RECKONING Method: Bi-Level Optimization for Dynamic Knowledge Encoding and Robust Reasoning

This article introduces RECKONING, a novel method utilizing bi-level optimization to teach language models to reason by dynamically encoding contextual knowledge.