Simon Willison’s Weblog

Subscribe

LLMs are bad at returning code in JSON (via) Paul Gauthier's Aider is a terminal-based coding assistant which works against multiple different models. As part of developing the project Paul runs extensive benchmarks, and his latest shows an interesting result: LLMs are slightly less reliable at producing working code if you request that code be returned as part of a JSON response.

Coding skill by model and code wrapping strategy - four models, each showing their pass rate % average of five runs. Claude 3.5 Sonnet gets 60.5% with Markdown, 54.1% with JSON. DeepSeek-Coder V2 0724 gets 60.6% with Markdown, 51.1% with JSON. GPT-4o-2024-05-13 gets 60.0% with Markdown, 59.6% with JSON. GPT-4o-2024-08-06 gets 60.8% with Markdown, 57.6% with JSON, and 56.9% with JSON (strict). Markdown consistently performs better than JSON across all models.

The May release of GPT-4o is the closest to a perfect score - the August appears to have regressed slightly, and the new structured output mode doesn't help and could even make things worse (though that difference may not be statistically significant).

Paul recommends using Markdown delimiters here instead, which are less likely to introduce confusing nested quoting issues.

Posted 16th August 2024 at 5:04 pm

Recent articles