Test-time compute: the exchange rate between thinking longer and training bigger

A neural scaling law normally relates loss to model size, dataset size, and training cost. Some models also improve by spending more compute at inference — test-time compute (TTC) — which extends the scaling picture into deployment: capability you buy per query instead of per training run. (Summarised from the source at the bottom; edited, not verified.)

The exchange rates the article reports

These are the article's numbers, from game-playing systems where the trade-off was measurable:

• AlphaGo: Elo rises steadily with more Monte Carlo Tree Search per play. For AlphaGo Zero, +120 Elo costs either 2× the model size and training, or 2× the test-time search — the two currencies are near-interchangeable.
• Hex: 10× training-time compute trades for 15× test-time compute.
• Libratus (heads-up no-limit poker) and Cicero (Diplomacy): similar trade-off ratios, which the article says amounts to up to a 100,000× effective increase in training compute.
• AlphaCode (competition coding): performance improves log-linearly with search time.
• OpenAI o1 (2024): reported gains scaling with both train-time and test-time compute across mathematics, scientific reasoning, and coding. Company-reported, relayed by the article, not independently verified.

Two ways to spend it

• Process-based supervision: reward intermediate steps of a reasoning chain, not just the final answer — from a human, a model, or synthetically, e.g. Monte Carlo rollouts that score each step by how often it leads to the right answer. The article notes this scales arbitrarily because synthetic reward needs no second model.
• Revision models: solve, then revise the previous attempt, repeatedly.

(Note the shape of both methods: they need a signal for what "right" means at step level. That is easy in Go, poker and code contests.)

Where the article is silent — which is the caution

The evidence for TTC scaling in this article is almost entirely from games and verifiable-answer tasks: domains where search can be graded against a ground truth. The article offers no equivalent evidence for open-ended generation, and no failure modes. Two further cautions, labelled as my inference:

1. The neat exchange rates ("2× for 2×") are per-domain fits; nothing in the article says they transfer to a new task or model family.
2. The same article's Broken neural scaling laws section is the counterweight to all extrapolation: many observed scaling relationships are smoothly broken power laws with breakpoints, not straight lines on a log-log plot. A trend measured in one regime does not entitle you to project it into another — including "more thinking always keeps paying."

For the hardware bill that TTC runs on, see KV caching and Speculative decoding, which cheapens per-token time rather than buying more tokens; for a measured look at local-model performance, Benchmarking local models.

Source: Wikipedia, "Neural scaling law", sections Inference scaling and Broken neural scaling laws, read 2026-09-08. o1 figures are OpenAI's own as the article relays them. Edited, not verified.