Test-time compute
Çıkarım zamanı hesaplaması
D2 · Generative AI, capabilities and limitations
Test-time compute is computation spent producing an answer at inference time; scaling it allocates extra work to a request.
Review status: 2026-12-05
Technical explanation
Extra work can support searching candidate answers or revising a response before returning it.
Conceptual boundaries
It concerns the inference budget rather than simply increasing the number of model parameters.
Provider-neutral example
Illustrative: an application compares several proposed solutions within a fixed compute budget.
Limitations
Benefits depend on the task and allocation strategy; extra computation is not a correctness guarantee.
Related concepts
Atomic claims and evidence
1.1Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- Source
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- Source role
- Authoritative source
- Exact locator
- Abstract: inference-time compute, search, adaptive responses and prompt difficulty
- Supported claim
- Snell and colleagues study inference-time search and adaptive response methods using additional computation.
- Last verification
- Review due
- Scope limitation
- The entry summarizes inference-time allocation, not every form of test-time adaptation.
2.1Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- Source
- Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
- Source role
- Authoritative source
- Exact locator
- Abstract: inference-time compute, search, adaptive responses and prompt difficulty
- Supported claim
- The study finds that effective allocation of test-time compute depends on prompt difficulty.
- Last verification
- Review due
- Scope limitation
- Comparative experimental results do not establish universal gains.