Paraphrase-Induced Output-Mode Collapse: When LLMs Break Character Under Semantically Equivalent Inputs

TL;DR

This paper investigates how large language models often fail to maintain the original output format when prompts are paraphrased, revealing a systematic collapse in output mode across various models and tasks.

Contribution

The authors introduce PARACONSIST, a benchmark with 900 prompts and a Semantic Consistency Score to measure output-mode robustness in LLMs.

Findings

Only about 22% of responses preserve the original label under prompt variations.

Model task structure influences output-mode collapse more than model identity.

Response-mode preservation is crucial for reliable LLM deployment.

Abstract

When the substantive content of a request is rewritten, do large language models still answer in the format the original task asked for? We find that they often do not, even at temperature zero. On a 150-query evaluation over five compact 2025-era LLMs and four task types, we observe a systematic failure mode we call prompt-variant output-mode collapse: when a closed-form prompt asks for a bare label or a single choice token, content-preserving prompt variants can push the model into conversational prose, the requested format dissolves, and exact-match evaluation pipelines silently misjudge the result. To make this measurable, we release PARACONSIST, a 900-prompt benchmark of 150 base queries with five lexical, syntactic, and semantic-expansion prompt variants each, and a Semantic Consistency Score that decomposes prompt-variant robustness into answer consistency, sentence-BERT semantic similarity, and length stability. Under a whole-word answer-set match, only ~22% of closed-form variant responses preserve the ground-truth label inside their output, while ~78% drift away from the answer space entirely. In our pool, the dominant predictor of collapse is task structure rather than model identity, with model differentiation jointly carried by answer consistency and length stability. Robustness audits should therefore track response-mode preservation as a first-class reliability target alongside answer accuracy.