Reproduce “EpiCaR: Knowing What You Don't Know Matters for Better Reasoning in LLMs” on OpenResearch

About this paper

Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park, Taesup Kim

The EPICAR framework, developed by researchers at Seoul National University, introduces an epistemic learning objective for large language models to jointly optimize reasoning performance and uncertainty calibration. It achieves Pareto-superiority in both metrics and enables a 3x reduction in inference compute by effectively integrating negative self-evaluation signals during supervised fine-tuning.

Key insights

• Reframing reasoning training as an "epistemic learning problem" allows models to learn not only how to reason but also when their reasoning should be trusted.
• Integrating negative self-evaluation signals during training is crucial for mitigating model collapse and the "calibration cost" often associated with positive-only self-training methods.
• Uncertainty calibration, when intrinsically learned, serves as a fundamental prerequisite for developing compute-optimal reasoning systems, rather than just an external adjustment.

Summary by alphaXiv.