prompt-sensitivity-reorder-40pct-format-76pts

IN premise — entries/2026/06/21/wiki-Prompt_engineering-chunk-1.md

Created 2026-06-21T09:50:10+00:00

LLMs are highly sensitive to prompt formatting: reordering few-shot examples can shift accuracy by 40%+, and formatting changes can cause up to 76 accuracy points of variation

Summary

The way a prompt is structured matters far more than the raw content itself — simply rearranging example sentences or tweaking formatting can swing a model's accuracy by dozens of percentage points. This means LLM outputs are fragile and hard to reproduce, so any system built on them needs to treat prompt layout as a first-class engineering concern rather than a cosmetic detail.

Dependents

These beliefs depend on this one:

IN prompt-sensitivity-is-fundamental-not-solvable-by-scale — LLM performance sensitivity to prompt formatting (40%+ accuracy shifts from reordering, 76-point swings from format changes) persists across model sizes, few-shot counts, and instruction tuning — it is an intrinsic property, not a fixable deficiency.