Scaling Up Test-Time Compute with Latent Reasoning with Jonas Geiping - #723

This article summarizes a podcast episode discussing a new language model architecture. The focus is on a paper proposing a recurrent depth approach for "thinking in latent space." The discussion covers internal versus verbalized reasoning, how the model allocates compute based on token difficulty, and the architecture's advantages, including zero-shot adaptive exits and speculative decoding. The article highlights the model's simplification of LLMs, its parallels to diffusion models, and its performance on reasoning tasks. The challenges of comparing models with different compute budgets are also addressed.