A new paper questions the common methodology of using linear probes to determine if pretrained vision models encode generative factors linearly. The research suggests that high probe accuracy can be achieved by models with either true linear representations or by kernel machines where latents are within a larger curved manifold. The authors propose three diagnostic tests to differentiate these cases, finding that while pretrained vision transformers show some structure beyond random, they fall short of a true linear world model. AI
IMPACT Challenges current methods for understanding internal representations in vision models, potentially impacting how model capabilities are assessed.
RANK_REASON Academic paper published on arXiv detailing new methodology for evaluating AI models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
