Skip to main content

Prompt Sensitivity Sweep

Same model, same questions, five wordings — five different scores.

ELI5 (Explain Like I'm 5)​

  • The Big Idea: We take the exact model and the exact MMLU questions from project 51 and change only one thing — how the question is phrased on the page. Five phrasings a person would call equivalent give five different scores.
  • The wordings: a bare Answer: cloze, a Q:/A: format, parenthesized (A) choices, a wordy "you are an expert taking an exam" instruction, and the model's own chat template. Nothing about the model's knowledge changed between them.
  • What we find: the scores swing 7.5 points (0.443 to 0.518) purely from wording. That is larger than the gap that separates many models on a leaderboard — so a benchmark comparison is only meaningful if both models were tested with the identical prompt.

Key Insight​

This project holds the model and the benchmark fixed and changes only how the question is phrased, running a sweep over five prompt formats and plotting how far the accuracy moves.

Why This Matters​

If rewording a prompt swings the score by ten points, then comparing two models is meaningless unless both are tested with the same template — a trap that has produced many bogus "model A beats model B" claims.

What's in this directory​

FileRole
prompt_sweep.pyScores Qwen2.5-0.5B-Instruct on a fixed 228-question MMLU sample under five prompt formats and plots the spread.

Reuses eval_lib.py and the mmlu_cloze(..., style=...) formats from project 51.

python prompt_sweep.py     # ~7 min on CPU

The five wordings​

Every format ends just before the answer letter, so all five are scored the same way (log-prob over the letter tokens) — the only variable is the surrounding text.

FormatShape
plainQuestion: … / A. … / Answer:
no_labelsbare question, then A. … / Answer:
parenthesized(A) … choices, then The answer is (
verbosea wordy "you are an expert taking a multiple-choice exam" wrapper
chatthe model's own Qwen chat template, answered on the assistant turn

Results​

Wording alone moves the score 7.5 points — and the wordy, "helpful" instruction does the worst.

prompt sweep

format          accuracy   95% CI
no_labels       0.518      [0.453, 0.582]
chat            0.496      [0.431, 0.560]
parenthesized   0.482      [0.418, 0.547]
plain           0.474      [0.410, 0.538]
verbose         0.443      [0.380, 0.508]

spread = 7.5 points  (min verbose 0.443 -> max no_labels 0.518)

Two things are worth sitting with:

  1. The verbose "expert exam" instruction is the worst format. The wrapper most people would assume helps a small model — spelling out the task, telling it to answer with a letter — costs it 7.5 points versus the terse no_labels format. Bigger models are more robust to this; small ones are easily distracted by extra tokens. You cannot guess the best format; you have to sweep it.

  2. The confidence intervals overlap. At 228 questions each format's 95% interval is about ±6 points, so we cannot confidently rank no_labels above plain — the point estimates differ by less than the noise. This is the sharper version of the lesson: if a 7-point gap can be half wording and half sampling luck, then a leaderboard that ranks two models 3 points apart, each run with its own prompt, is reporting noise with a decimal point.

The connection to project 51​

Project 51's reproduction of the published number was suspiciously exact (0.482 vs. 0.475). This sweep explains why that agreement was partly luck: the choice of prompt was worth several points on its own, comparable to the entire gap between our protocol and the authors'. A matching number means "nothing looks broken," not "confirmed" — and the only defensible way to compare models is to fix the prompt for all of them and report it.

Things to try​

  • Add a 6th format that inserts an irrelevant sentence before the question ("The following question is important.") and watch the score move again.
  • Re-run with per_subject raised to shrink the confidence intervals until the format ranking becomes statistically distinguishable — then note how many questions that took.
  • Sweep the same five formats on a larger model (if you have the compute); the spread shrinks, which is exactly why prompt robustness is itself a capability.