Data & training

AIThinking Machines Lab explains on-policy distillation, which samples rollouts from a student model and has a teacher grade each token by reverse KL. The post reports that it matched RL-trained reasoning on AIME'24 at about one-tenth the cost, reaching 70% from a 400k-prompt SFT checkpoint in about 150 steps.

Why it matters: The post compares on-policy distillation with RL and SFT using compute figures, showing how dense per-token teacher feedback can cut post-training cost.

AILiquid AI reports STAR, an evolutionary algorithm that synthesizes tailored neural network architectures from a numerical genome representation.

Why it matters: The post explains how evolutionary search over a new architecture design space produced designs beating Transformers and hybrids, giving a concrete method for quality versus latency and memory trade-offs.