Wording changes moved benchmark scores by dozens of points

Start here

A prompt that scores well once, on one model, is not the same as a prompt that is reliable. Two studies measured exactly how much a meaning-preserving change, the same instruction reworded or reformatted, can move a model's accuracy, and the answer in both cases is more than most people testing a single prompt would expect.

What the documents say

Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design, submitted 17 October 2023 by Melanie Sclar and colleagues, tested several widely used open-source models on few-shot tasks and reports that formatting changes alone, not the content of the instruction, produced performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. The paper states the sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning, and that which format works best only weakly correlates between models, meaning a prompt tuned on one model is a weak guide to another. A follow-up study, The Unreasonable Effectiveness of Eccentric Automatic Prompts, submitted 9 February 2024 by Rick Battle and Teja Gollapudi, tested 60 system-message variants, with and without chain-of-thought, across three models from 7 to 70 billion parameters on the GSM8K math benchmark. It found that results do not universally generalize across models: adding a 'positive thinking' phrase helped most models, but the 70-billion-parameter Llama2 model performed best with no system message at all when not using chain-of-thought. An automated prompt search beat the best hand-written variant, including on smaller open models, and the winning automated prompt was, in the paper's own word, unexpectedly eccentric.

Check this

The mechanism is testable directly: take one working prompt, rewrite it with a different but meaning-equivalent format, a different delimiter, spacing, or label style, and run both versions on the same task twice each. If the two versions disagree noticeably, the first score was measuring the format as much as the task, which is what both studies found at a larger scale.

What holds and what fails

The evidence holds as a caution against trusting one good result from one prompt on one model as proof of a reliable method. It does not establish that formatting is the largest lever available, only that it is a larger one than a single successful attempt would reveal, since neither paper compares its effect size against every other prompting choice. Carrying a prompt tuned on one model to a different model without retesting is the specific failure both papers document, not a hypothetical one.

• Run any important prompt at least twice, with a reformatted but equivalent version, before trusting the first result.
• Do not assume a prompt tuned on one model performs the same way on a different model or a new version of the same one.
• Treat a single successful test as a sample of one, not as a validated method.

Both studies are about brittleness that is invisible in a single test and only shows up once someone deliberately reruns the same request in a different shape, a check most everyday prompting skips.