Search papers, labs, and topics across Lattice.
This study investigates the phenomenon of "drift" in large language models (LLMs), where rephrasing benchmark problems can lead to significant changes in model performance. By generating meaning-preserving variations of benchmark problems across linguistic, referential, pragmatic, and structural axes, the authors reveal that stronger models are more sensitive to wording changes, often resulting in a greater loss of correctness compared to weaker models. The findings indicate that the fragility lies in the phrasing rather than the models themselves, highlighting a critical aspect of benchmark evaluation in AI.
Rephrasing benchmark problems can flip model answers, revealing that stronger LLMs are paradoxically more fragile to wording changes than weaker ones.
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
