Reproduce “Structured Output Collapses Answer Diversity Across 44 Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tapan Parikh
Researchers at Cornell Tech quantified how requesting structured output like JSON from language models significantly reduces answer diversity, particularly for initially distinctive models, revealing a 'format tax' that amplifies response conformity. The study found that models become more homogeneous in structured formats compared to free-form prose, indicating that evaluations based on chat interfaces may overstate actual diversity in deployment contexts.
Key insights
- A 'progressive format tax' exists, where requesting structured output (especially JSON) significantly increases conformity across LLMs, with models that were initially more distinctive experiencing the largest drops in diversity.
- The structured output format acts as a 'sharpener' that reinforces existing modal answers rather than a 're-indexer' that causes models to select entirely new answers.
- LLM 'defaults' (idiosyncratic, stable answer choices) are 'register-indexed', meaning the requested output format can significantly erode or install specific preferences.
- A 'register gradient' shows that compression effects are most pronounced for specific serialization formats (JSON, XML) typically associated with tool-use post-training, distinguishing them from generic structural requests (e.g., brackets).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.