Reproduce “A Survey of Test-Time Compute: From Intuitive Inference to Deliberate Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yixin Ji, Juntao Li, Yang Xiang, Hai Ye, Kaixin Wu, Kai Yao, Jia Xu, Linjian Mo and 1 more
This survey provides the first comprehensive taxonomy of test-time compute methods, organizing them along a continuum from intuitive inference (System-1) to deliberate reasoning (System-2) for AI models. It consolidates fragmented research, offering a unified framework and identifying future research directions for more capable AI systems.
Key insights
- Test-time compute is a critical and evolving dimension in AI development, capable of transitioning models from basic System-1 capabilities to strong System-2 thinking, thereby unlocking further potential beyond training-time scaling.
- Kahneman's System-1 and System-2 thinking paradigm offers a powerful and coherent framework for classifying and understanding the diverse array of AI inference strategies.
- Many test-time compute strategies, particularly those for reasoning, can be integrated into model training (e.g., 'improvement training') to enhance models' inherent reasoning capabilities, demonstrating a feedback loop between inference and training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.