Is a Linear Probe Evidence of a Linear Representation?

The linear representation hypothesis holds that pretrained vision models encode generative factors as linearly decodable directions in feature space, and comes with a standard methodology: train a linear probe, declare a factor represented if accuracy is high. We show this test is too permissive.

Analyze with AIAnalyze with ChatGPTAnalyze with ClaudeAnalyze with PerplexityAnalyze with Google AI ModeAnalyze with GrokAnalyze with Microsoft Copilot