A survey of test-time compute: From intuitive inference to deliberate reasoning

Analysis

Why This Paper Matters

Test-time compute scaling is an emerging paradigm that shifts computational effort from training to inference, enabling models to adapt dynamically. This survey is timely as the field moves beyond static, feed-forward models toward systems that can reason more deliberately at inference time. By tracing the concept back to System-1 models, the paper connects foundational ideas with modern scaling techniques, offering a unified view that can guide both research and practical deployment.

Technical Contributions

The paper's main contribution is a comprehensive taxonomy of test-time compute methods. Key innovations include:

• System-1 models as a starting point: Establishes that even early models used test-time compute to handle distribution shifts.
• Categorization of approaches: Distinguishes between intuitive inference (fast, automatic) and deliberate reasoning (slow, iterative) paradigms.
• Scaling strategies: Surveys techniques such as iterative refinement, chain-of-thought, and adaptive computation.
• Distribution handling: Focuses on how test-time compute addresses out-of-distribution inputs without retraining.

Results

As a survey, the paper does not present new experimental results. Instead, it synthesizes findings from numerous prior works, providing a structured overview of the field. The main outcome is a clear mapping of techniques and their relationships, which can serve as a reference for future research.

Significance

This survey has broad implications for AI deployment, especially in safety-critical or resource-constrained settings where test-time adaptation is crucial. By formalizing the landscape, it enables practitioners to choose appropriate scaling strategies and inspires new research directions in dynamic inference.