Session brief
The Unreasonable Effectiveness of Prompt Learning – Aparna Dhinakaran, Arize
Overview
This talk explores prompt learning as a method to improve coding agents, contrasting it with traditional Reinforcement Learning (RL). Prompt learning leverages English feedback on agent outputs to iteratively refine system prompts, offering a potentially more efficient and data-light approach for building agents compared to RL's reliance on scalar rewards and extensive data. The core idea is to use LLM-based evaluations to generate actionable feedback that directly informs prompt adjustments.
Who should watch
- AI engineers building or refining coding agents.
- Product managers evaluating agent performance and development strategies.
- Developers seeking to improve the effectiveness of tools like Claude and Cursor.
- Researchers interested in alternative methods to fine-tuning LLMs for agentic tasks.
Key takeaways
- System prompts are critical for the success of coding agents and require continuous iteration.
- Prompt learning uses English feedback from evaluations to update system prompts, mimicking human learning.
- This approach is presented as an alternative to Reinforcement Learning, which can be sample inefficient and data-hungry.
- The process involves generating code, running unit tests, using an LLM as a judge for evaluations, and then feeding these evaluations into a meta-prompt to refine the system prompt.
- Testing on SWEBench showed significant improvements in GitHub issue resolution rates for Claude and Cursor after prompt learning, without model fine-tuning.
- Developing high-quality LLM-as-a-judge evaluation prompts is crucial for generating effective feedback.
- The prompt learning approach was benchmarked against DSPI's GEA, showing comparable results with fewer iteration loops due to a focus on eval quality.
- *Iterating on system prompts is a key lever for improving agent performance.*
Notable quotes
Aparna Dhinakaran noted that *system prompts are repeatedly iterated on and are an important piece of context for making coding agents successful.*
She also stated that *evaluating and iterating on the eval prompts really mattered to making sure that you gave really good explanations back to the agent.*
Unofficial community note. Prefer the recording for nuance.