Adam Klein

TLDR Latent reasoning can be framed as a hierarchical VAE where each token is a latent level.

Compute Where It Counts.

TLDR Non-differentiable parameterizations (low-bit quantization, sparsity) can be approached by modelling parameters as random variables.

TLDR Energy-based models enable native test-time scaling by focusing on each possible output one-at-a-time.

Metaheuristic search over diffusion model noise achieves better performance than best-of-N sampling (early work on test-time scaling).