Structured output may cut answer diversity across 44 language models
vista8 · x · 2026-07-27
A new paper claims that asking language models for structured output can reduce answer diversity across many models.
- The study re-runs the One-Word Census on 44 models, comparing unconstrained prompts with requests such as “Reply with JSON only.”
- Under JSON requests, the modal answer share rises from 41% to 64%, distinct answers drop from 52 to 36, and mean answer-choice surprisal falls from 1.80 to 1.58 bits.
- The authors argue the effect is progressive, largely tied to the format register rather than the decoder itself: JSON and XML show the strongest compression, while YAML and CSV do not show the same effect.
- They conclude that structured output makes models converge more toward the mode, meaning software systems that depend on schema-filled responses may see less diversity than chat evaluation suggests.
Related event: Structured Outputs Reduce LLM Diversity, Study Finds(2 posts)→
More from Research
- Minerva paper details: ncRNA arrays template short DNA hairpins — BrianHie · 2026-09-23
- Genome language models uncover new class of reverse-transcriptase mechanisms — BrianHie · 2026-09-23
- Mathematician shares a cheap 4-step heuristic for hyperparameter tuning — dejanseo · 2026-09-23
- Burkov skew AI hype: 'deterministic LLMs' and 'first agents' are old tricks rebranded — burkov · 2026-09-23
- Continuous diffusion beats discrete on random k-SAT, proposed as standard benchmark — ArashVahdat · 2026-09-23
- Grady Booch: Contemporary AI Still Lacks Abductive Reasoning, Just 'Next-Token Prediction' — Grady_Booch · 2026-09-23