English

LLMs Show Surface-Form Brittleness Under Paraphrase Stress Tests

Computation and Language 2025-10-13 v1

Abstract

Benchmark scores for Large Language Models (LLMs) can be inflated by memorization of test items or near duplicates. We present a simple, protocol that probes generalization by re-evaluating models on paraphrased versions of benchmark questions. Using Mistral-7B-Instruct and Qwen2.5-7B-Instruct, we measure the accuracy gap between original and paraphrased items on ARC-Easy and ARC-Challenge. Our pipeline controls decoding, enforces multiple-choice output format, and includes a robust paraphrase-cleaning step to preserve semantics. We find that paraphrasing induces a non-trivial accuracy drop (original vs. paraphrased), consistent with prior concerns about contamination and brittle surface-form shortcuts.

Keywords

Cite

@article{arxiv.2510.08616,
  title  = {LLMs Show Surface-Form Brittleness Under Paraphrase Stress Tests},
  author = {Juan Miguel Navarro Carranza},
  journal= {arXiv preprint arXiv:2510.08616},
  year   = {2025}
}

Comments

NeurIPS 2025 Workshop. Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling. Selected for contributed talk

R2 v1 2026-07-01T06:27:43.757Z