Cornell study: JSON mode cuts answer diversity across 44 language models

omarsar0 · x · 2026-07-23

Structured output can narrow model answers more than expected

A Cornell paper tests whether asking for structured output changes not just formatting, but the content distribution itself. Across 44 language models and 31 broad-answer prompts, the authors find that a simple instruction like “Reply with JSON only” measurably reduces diversity.

Key findings

On an open-ended pick-a-word prompt, asking for JSON shifts the modal answer from 41% to 64% and cuts distinct answers from 52 to 36.
The collapse appears for JSON and XML, both formats models are trained to emit, but not for YAML and CSV.
Forcing a schema at the decoder does not compress answers further, suggesting the narrowing happens when the model forms the response, not when the decoder enforces structure.
In a within-run re-sample, 53% of stable chat defaults shift under JSON mode, often back toward the majority answer.
The effect survives in 28 of 31 categories and shows a register-specific gradient: compression is present in answer-delivery formats, absent for YAML/CSV, and reversed for a generic bracket wrapper.

Why it matters

If your agent or pipeline relies on JSON mode, you may be sampling from a smaller and more homogeneous pool than the chat surface suggests.

Original post →

Cornell study: JSON mode cuts answer diversity across 44 language models

More from coding & agent