Insights from Benchmarking Frontier Language Models on Web App Code Generation

TL;DR

Insight from evaluating 16 frontier large language models on the WebApp1K benchmark reveals that writing correct code is more complex than generating incorrect code and prompt engineering shows limited efficacy in reducing errors beyond specific cases.

Abstract

This paper presents insights from evaluating 16 frontier large language models (LLMs) on the WebApp1K benchmark, a test suite designed to assess the ability of LLMs to generate web application code. The results reveal that while all models possess similar underlying knowledge, their performance is differentiated by the frequency of mistakes they make. By analyzing lines of code (LOC) and failure distributions, we find that writing correct code is more complex than generating incorrect code. Furthermore, prompt engineering shows limited efficacy in reducing errors beyond specific cases. These findings suggest that further advancements in coding LLM should emphasize on model reliability and mistake minimization.

Figures (15)

• Figure 1: Failures per Problem
• Figure 2: LOC Distribution by Model (BiModal)
• Figure 3: LOC Distribution by Model (UniModal)
• Figure 4: LOC Distribution by Model of High $pass@1$: Success vs Failure
• Figure 5: LOC Distribution by Model of Low $pass@1$: Success vs Failure