arXiv cs.AI· Vineet Kumar, Kanishka, Bhuvanesh Mandora·· 9 天前AI 评分29
Structure Tax：结构化输出如何影响 LLM 性能
Structure Tax: How Structured Output affects LLMs Performance
AI 导读
研究重新审视了 LLM 结构化输出导致精度损失的"结构税"说法，发现损失取决于 schema 设计而非结构本身。推理优先的字段排序可匹配甚至超过自由格式精度，答案优先排序则导致精度骤降，小模型尤为明显。保留推理顺序的 schema 还能改善置信度校准，CKA 显示中间 Transformer 层正确与错误表征的可分离性更强。
正文 · 原文
Abstract:Deploying large language models in production often requires constraining outputs to structured formats such as JSON or XML, and prior work treats the resulting accuracy loss as an inherent `structure tax'. We re-examine this claim by evaluating a battery of models, datasets and schemas, measuring task accuracy, confidence calibration, and hidden-state geometry. The tax turns out to depend on schema design rather than on structure per se: reasoning-first field ordering matches or exceeds free-form accuracy, while answer-first ordering causes steep drops, particularly in smaller models. Format sensitivity scales inversely with a task's own structural constraints, and schemas that preserve reasoning order also improve calibration with CKA showing greater separability between correct and incorrect representations in middle transformer layers. Our findings indicate that properly designed structured formats can match or exceed free-form performance, reframing the critical question from `whether to structure' to `how to structure' for optimal reasoning preservation.
| Comments: | AACL IJCNLP 2026 (Main) |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.12056 [cs.AI] |
| (or arXiv:2610.12056v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.12056 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Vineet Kumar [view email]
[v1]
Thu, 8 Oct 2026 14:39:31 UTC (1,391 KB)
来源：arXiv cs.AI · arxiv.org