Why Semantic Entropy Fails: Geometry-Aware and Calibrated Uncertainty for Policy Optimization
Researchers have developed a new framework called Geometric-aware Calibrated Policy Optimization (GCPO) to improve post-training methods for large language models. Current approaches using semantic entropy for uncertainty signals are unstable and unclear in their impact on optimization. GCPO addresses this by integrating geometry-aware measures and reward-based calibration to better capture semantic disagreement and align uncertainty with learning signal strength. Experiments demonstrate that GCPO more accurately tracks gradient variability and consistently enhances post-training performance. AI
IMPACT This research offers a more principled approach to improving LLM reasoning and alignment through better uncertainty estimation in post-training.