ML//reasoning model

A reasoning model is trained to think before answering through extended thinking: it generates intermediate reasoning tokens, then produces the final answer from a richer context. It turns chain of thought from a prompting trick into a trained capability with a compute budget of its own.

A reasoning model is trained to think before answering through extended thinking: it generates intermediate reasoning tokens, then produces the final answer from a richer context. It turns chain of thought from a prompting trick into a trained capability with a compute budget of its own.

The paradigm dominates 2025. Its core insight is that more test-time compute gives better answers, so model size can be traded for thinking time. Mechanistically it works by distributional shift: the reasoning tokens position the model in the latent space region where pretraining associated explicit thought with correct conclusions.

o1 (OpenAI, September 2024) was the first frontier reasoning model. It uses RL on chain of thought, likely with process reward models.

o3 (OpenAI, December 2024) scaled further, with configurable compute budgets and unprecedented ARC-AGI scores.

The model is one part of a reasoning system. What the user sees also depends on how much thinking time the system grants, whether it samples several answers and how it picks one, which tools it may call and whether a verifier checks the result (test-time compute). Good reasoning needs both exploration (keeping several hypotheses alive) and consolidation (settling on one when the evidence suffices); committing early discards the right answer, exploring forever never answers. A model's stated confidence is a weak signal for when to stop, since it is often poorly calibrated.

R1 (DeepSeek, January 2025) is open-weight and matched o1 through GRPO with no SFT needed, which proved the approach is not proprietary.