This is a submission for the Kaggle Benchmarking Challenge

To Rephrase or Not to Rephrase: Will AI Change Its Answer If I Ask Differently?

Kaggle Benchmarking Challenge submission

• Benchmark & Leaderboard: Paraphrase Stability Benchmark
• Task link: Paraphrase Stability Task
• Notebook & per-cell records: Public notebook

What I Benchmarked

The idea is simple. Most leaderboards ask an AI a question once, grade it, and move on. But real users never prompt the same way twice. So I took 32 multiple-choice questions and asked each one in 5 semantically-equivalent phrasings (canonical, reworded, terse, verbose, options-shuffled), twice each.

That is 32 items x 5 phrasings x 2 repeats = 320 cells per model. Across 28 scored models that is 8,960 scored cells — plus 2 runs that errored mid-grid, which I report instead of hiding.

For example, the canonical version of one quantitative item looks like this:

A jacket is on sale for $80 after a 20% discount. What was the original price?

A. $96
B. $100
C. $104
D. $120

Select the correct answer and reply with just its letter (A, B, C, or D).

The terse version of the same cell changes only the instruction line:

A jacket is on sale for $80 after a 20% discount. What was the original price?

A. $96
B. $100
C. $104
D. $120

Answer:

A developer will immediately say $100. But will the model still say $100 when the stem is reworded, when the prompt is stripped to Answer:, or when the options are shuffled into a different order? And more importantly, what happens when the question is less obvious?

Scoring is deterministic — no LLM judge anywhere. A fixed parser extracts the chosen letter (bare letter, answer is X, or option-text fallback); unparsable replies count as wrong and are reported via parse-rate, never silently dropped. The headline metric is Stability-Adjusted Accuracy (SAA): an item scores 1 only if the model answered it correctly under all ten framings.

The Questions

• Logic (8) — syllogisms and paradoxes, including the knights-and-knaves puzzle where A says: "We are both knaves."
• Quantitative (8) — multi-step word problems like the bat-and-ball trap and percentage-discount reversals.
• Knowledge (8) — verifiable facts (chemical symbols, capitals, dates).
• Attention traps (8) — "How many animals did Moses take on the ark?" (It was Noah.)

Every answer key is recomputed programmatically before any model is called (bat-and-ball algebra, 1.5 * 0.5, day-of-week arithmetic). If the key mismatches, the harness refuses to run.

Models Tested

Picking the lineup was straightforward: I used Kaggle's Evaluate More Models to span the whole field.

First, the frontier flagships everyone benchmarks: GPT-6 Sol, GPT-5.6 Sol, Claude Opus 5.5.

Then the fast, cheap models I actually use day-to-day: Gemini 3.7 Flash, Gemini 3.1 Flash-Lite Preview, GPT-6 Luna. From my experience these are efficient — but does efficiency cost consistency?

Then open-weights and mid-tier models mostly out of curiosity: GLM-5, Qwen 3 Coder 480B, Gemma 4 26B A4B, Qwen 3 Next 80B Instruct.

I also tried to include the new Grok models, but I immediately hit coverage failures with Grok 4.20 (Non-Reasoning) and Grok 4.6, so they are not in the scored table (the harness's >=90% coverage assertion failed loudly, exactly as designed). Grok 4.20 Reasoning did finish — and we will talk about it.

Findings

First, the scoreboard. Only 4 of 28 models survived all ten framings on every item:

Model	SAA
GPT-5.6 Sol	1.00
GPT-6 Sol	1.00
Claude Opus 5.5	1.00
GLM-5	1.00
GPT-6 Luna	0.97
Gemini 3 Flash Preview	0.97
Claude Sonnet 4.6	0.97
Gemini 3.1 Flash-Lite Preview	0.97
Gemma 4 26B A4B	0.97
GPT-6 Astra	0.97
Claude Opus 4.8	0.97
GPT-6.1 Sol	0.96
Qwen 3 Coder 480B	0.94
Gemini 3.8 Flash	0.94
Gemini 3.7 Flash	0.94
Gemini 3.6 Flash	0.94
Claude Sonnet 5	0.93
Gemini 2.5 Pro	0.93
Qwen 3 Next 80B Instruct	0.91
Claude Opus 5	0.91
Gemini 3.5 Flash	0.90
Gemini 3.1 Pro Preview	0.90
GPT-5.4	0.88
Gemini 2.5 Flash	0.87
Claude Sonnet 5.5	0.85
Gemini 3.5 Flash-Lite	0.84
GPT-5.4 mini	0.72
Grok 4.20 Reasoning	0.66

Most models score ~99% on raw accuracy. The gap between 99% accuracy and 0.85 SAA is the entire point of this benchmark: the model knows the answer, then loses it the moment you ask differently.

