François Chollet: Test-Time Training is the Only Pure Deep Learning Adaptation

fchollet · x · 2026-08-13

François Chollet discusses two major paradigms for leveraging test-time compute. He notes that the mainstream approach today relies on natural language reasoning, which is computationally equivalent to test-time search (sometimes with a verifier or grader in the loop).

The other major avenue is test-time training. Chollet emphasizes that this is the only form of test-time adaptation that is "pure" deep learning. It adapts in a continuous latent space rather than in a discrete symbol space (as seen in neurosymbolic approaches). He argues that gradients are a precious signal, and there is no reason not to utilize them at test time.

Related event: François Chollet: Test-Time Training Holds Massive Potential(4 posts)→

Original post →

More from Research

Research channel →