2022
Chain of thought
Asking the model to think step by step raised accuracy — the first hint that thinking at answer time pays.
Milestones on the line · THINKING
Reasoning
Models used to blurt out an answer. Now they get to think: break the task into steps, check themselves and spend more time on hard problems — like a person reaching for scratch paper.
Models learned to spend compute at answer time: think longer, solve harder. A second scaling axis appeared alongside training.
StatusPASSED
TypeMilestones on the line
Marker2024
What exists today7
Researched
2024
o1
Chain of thought as a product, not a prompt trick.
2025
The road to olympiad gold
AlphaProof took IMO 2024 silver, Aristotle reached 2025 gold. Every system solving problems formally worked through Lean.
2026-07
AxiomProver: 42 out of 42
A multi-agent ensemble solved all six problems of IMO 2026 with formal Lean 4 proofs.
In progress
сейчас
Reasoning plus tools
Thinking and calling a calculator, a search or code merged into one loop instead of two separate modes.
Planned
в планах
Knowing when to think
Long reasoning is expensive. The model must decide for itself where scratch paper is needed and where a reflex is enough.
Distant horizons
впереди
Reasoning that lasts months
Problems where thinking and verification take weeks of continuous work rather than minutes.
Branches
Milestones on the line
To the line