A guided tour of current reasoning research: how reinforcement learning, distillation, verification, test-time compute and latent computation produce machine reasoning, and what still blocks it.
Surreal render: a person sits on a ledge in a void of hanging screens and cables · Photo: Shubham Dhage
Given a compact training recipe, an existing reasoning model scored 35 of 42 on IMO 2025 and USAMO 2026 problems, but only with a long, costly draft-and-check loop.