LLMs Show Surface-Form Brittleness Under Paraphrase Stress Tests

TL;DR

The paper investigates surface-form brittleness of LLMs under paraphrase stress tests to address concerns that benchmark scores reflect memorization or brittle phrasing. It introduces a protocol that paraphrases MCQ stems, a cleaning pipeline to preserve semantics, and a fully specified ARC-based setup using Mistral-7B-Instruct and Qwen2.5-7B-Instruct with 4-bit inference on an A100, revealing a measurable accuracy drop when questions are paraphrased. Across ARC-Easy and ARC-Challenge, paraphrasing yields consistent performance degradation (approximately 6–10 percentage points), suggesting residual memorization and reliance on surface patterns; results hold across cross-pairings of answerer and paraphraser. The findings motivate paraphrase-aware evaluation, contamination auditing, and broader-domain stress testing to better quantify true generalization in LLMs beyond surface-level cues.

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.