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

• URL: http://arxiv.org/abs/2605.04665v1
• Date: Wed, 06 May 2026 09:11:10 GMT
• Title: Paraphrase-Induced Output-Mode Collapse: When LLMs Break Character Under Semantically Equivalent Inputs
• Abstract summary: We observe a systematic failure mode we call prompt-variant output-mode collapse.<n>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.<n>We release PARACONSIST, a benchmark of 150 base queries with five lexical, syntactic, and semantic-expansion prompt variants each.
• Score: 0.9525172018746524
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• 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.

Related papers

• What Does Your Short-Answer VQA Score Actually Measure? Evaluator-Dependent Instability in Multimodal Short-Answer Benchmarks [25.519850997497283]
  Short-answer VQA benchmarks conflate two distinct quantities.<n>We use a human-validated semantic judge to audit over 37k official errors.
  arXiv  Detail & Related papers  (2026-07-11T10:01:29Z)
• When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment [6.48455847613423]
  Language models often generate reasoning before giving a final answer, but the visible answer does not reveal when the model's answer preference became stable.<n>We study this question through a narrow computable object: emphfinite-answer preference stabilization.<n>For a model state and specified answer verbalizers, we project the model's own continuation probabilities onto a finite answer set.<n>In binary tasks this yields an exact log-odds code, $()=S_(mathrmyesmid)-S_(mathrmnomid
  arXiv  Detail & Related papers  (2026-05-07T08:34:59Z)
• Instruction Complexity Induces Positional Collapse in Adversarial LLM Evaluation [0.0]
  Vague adversarial instructions produce moderate accuracy reduction with preserved content engagement.<n>A two-step answer-aware avoidance instruction produces extreme positional collapse with near-total concentration on a single response position.<n>Results suggest that instruction complexity can determine whether adversarial compliance uses content-aware or content-blind mechanisms.
  arXiv  Detail & Related papers  (2026-04-29T22:48:24Z)
• Prompt-Based One-Shot Exact Length-Controlled Generation with LLMs [56.47577824219207]
  We present a prompt-based strategy that compels an off-the-shelf large language model to generate exactly a desired number of tokens.<n>The prompt appends countdown markers and explicit counting rules so that the model "writes while counting"<n>On MT-Bench-LI, strict length compliance with GPT-4.1 leaps from below 30% under naive prompts to above 95% with our countdown prompt.
  arXiv  Detail & Related papers  (2025-08-19T13:12:01Z)
• Prompt-Response Semantic Divergence Metrics for Faithfulness Hallucination and Misalignment Detection in Large Language Models [0.0]
  This paper introduces Semantic Divergence Metrics (SDM), a novel framework for detecting Faithfulness Hallucinations.<n>A heatmap of topic co-occurances between prompts and responses can be viewed as a quantified two-dimensional visualization of the user-machine dialogue.
  arXiv  Detail & Related papers  (2025-08-13T20:55:26Z)
• Variability Need Not Imply Error: The Case of Adequate but Semantically Distinct Responses [7.581259361859477]
  Uncertainty quantification tools can be used to reject a response when the model is uncertain'<n>We estimate the Probability the model assigns to Adequate Responses (PROBAR)<n>We find PROBAR to outperform semantic entropy across prompts with varying degrees of ambiguity/open-endedness.
  arXiv  Detail & Related papers  (2024-12-20T09:02:26Z)
• Prompt Stability Scoring for Text Annotation with Large Language Models [0.0]
  Researchers are increasingly using language models (LMs) for text annotation.<n>These approaches rely only on a prompt telling the model to return a given output according to a set of instructions.<n>The diagnose of LM outputs may nonetheless be vulnerable to small changes in the prompt design.
  arXiv  Detail & Related papers  (2024-07-02T08:11:18Z)
• QLSC: A Query Latent Semantic Calibrator for Robust Extractive Question Answering [32.436530949623155]
  We propose a unique scaling strategy to capture latent semantic center features of queries. These features are seamlessly integrated into traditional query and passage embeddings. Our approach diminishes sensitivity to variations in text format and boosts the model's capability in pinpointing accurate answers.
  arXiv  Detail & Related papers  (2024-04-30T07:34:42Z)
• Look Before You Leap: A Universal Emergent Decomposition of Retrieval Tasks in Language Models [58.57279229066477]
  We study how language models (LMs) solve retrieval tasks in diverse situations. We introduce ORION, a collection of structured retrieval tasks spanning six domains. We find that LMs internally decompose retrieval tasks in a modular way.
  arXiv  Detail & Related papers  (2023-12-13T18:36:43Z)
• Answering Ambiguous Questions via Iterative Prompting [84.3426020642704]
  In open-domain question answering, due to the ambiguity of questions, multiple plausible answers may exist. One approach is to directly predict all valid answers, but this can struggle with balancing relevance and diversity. We present AmbigPrompt to address the imperfections of existing approaches to answering ambiguous questions.
  arXiv  Detail & Related papers  (2023-07-08T04:32:17Z)
• Compressive Summarization with Plausibility and Salience Modeling [54.37665950633147]
  We propose to relax the rigid syntactic constraints on candidate spans and instead leave compression decisions to two data-driven criteria: plausibility and salience. Our method achieves strong in-domain results on benchmark summarization datasets, and human evaluation shows that the plausibility model generally selects for grammatical and factual deletions.
  arXiv  Detail & Related papers  (2020-10-15T17:07:10Z)
• Generating Dialogue Responses from a Semantic Latent Space [75.18449428414736]
  We propose an alternative to the end-to-end classification on vocabulary. We learn the pair relationship between the prompts and responses as a regression task on a latent space. Human evaluation showed that learning the task on a continuous space can generate responses that are both relevant and informative.
  arXiv  Detail & Related papers  (2020-10-04T19:06:16Z)
• Robust Question Answering Through Sub-part Alignment [53.94003466761305]
  We model question answering as an alignment problem. We train our model on SQuAD v1.1 and test it on several adversarial and out-of-domain datasets.
  arXiv  Detail & Related papers  (2020-04-30T09:10:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.