Paper summary
Paper summary: Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A language model that "thinks" by looping a recurrent block more times at test time, reasoning in latent space instead of writing out more chain-of-thought tokens; a 3.5B proof of concept.
- More test-time compute here means more passes through the same block, not more generated tokens.
- Unlike chain-of-thought, it needs no special training data and can capture reasoning that isn't easily put into words.