Paper of the Week — Structured Output Collapses Answer Diversity Across 44 Language Models

Structured Output Collapses Answer Diversity Across 44 Language Models

Tapan Parikh. Published 2026-07-20. arXiv:2607.18476

One sentence summary

Adding a JSON-only format clause to a prompt changes which answer a model selects from a wide-open option space — not just how it wraps the response — across 44 models tested.

Why this paper

Every production LLM system uses structured output. If format constraints are silently biasing which answer gets chosen, that’s a correctness bug hiding behind a formatting feature — and it’s flying under the radar.

What they did

The author re-ran the One-Word Census benchmark (31 wide-answer-space category prompts, like “name a country”) across 44 models, once with free-form prompting and once with a “Reply with JSON only” clause appended. The benchmark was designed to surface answer diversity across a broad option space, making it a clean instrument for detecting format-induced answer shift rather than just output-structure change.

Key findings

• Structured output prompts measurably collapsed answer diversity — models converged on fewer, more predictable answers when asked to respond in JSON
• The effect was consistent across all 44 models tested, from small open-weights to frontier-class models, suggesting this is architectural rather than a quirk of one model family
• The shift is in the semantic content of answers, not just token arrangement — models pick different words from the option space under format constraints
• Homogenization was strongest on prompts with the widest legitimate answer spaces (geography, foods, animals), exactly where you’d want diversity preserved
• The finding replicates and extends prior One-Word Census work by isolating format as the sole variable

Why it matters for practitioners

If you’re using structured output for tool-calling, classification, or any task where the model should draw from a broad option space, your format instructions may be systematically biasing which options get selected — you’re not just changing the wrapper, you’re changing the content. This is especially sharp for use cases like recommendation diversity, creative generation under constraints, or any eval pipeline where you force JSON to make parsing easier.

What you can use today

• Run a quick A/B on your own prompts: strip the format clause and compare answer distributions over 20–50 samples to see if you have silent homogenization in production
• If you need structured output and diversity, consider generating free-form first and parsing afterward with a lightweight extraction step, rather than constraining generation directly
• When benchmarking models for tasks requiring broad coverage (e.g., recommendation or brainstorming agents), make sure your eval prompts don’t include format constraints that would artifically flatten the score distribution you’re measuring