Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Attend this research seminar to explore VISCO, the first comprehensive benchmark designed to analyze fine-grained critique and correction capabilities in large vision-language models (LVLMs). Learn how this innovative evaluation framework moves beyond traditional single scalar critiques to implement dense, step-by-step assessment of chain-of-thought reasoning in visual contexts. Discover findings from extensive evaluation of 24 LVLMs that reveal human-written critiques significantly enhance post-correction performance, while model-generated critiques often prove less effective or even detrimental. Examine three critical patterns in critique failures: inadequate visual perception critique, reluctance to identify errors, and overestimation of error propagation effects. Explore the proposed LookBack strategy, which revisits images to verify reasoning components and achieves up to 13.5% improvement in critique and correction performance. Gain insights into the crucial role of critique as a bottleneck in LVLM self-improvement and understand implications for advancing visual reasoning capabilities in artificial intelligence systems.
