Scaling with Confidence: Calibrating Confidence of LLMs for Adaptive Test Time Scaling

Researchers have introduced a method to improve large language model performance by calibrating confidence levels during the inference process, rather than relying solely on reinforcement learning rewards. This approach allows models to dynamically adjust the computational resources allocated to a task based on their internal certainty. By effectively scaling test-time compute, the technique aims to enhance reasoning accuracy and reduce errors in complex problem-solving scenarios without requiring further model retraining.