test-time compute
Test-time compute refers to the computational resources and time required to evaluate a trained machine learning model on new, unseen data, which can impact the model's practicality in real-world applications, especially for those requiring quick inference.
- Learning to Better Search with Language Models via Guided Reinforced Self-Training
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation
- Provable Scaling Laws for the Test-Time Compute of Large Language Models
- Reasoning as an Adaptive Defense for Safety
- Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning
- Reward Reasoning Models
- Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning