Think longer at answer-time

So far the whole story has been about training — spending compute ahead of time to shape the model.

• Pretraining taught it to predict the likely next word.
• Instruction tuning taught it to follow what you ask.
• Preference alignment taught it to lean toward answers people prefer.

All of that is baked in before you ever type a prompt. But there is one more lever, and it isn’t training at all.

Spend compute at answer-time

When the trained model finally answers, it runs forward through its weights and produces words. You can choose to spend more computation in that moment — let it generate more before it commits to a final answer. That class of methods is test-time compute: extra thinking at answer-time, no weights changed.

Why can intermediate text help? Later tokens can use earlier calculations as part of their input. Even a short answer is generated token by token; the comparison is between answering directly and generating additional reasoning before the answer.

Let it write the steps

Chain-of-thought prompting asks for intermediate reasoning before the answer. In Wei et al. (2022), worked examples in the prompt demonstrated that format. The model can then use earlier steps as context for later ones. This improved results in their experiments, but written reasoning can still contain mistakes.

Here are two scripted responses to the same math word problem:

Straight to the answer: “The answer is 27.” — wrong.

With steps: “The cafeteria had 23 apples. They used 20 for lunch, so 23 − 20 = 3. They bought 6 more, so 3 + 6 = 9. The answer is 9.” — right.

A scripted illustration of two answer formats; no model runs here. One example is wrong, and the worked example shows the calculation 23 − 20 + 6 = 9.

Try it

compare the two scripted responses, then check each arithmetic step yourself. The animation illustrates how intermediate results can be carried forward; it does not measure a model’s accuracy.

Generating additional intermediate tokens takes additional computation. This can help when those tokens contain useful calculations or checks. It can also extend a mistaken argument, so more text alone is not a guarantee of a better answer.

How intermediate text can help

In autoregressive generation, each new token is conditioned on the tokens already available. Intermediate text can act as scratch work: after writing 23 − 20 = 3, the model can use that result when adding 6. This explanation describes a possible benefit, not a promise that every intermediate prediction is easy or correct.

Today

This simple idea — give the model room to reason before answering — grew into today’s reasoning models, which are trained to write long internal chains of thought and can spend dramatically more answer-time compute on the hardest problems.

Training changes the model’s weights. Additional computation at answer time changes how much work the trained model does on this particular question. Both can affect accuracy, and both need evaluation.

Sources · 4

1. Jurafsky, D., & Martin, J. H. (2026). Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition with Language Models (3rd ed., draft of January 6, 2026). Stanford University. Ch. 9 (Post-training: Instruction Tuning, Alignment, and Test-Time Compute).
2. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems 35 (NeurIPS 2022). arXiv:2201.11903.
3. Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., & Schulman, J. (2021). Training Verifiers to Solve Math Word Problems. arXiv:2110.14168.
4. Suzgun, M., Scales, N., Schärli, N., Gehrmann, S., Tay, Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, E. H., Zhou, D., & Wei, J. (2023). Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them. Findings of ACL 2023. arXiv:2210.09261.

Full bibliography →