Test-Time Compute

Entity Type: Glossary ID: test-time-compute

Definition: A scaling paradigm that improves AI model performance by allocating additional computational resources during inference (test time) rather than just during training. This approach includes techniques like iterative refinement, multiple sampling and selection, chain-of-thought reasoning, tree search methods, and verification processes. Test-time compute allows models to 'think harder' on difficult problems by spending more time on reasoning and validation, often leading to better results on complex tasks.

Related Terms: - inference-scaling - reasoning-time - iterative-refinement - beam-search - compute-scaling

Source Urls: - https://arxiv.org/abs/2408.03314 - https://openai.com/index/learning-to-reason-with-llms/ - https://arxiv.org/abs/2407.21787

Tags: - inference - scaling - reasoning - compute-optimization

Status: active

Version: 1.0.0

Created At: 2025-09-10

Last Updated: 2025-09-10