Sclar et al. (2023), "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design," measure how much purely cosmetic formatting choices (such as separators and spacing) in a few-shot prompt affect accuracy, holding the semantic content constant. What did they find?
- Purely formatting-level changes to a semantically identical few-shot prompt caused accuracy swings of up to tens of accuracy points (as much as 76 points for one open-source model tested), and this sensitivity persisted even with larger models, more few-shot examples, and instruction tuning
- Formatting choices had no measurable effect on accuracy once a model exceeded roughly one billion parameters, fully resolving the issue at modern model scales
- Formatting only matters for image-based prompts and has no measurable effect on plain-text few-shot prompts
- Accuracy differences from formatting disappeared entirely once few-shot examples were replaced with zero-shot instructions
Correct answer: A. Purely formatting-level changes to a semantically identical few-shot prompt caused accuracy swings of up to tens of accuracy points (as much as 76 points for one open-source model tested), and this sensitivity persisted even with larger models, more few-shot examples, and instruction tuning
Sclar et al. found that large language models are highly sensitive to spurious, purely cosmetic formatting choices in few-shot prompts, such as separators, spacing, and other superficial template details that do not change the semantic content, with accuracy swings of up to 76 points observed for one model studied (LLaMA-2-13B), and this sensitivity did not disappear as model size, number of few-shot examples, or instruction tuning increased. The second option is wrong because the paper explicitly reports that this sensitivity persists even as models scale up, rather than resolving at some parameter threshold. The third option is wrong because the study concerns plain-text prompt formatting rather than image inputs, and it found the effect specifically in text-based few-shot prompting. The fourth option is wrong because the paper studies few-shot formatting sensitivity directly; it does not report that switching to zero-shot prompting eliminates formatting-driven accuracy differences.
Source: Sclar, Choi, Tsvetkov, Suhr, "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design" (ICLR 2024), arXiv:2310.11324