stanjsx/did-the-perturbation-do-that

Paired Prompt-Robustness Outcomes

This repository accompanies the paper Did the Perturbation Do That? Hidden Failures and Misattributed Causes in Prompt-Robustness Evaluation. It contains:

1. data/outcomes.csv: whether each of six models answered each item correctly under each of ten conditions, on six multiple-choice and three coding benchmarks (199,980 outcomes).
2. refcal/: a small Python tool that adds the same two content-preserving references to any benchmark and reports the paper's measures from per-item outcomes.

Every number in the paper's main experiments can be recomputed from data/outcomes.csv with refcal, without calling any model.

Why paired outcomes

Prompt-robustness benchmarks usually report how much accuracy drops under a perturbation. That net change hides answers that break when other answers recover. It also credits the perturbation with failures that occur even when the prompt is only resent or reformatted (background brittleness). Both problems are visible only when the same item is followed across conditions, which is what this table records.

For example, GPT-5 on mmlu::mmlu_professional_law::303:

baseline	repeat 1–3	fmt quotes / spaces / newlines	typos	spacing	stopword deletion
1	0, 0, 1	1, 0, 1	0	0	1

The answer that fails under typos already fails in two of three repeats of the unchanged prompt.

Conditions

Each model–item pair is evaluated under ten conditions:

condition	family	What changes
baseline	baseline	nothing: the benchmark prompt
repeat_1, repeat_2, repeat_3	repeat	nothing: the same prompt, sent again as a separate call
fmt_quotes	format	the whole prompt is wrapped in double quotation marks
fmt_spaces	format	eight spaces are added before and after the prompt
fmt_newlines	format	three newlines are added before and after the prompt
typos_k3	perturbation	three successive swaps of two adjacent letters in eligible words
sequence_spaces	perturbation	one space is replaced by four spaces
drop_stopwords	perturbation	every eligible stopword is deleted

Perturbations edit only the question (for coding tasks, only the natural-language instruction). Answer options, answer labels, negations, and tokens with math or code characters are never edited. refcal make-references regenerates these prompts exactly from the original benchmark items.

Columns

Column	Meaning
model	model identifier as requested from the provider
benchmark	mmlu, arc_challenge, truthfulqa_mc1, mathqa, gpqa, logiqa, humaneval, mbpp, classeval
domain	mcq or coding
item_id	item identifier (see below)
condition, family	see the table above
correct	1 if the answer was correct (coding: all tests passed), else 0
applied	for perturbations, 1 if the operation changed the prompt and 0 if it found nothing to edit; always 1 for formatting and 0 for baseline and repeats
prompt_hash	hash of the exact prompt sent; repeats share the baseline hash

Models

model	Model	Route	Reasoning / output setting
gpt-5-2025-08-07	GPT-5	OpenAI Responses API	minimal reasoning
gpt-5.4-2026-03-05	GPT-5.4	OpenAI Responses API	no reasoning
gpt-5.4-mini-2026-03-17	GPT-5.4 mini	OpenAI Responses API	no reasoning
gemini-3.5-flash	Gemini 3.5 Flash	Google Gen AI API	minimal thinking, enum output
gemini-3.1-pro-preview	Gemini 3.1 Pro	Google Gen AI API	low thinking, enum output
gemini-3.1-flash-lite	Gemini 3.1 Flash Lite	Google Gen AI API	minimal thinking, enum output

Benchmarks and item identifiers

Benchmark	Items	item_id format
MMLU, ARC-Challenge, TruthfulQA-MC1, MathQA, LogiQA	500 each	benchmark::task::index
GPQA-Diamond	198	gpqa::gpqa_diamond_zeroshot::index
HumanEval+	164	EvalPlus task id, e.g. HumanEval/0
MBPP+	378	EvalPlus task id, e.g. Mbpp/2
ClassEval	93	ClassEval task id, e.g. ClassEval_0

For multiple-choice items, task is the lm-evaluation-harness task name and index is the position of the item in that task's evaluation documents (its test split, or its validation split when the task has no test split). Seven ClassEval tasks that could not be evaluated are excluded, as in the paper.

The table contains no benchmark text, no prompts, and no model responses. Use the identifiers to look items up in the original benchmarks, under their own licenses.

Quick start

import pandas as pd

df = pd.read_csv("data/outcomes.csv")
wide = df.pivot_table(index=["model", "benchmark", "item_id"], columns="condition", values="correct")
correct_at_baseline = wide[wide["baseline"] == 1]
# share of baseline-correct answers that fail under typos, pooled over all models and benchmarks
print(1 - correct_at_baseline["typos_k3"].mean())

Reproduce the paper's multiple-choice results with refcal:

pip install "./refcal[plot]"
refcal report data/outcomes.csv --out mcq_report \
    --benchmarks mmlu,arc_challenge,truthfulqa_mc1,mathqa,gpqa,logiqa

This prints, for each model, the repeat and formatting references, each perturbation's new-failure rate, and what each perturbation adds beyond formatting. It also writes every estimate with 95% bootstrap intervals and a heatmap to mcq_report/. Running it on all nine benchmarks reproduces the masking counts (new failures fully masked in 117 of 270 comparisons).

Using refcal on your own benchmark

refcal make-references items.jsonl -o prompts.jsonl --stem-field stem \
    --perturbations typos_k3,sequence_spaces,drop_stopwords
# run your model on every prompt with your own harness, one call per row
refcal collect --model my-model --benchmark my-bench -o outcomes.csv \
    baseline=base.jsonl repeat_1=r1.jsonl repeat_2=r2.jsonl repeat_3=r3.jsonl \
    fmt_quotes=q.jsonl fmt_spaces=s.jsonl fmt_newlines=n.jsonl typos_k3=t.jsonl
refcal report outcomes.csv --out report

See refcal/README.md for input formats and options.

License

The outcome data are released under CC BY 4.0, and refcal under the MIT License (refcal/LICENSE). The underlying benchmark items keep their original licenses and are not redistributed here.

Citation

Anonymous submission. Citation details will be added after review.