Continual System Prompt Learning for Code Agents – Aparna Dhinakaran, Arize
The talk by Aparna Dhinakaran introduces "system prompt learning" as an efficient alternative to traditional Reinforcement Learning for improving large language model-based coding agents. By leveraging LLM-as-a-judge evaluations to generate English feedback and explanations for code failures, agents can automatically refine their system prompts and rules. This method, demonstrated on Claude and Klein, significantly boosts performance on benchmarks like SWEBench with minimal data, highlighting the critical role of high-quality evaluation prompts.
System Prompt Learning for Coding Agents: Beyond Traditional RL
Aparna Dhinakaran’s presentation focuses on a paradigm shift in optimizing large language model (LLM) based coding agents, moving from traditional Reinforcement Learning (RL) to an innovative approach called System Prompt Learning (SPL). The core idea is to enable agents to learn and iteratively refine their own system prompts by incorporating human-like English feedback, much like a student improving from detailed critique rather than just a score.
The Overlooked Importance of System Prompts
The talk begins by highlighting the critical yet often underestimated role of system prompts in the success of sophisticated coding models. Citing a viral leak of Claude’s system prompt and observations from figures like Karpathy, it’s emphasized that these prompts are not static. Instead, they are