The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark Performance

2026-08-12

AI Summary

This work addresses the high sensitivity of large language models’ performance on benchmark tasks to question phrasing, noting that a single formulation inadequately captures their true capabilities. The authors propose an automated paraphrasing framework grounded in four linguistic dimensions—lexicon, reference, pragmatics, and syntax—to generate semantically equivalent yet lexically diverse question variants, enabling systematic evaluation of model response stability. They identify and quantify a previously uncharacterized phenomenon termed “bidirectional drift”: stronger models exhibit greater sensitivity to rephrasing, with accuracy significantly declining under certain paraphrases. Crucially, this fragility stems from the rephrasings themselves rather than inherent model differences. Experiments across eight prominent models and three established benchmarks consistently demonstrate the prevalence and robustness of this effect.

Abstract

A benchmark score comes from a single phrasing of each problem. That single phrasing is treated as if it stood for the whole space of ways the same problem could be asked, but it does not. We show that rephrasing a problem while keeping its meaning and answer fixed routinely flips a model's answer in both directions, so some failures become successes and some successes become failures. We call this drift. BenchDrift generates meaning-preserving variations of benchmark problems along four axes, namely linguistic, referential, pragmatic, and structural, and measures how often, and why, correctness flips under each. Across eight models and three benchmarks (GSM8K, MMLU, MATH-Hard), we observe that drift is large in both directions. Two findings stand out. First, phrasing sensitivity does not fade as models get better. Instead, it changes sign. Weak models gain more from rephrasing than they lose, while strong models lose far more than they gain. We find that the best models on a benchmark are therefore the ones whose scores depend most on the wording they happened to be given. Second, the models largely agree on which rephrasings cost the most correct answers even though they differ in how much they drift, so fragility belongs to the rephrasing and not to the model. Furthermore, rephrasing breaks answers a model was confident about, whether the problem is made shorter or longer. Code and Data: https://github.com/IBM/BenchDrift/tree/demo-ui

Problem

• wording effect
• benchmark drift
• large language models
• rephrasing sensitivity
• evaluation reliability

Innovation

• Wording Effect
• Benchmark Drift
• Meaning-Preserving Rephrasing
• LLM Robustness
• BenchDrift