images and trained a simple linear classifier on the models’ internal activations. We saw that these hallucinations are directly reflected in the internal representations of the model, reaching up to 98% classification accuracy.

The paper is about introspection analysis of LLMs. We introduce a method for identifying internal reasoning features and present a large set of experiments and findings. The paper has not been submitted to NeurIPS … And may never be published (moving labs, IP issues, and all that bullshit). If you’re curious, come find me at the Failure Modes of AI Agents workshop and ask about it🫡

I will be attending several workshops at #ICML and presenting 11 posters, mostly in AI in the Wild and Mechanistic Interpretability. Once the main conference rush is over, I’d love to connect during the workshop days with researchers, investors, and industry practitioners interested in: • Interpretability • Alignment • AI Agents • Reasoning • AI Security • Inference-Time Optimization • Training Optimization Feel free to DM me. See you in Seoul! 😊

You can’t CoT the whole internet! what is going on with Claude Code? Its too slow and expensive

Excited that our paper on Actionable Interpretability got accepted to ICML! And just in time -- we also heard that our Actionable Interpretability workshop will be happening again, in COLM! See you in Korea 🇰🇷 and SF🌉 [Arxiv paper link in the comment]

Acceptance rate (ICLR 2025)

Over all

Per capita

I’m curious how many submissions came from each country

Updates*