Generative AI · Fast-moving · Advanced

Test-Time Compute

Also known as: Inference-time scaling

Improving answer quality by spending more computation at inference — thinking longer, sampling more, or verifying — rather than by training a bigger model.

What Test-Time Compute is

This is a second scaling axis. A model can be given a budget to deliberate, generate candidates, critique and revise before answering.

How it works

Techniques include extended internal reasoning, best-of-n sampling with a verifier, self-critique loops and search over candidate solutions. Providers increasingly expose an explicit effort or thinking-budget setting.

Why it matters

It underpins the current generation of reasoning models and changes cost planning: the expensive requests are the hard ones, not simply the long ones.

Common uses

→Competition-level maths and coding
→Complex agent planning
→High-stakes analysis

Strengths

✓Better accuracy without retraining
✓Budget can be tuned per request

Watch for

✓Latency and cost grow quickly
✓Returns diminish on simple tasks

Continue exploring

Reasoning Model Chain-of-Thought Prompting Self-Consistency Scaling Laws