Skip to content

Structured Output Collapses Answer Diversity Across 44 Language Models

🕒 Published (v1): 2026-07-20 19:47 UTC · Source: Arxiv · link

Why this paper was selected

Structured-output format shifts answers across 44 LLMs; critical for every agent builder

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Requesting answers in JSON (or XML) measurably collapses answer diversity across 44 LLMs — field-mean answer-choice surprisal drops from 1.80 to 1.58 bits — without any constrained decoding, purely from the format instruction. The effect is progressive (distinctive models lose the most), register-specific (JSON/XML compress; YAML/CSV do not; brackets reverses the effect), and lives in the model's trained response to the register, not the sampler. The surface on which software consumes LLMs is measurably more homogeneous than the chat surface on which models are benchmarked.

Problem

LLMs are evaluated and ranked in prose (chat) but deployed through structured output (JSON, tool calls, pipelines). Whether the same model gives qualitatively different — and specifically less diverse — answers when asked for JSON has not been measured independently of constrained decoding or accuracy. Existing work on format taxes focuses on task accuracy; no prior work quantifies the effect on answer diversity when the format is merely requested, not enforced at the decoder.

Method

The authors re-run the One-Word Census instrument — 31 single-turn prompts ("Name a[n] X. Reply with one word only.") across 44 models, temperature 1.0, four samples per cell — with one manipulation: appending a serialization clause that requests the answer in a given format (JSON, XML, YAML, CSV, or square brackets). No schema enforcement, no token masking, no constrained decoding.

The metric is add-one-smoothed leave-one-out answer-choice surprisal in bits: how unlikely a model's answers are under the pooled answers of the other 43 models, computed by exact-match on normalized tokens. Surprisal is computed within each format column (JSON answers scored against the JSON pool), so the delta \(\Delta S = S_{\text{JSON}} - S_{\text{plain}}\) is a pure within-format measure, not a cross-register artifact.

To separate temperature from preference, the authors track self-distinctness (distinct answers / samples within a model-category cell). To separate real individual movements from noise, they apply a BH-FDR permutation test (\(q = 0.10\)) on both exchangeable units (31 categories and 44 models). A within-run \(n=20\) re-sample validates which stable chat defaults the register erodes or installs. A panel-free Jensen–Shannon divergence check distinguishes passive stranding (field collapses around a fixed model) from genuine divergence.

Key Contributions

  • Progressive format tax: Field-mean surprisal falls \(-0.22\) bits under JSON (\(p = .0002\)); distinctive models lose the most (deepseek-v3.2: \(-1.31\) bits, roughly half its measured distinctiveness) while the conformist floor is immobile.
  • Sharpener, not re-indexer: The JSON mode matches the plain-chat mode in 28 of 31 categories; only 3 flips occur, each in categories where the plain mode was weakest (<43% share). Less than 4% of JSON answer-mass lands on words the plain census never produced.
  • Register-indexed defaults: 53% of a model's stable chat defaults are significantly shifted by JSON (BH-FDR \(q=.10\)); 29% revert to the field modal answer. The register also installs register-only defaults (Claude Fable 5: cerulean for colour 0% → 100% in JSON, \(p \approx 10^{-11}\)).
  • Register gradient: Compression is significant for JSON (\(-0.22\) bits) and XML (\(-0.19\) bits), absent for YAML and CSV (\(p = .75\), \(.46\)), and reversed for a bracket wrapper (\(+0.13\) bits, \(p = .009\)) — pointing to tool-use post-training as the dominant mechanism.
  • Enforcement adds little: Enforcing JSON via response_format json_schema adds only \(-0.03\) bits beyond the request alone (\(-0.22\) bits); the collapse is in the register response, not the decoder.

Results

  • Field-mean surprisal: 1.80 bits (plain chat) → 1.58 bits (JSON), \(\Delta = -0.22\), \(p = .0002\).
  • On the unconstrained "Pick a word" prompt: modal word serendipity rises from 41% → 64% share; distinct answers fall from 52 → 36.
  • Individually significant compressors (BH-FDR \(q=.10\), all six negative): deepseek-v3.2 (\(-1.31\)), gpt-4o-mini (\(-0.92\)), hermes-4 (\(-0.91\)), gpt-4-turbo (\(-0.87\)), gpt-5.6-sol (\(-0.65\)), qwen3-235b (\(-0.45\)), gemini-3.1-pro (\(-0.31\)); every other model's delta is individually indistinguishable from zero.
  • Self-distinctness is nearly flat across formats (plain 0.42, JSON 0.39), ruling out a quiet temperature reduction as the mechanism.
  • Split-half test-retest reliability of census scores: \(r = 0.94\); regression-to-the-mean predicts \(\leq 0.05\) bits shrinkage for the most distinctive model vs. 1.31 bits observed (clearance \(> 3.7\sigma\)).
  • Register-enforcement comparison on 36-model subset: plain 1.79 bits → request 1.56 → enforcement 1.53 (\(\Delta_{\text{enforcement}} = -0.03\)).
  • n=20 re-sample: 81% of JSON-only defaults survive as genuine register-only behaviors.

Limitations

  • Single-channel snapshot (OpenRouter); temperature 1.0 is not uniformly honored across providers.
  • Surprisal is panel-relative — absolute bit values are not portable to other panels; only within-column and \(\Delta\) comparisons are internally valid.
  • Each format probed with a single clause wording; phrase-level confounds cannot be fully excluded (though the JSON/XML vs. YAML/CSV contrast already rules out key-priming and fill-in framing as sole explanations).
  • Tool-call framing and other structured-output pathways (beyond response_format) remain unmeasured.
  • response_format enforcement is served heterogeneously across providers; a provider-controlled native-enforcement replication is needed for per-model resolution.
  • Four samples per cell in the main battery; the \(n=20\) re-sample is limited to the default-validation sub-study.

Relevance to Agentic AI / LLM Agents

Agents almost universally consume LLM outputs in structured form (tool calls, JSON schemas, field extraction) rather than prose — this paper directly characterizes the distribution the agent is actually sampling from, which is measurably narrower than what benchmarks measure. The finding that the collapse is concentrated in the most distinctive models and is specific to tool-use-trained registers (JSON/XML but not YAML/CSV) implies that diversity-sensitive agentic tasks — brainstorming, multi-agent disagreement, tool selection among near-equivalent options — are operating on a compressed distribution the chat-surface evaluations do not capture. For multi-agent system designers, the result motivates explicitly testing diversity of agent responses in the deployment register, not the evaluation register, and suggests that format-switching could be a lever for either increasing or decreasing response diversity.