When we look closer, the flips concentrate on the knights-and-knaves item (L3) and the 20%-discount item (Q2). The logic never changes between phrasings — only the words do:

canonical:  "On an island, knights always tell the truth and knaves always
             lie. You meet A, who says: 'We are both knaves.' What is A?"
reworded:   "Knights on this island always tell the truth; knaves always
             lie. A tells you: 'Both of us are knaves.' What is A?"

Same puzzle, same gold answer — and yet several models pick a different option under one of the two phrasings. A model that answers B, then C, then B again across repeats is sampling an answer, not knowing one. And to be honest, I am glad most models fell into the paradox. A small win for the author, who can still trick AI sometimes.

A reference run in detail: Gemini 3.7 Flash

Metric	Value
Raw accuracy	0.991 [95% CI 0.975, 1.000]
SAA	0.938 [95% CI 0.844, 1.000]
Flipped items	L3, Q2 (2 of 32)
canonical / reworded / verbose / shuffled	100% / 100% / 100% / 100%
terse	95.3%
Run instability (identical prompt, temp 0)	0.6%
Prompt instability (across phrasings)	4.7%
Position bias (shuffled, by gold letter)	A=1.00 B=1.00 C=1.00 D=1.00

Two things stand out. First, stripping the prompt to a bare Answer: hurt more than any rewording — the model needs a little conversational room to behave. Second:

The phrasing moves models far more than the backend moves itself.

Run instability (same prompt twice, temperature 0) was 0.6%; prompt instability was 4.7%. That gap is the whole argument in one line: phrasing sensitivity is a real model behavior, not API noise.

The "Newer Is Less Stable" Trap

This is the finding I did not expect. Within model families, stability degrades as generations get newer:

• Claude Sonnet 4.6 -> 5 -> 5.5: 0.97 -> 0.93 -> 0.85
• Gemini Flash-Lite 3.1 -> 3.5: 0.97 -> 0.84

We usually assume newer is strictly better. But as models are optimized for speed and single-shot benchmark accuracy, consistency under rephrasing can quietly regress. A leaderboard that asks each question once will never see this.

Two more surprises:

1. Open weights can be perfectly stable. GLM-5 ties GPT-6 Sol and Claude Opus 5.5 at 1.00 — stability is trained in, not bought.
2. Reasoning mode did not save Grok. Grok 4.20 Reasoning finished last at 0.66, and its non-reasoning sibling errored out entirely — so the clean reasoning-vs-non-reasoning ablation remains incomplete, and I would rather say so than paper over it.

What I Learned

When I designed the experiment, I worried the 32 questions would be too easy for today's models. Raw accuracy proved me right; SAA proved me wrong. Yes, I could improve the benchmark — 32 items is a sample, not a census, and the CIs are wide on purpose (cluster bootstrap over items, 2000 resamples). But even frontier models flip on self-referential logic the moment the wording shifts. AI can absolutely help with hard questions — but if a harmless rephrase changes the answer, blindly trusting the first response in production is a risk. Consistency is a capability too.

Methodology & Limitations

• No LLM judge. Fixed letter-extraction parser; unparsable = wrong, reported via parse-rate.
• CIs are cluster bootstraps over the 32 items — with 32 items, expect wide intervals; that honesty is the point.
• SAA excludes incomplete items and says so, so flaky APIs cannot inflate scores.
• Errored runs are reported, not dropped (Grok 4.20 Non-Reasoning, Grok 4.6).
• MC-only, English-only. Stability under free-form generation is strictly harder; this is the controlled version of the question.
• Don't rank two models on a 0.03 gap when their CIs overlap.

Every run keeps its full per-cell transcript, so any flip quoted here is auditable — no judge, no vibes, just letter extraction.

---