Eliciting Code Capability

Eliciting code capability focuses on unlocking the full potential of large language models (LLMs), particularly in code generation and related tasks, by overcoming limitations in training data and prompting strategies. Current research explores iterative prompting techniques using unlabeled data, fine-tuning with diverse multi-source datasets, and employing reinforcement learning methods to elicit even hidden or "password-locked" capabilities. These advancements are crucial for improving the safety and reliability of LLMs in various applications, including autonomous driving and software development, by enabling more robust and comprehensive evaluation of their abilities.