Calibrating Language Models with Adaptive Temperature Scaling
This work is significant for researchers and practitioners working with RLHF-tuned LLMs, as it provides a method to restore crucial calibration properties that are often lost during fine-tuning.
This paper addresses the degradation of calibration in large language models (LLMs) after fine-tuning with reinforcement learning from human feedback (RLHF). The authors introduce Adaptive Temperature Scaling (ATS), a post-hoc method that predicts a token-specific temperature scaling parameter, improving calibration by 10-50% across three NLP benchmarks without hindering RLHF performance gains.
The effectiveness of large language models (LLMs) is not only measured by their ability to generate accurate outputs but also by their calibration-how well their confidence scores reflect the probability of their outputs being correct. While unsupervised pre-training has been shown to yield LLMs with well-calibrated conditional probabilities, recent studies have shown that after fine-tuning with reinforcement learning from human feedback (RLHF), the calibration of these models degrades significantly. In this work, we introduce Adaptive Temperature Scaling (ATS), a post-hoc calibration method that predicts a temperature scaling parameter for each token prediction. The predicted temperature values adapt based on token-level features and are fit over a standard supervised fine-tuning (SFT) dataset. The adaptive nature of ATS addresses the varying degrees of calibration shift that can occur after RLHF fine-tuning. ATS improves calibration by over 10-50% across three downstream natural language evaluation benchmarks compared to prior calibration methods and does not impede performance improvements from RLHF.