SWE-bench

The Shift from Static Prompts to Dynamic Learning Software development is hitting a wall with Large Language Model (LLM) agents. We have built systems that work 80% of the time, but the remaining 20%—the "reliability gap"—remains stubbornly open. Traditionally, we have tried to close this gap by manually tweaking prompts, a process that is both unscalable and fragile. SallyAnn DeLucia and Fuad Ali from Arize AI argue that the industry needs to move away from static instructions entirely. Instead, developers should implement **prompt learning**, a technique that borrows principles from Reinforcement Learning to create a self-correcting optimization loop. Unlike traditional prompt engineering, where a human tries to guess what words might steer the model better, prompt learning treats the prompt as a set of weights that can be updated based on structured feedback. The core philosophy is that the most valuable data in your system isn't just the final output; it is the **English feedback** explaining *why* an output failed. By capturing human or LLM-based explanations of failures and feeding them back into an optimizer, you can achieve performance gains—like a 15% improvement in coding accuracy—without touching the underlying model architecture or training data. Prerequisites and the Optimization Stack To build a prompt learning loop, you need a baseline understanding of Python and Jupyter Notebooks. Conceptually, you should be familiar with evaluation frameworks and the idea of "LLM-as-a-judge." Key Libraries & Tools * **Arize Phoenix**: An open-source observability library used for tracing and evaluating LLM applications. * **OpenAI SDK**: Used here for both the core agent logic and the evaluators (specifically GPT-4o or newer models). * **Nest-asyncio**: A utility to allow nested asynchronous loops in Jupyter, which is critical for running parallel evaluations quickly. * **Pandas**: Necessary for managing the training and testing datasets that drive the optimization process. Architecting the Multi-Step Optimization Loop Setting up the environment requires specific attention to library versions. A common pitfall in these rapidly evolving ecosystems is version mismatch. For this tutorial, ensure you are using `arize-phoenix >= 2.2.0` to avoid package conflicts during evaluation. ```python import phoenix as px import nest_asyncio Patch for Jupyter environments to handle async calls nest_asyncio.apply() Configuration parameters NUM_SAMPLES = 50 TRAIN_SPLIT = 0.8 OPTIMIZATION_LOOPS = 5 ``` The loop consists of three logical stages: **Generation**, **Evaluation**, and **Refinement**. You start by splitting your dataset into a training set (used to generate the new prompt) and a test set (used to verify that the new prompt actually performs better). Building Custom Evaluators as High-Fidelity Signals A prompt learning loop is only as strong as its evaluators. If your evaluator provides a simple "Incorrect" label without context, the optimizer has no idea how to fix the instruction. You must initialize evaluators that provide **detailed explanations**. ```python Initializing the Classification Evaluator evaluate_output = px.evals.OpenAIModel( model="gpt-4o", template=EVAL_TEMPLATE, # A template defined in external files choices=["correct", "incorrect"] ) ``` In this workshop, SallyAnn DeLucia highlights the "Rule Checker"—a specialized evaluator that performs a granular, rule-by-rule analysis of the output. This creates a high-dimensional feedback signal. Instead of telling the optimizer "this failed," it says "this failed because it didn't adhere to the JSON schema in rule #3." This level of specificity is what allows the Prompt Learning SDK to rewrite the system prompt effectively. Syntax Notes and Implementation Details When writing the optimization logic, pay attention to the **response format** and **temperature**. For consistent results during an automated optimization loop, setting `temperature=0` is standard practice. ```python async def generate_output(data, system_prompt): response = await client.chat.completions.create( model="gpt-4o", messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": data} ], response_format={"type": "json_object"}, temperature=0 ) return response.choices[0].message.content ``` The `response_format` parameter is a critical language feature in the OpenAI API that ensures the model outputs valid JSON. This is vital when the task involves web page creation or structured data, as it prevents the optimizer from getting distracted by formatting errors and allows it to focus on logic and content. Practical Case Study: The 15% Performance Jump To prove the efficacy of this method, Arize AI applied prompt learning to OpenDevin, an open-source coding agent. The original system prompt was remarkably simple, lacking specific rules for error handling or test requirements. By running this exact optimization loop, the system generated a new prompt that included a robust "Rules" section. This optimized prompt improved the agent's performance on the SWE-bench benchmark by 15%. Most importantly, the optimized agent (using GPT-4o) approached the performance of much more expensive models like Claude 3.5 Sonnet while costing two-thirds less. This demonstrates that "expertise" can be engineered into a prompt through data-driven iterations, often negating the need for expensive fine-tuning. Tips and Debugging Your Loop 1. **Avoid Over-Optimization**: There is a temptation to run 20 or 30 loops. However, Fuad Ali notes that significant gains usually occur within the first 3-5 loops. Beyond that, you risk overfitting to the specific quirks of your training set. 2. **Optimize the Evaluator First**: If your prompt learning loop isn't working, the problem is likely your evaluator. You should optimize the evaluator's prompt with the same rigor as your agent's prompt. 3. **Use Logprobs for Confidence**: If you aren't sure if the model's "Incorrect" label is reliable, look at the logprobs (logarithmic probabilities) of the token. Low confidence in the evaluator's label should trigger a human review. 4. **Handling Multi-Agent Systems**: While the current SDK focuses on independent tasks, you can optimize multi-agent systems by treating each agent's hand-off as a discrete step for prompt learning. By treating prompts as software that requires a CI/CD-like iteration cycle, developers can finally build agents that aren't just "cool prototypes" but reliable production tools.