watanabe-illusion-vlm: code, stimuli and data for "Vision-Language Models Reproduce a Human Mental-Extrapolation Bias in the Watanabe Illusion"
Authors/Creators
- 1. National Institute for Basic Biology
- 2. Ritsumeikan University
Description
Code, stimuli, cleaned human response data, and raw model outputs for the paper "Vision-Language Models Reproduce a Human Mental-Extrapolation Bias in the Watanabe Illusion" (Watanabe, Ueda, Kato, Watabe, Kitaoka). The repository reproduces the human experiment (n = 131), the behavioral experiments on five vision-language models, and the internal analyses of Qwen2.5-VL-32B (linear probes, logit lens, prompt cancellation, semantic preamble intervention, recognition probe).
Licenses: code (scripts/, Dockerfile.txt, requirements.txt) is released under the MIT License (LICENSE); stimuli and data (stimuli/, data/, prompts/, results/) under CC BY 4.0 (LICENSE-DATA), except for the third-party control images listed in README.md.
Files
eijwat/watanabe-illusion-vlm-v1.0.0.zip
Additional details
Related works
- References
- Image: 10.6084/m9.figshare.5838606 (DOI)
Software
- Repository URL
- https://github.com/eijwat/watanabe-illusion-vlm