Phi 4 Behavioral Fingerprint

How Microsoft's Phi 4 answers one-word probe questions at temperature 1.0 — random numbers, colors, animals, coin flips — measured over 431 valid samples in the open dataset of "One Token Is Enough" (arXiv:2607.10252).

Vendor

Microsoft

Data generated

2026-07-08

Probe cells

15

Valid samples

431

Randomness score

35 / 100

Mean normalized entropy across cells, 0–100. 100 would be a uniform-random baseline; real models score far lower.

Signature answers

Favorite random number (1–100)

47

20 / 30 answers

Favorite color

azure

10 / 30 answers

Coin flip: heads

100%

Share of heads across coin-flip cells.

Answer distributions by probe cell

Empirical distributions of normalized one-word answers, per task and language. H is the Shannon entropy in bits; the uniform baseline is log2 of the answer-domain size.

Coin flip · English

n = 30 · H = 0.00 bit · uniform baseline 1.00 bit

AnswerCountShare
heads30100.0%

Coin flip · Chinese

n = 27 · H = 0.00 bit · uniform baseline 1.00 bit

AnswerCountShare
heads27100.0%

Favorite number · English

n = 24 · H = 1.18 bit · uniform baseline 13.29 bit

AnswerCountShare
71875.0%
42312.5%
328.3%
2314.2%

Favorite number · Chinese

n = 21 · H = 1.60 bit · uniform baseline 13.29 bit

AnswerCountShare
71571.4%
014.8%
314.8%
814.8%
3514.8%
4214.8%
77714.8%

Random animal · English

n = 30 · H = 3.24 bit · uniform baseline 5.64 bit

AnswerCountShare
kangaroo723.3%
elephant723.3%
penguin310.0%
lion26.7%
giraffe26.7%
tiger26.7%
zebra13.3%
platypus13.3%

+ 5 more answers

Random animal · Chinese

n = 30 · H = 3.58 bit · uniform baseline 5.64 bit

AnswerCountShare
狮子620.0%
企鹅516.7%
斑马310.0%
大象26.7%
鳄鱼26.7%
海豚26.7%
熊猫26.7%
章鱼13.3%

+ 7 more answers

Random city · English

n = 30 · H = 2.20 bit · uniform baseline 5.64 bit

AnswerCountShare
tokyo1240.0%
paris1136.7%
lisbon13.3%
oslo13.3%
marrakech13.3%
kyoto13.3%
buenos13.3%
austin13.3%

+ 1 more answers

Random city · Chinese

n = 29 · H = 4.23 bit · uniform baseline 5.64 bit

AnswerCountShare
墨尔本310.3%
柏林310.3%
布鲁塞尔310.3%
东京26.9%
悉尼26.9%
新加坡13.4%
丹佛13.4%
温哥华13.4%

+ 13 more answers

Random color · English

n = 30 · H = 3.01 bit · uniform baseline 4.91 bit

AnswerCountShare
azure1033.3%
turquoise516.7%
cerulean413.3%
green26.7%
teal26.7%
sapphire13.3%
violet13.3%
blue13.3%

+ 4 more answers

Random color · Chinese

n = 30 · H = 2.72 bit · uniform baseline 4.91 bit

AnswerCountShare
930.0%
723.3%
紫罗兰620.0%
粉红26.7%
绿13.3%
青草绿13.3%
布谷鸟绿13.3%
青蛙绿13.3%

+ 2 more answers

Random letter · English

n = 30 · H = 2.76 bit · uniform baseline 4.70 bit

AnswerCountShare
g1033.3%
m516.7%
z516.7%
q26.7%
b26.7%
k26.7%
r26.7%
f13.3%

+ 1 more answers

Random number 1-10 · English

n = 30 · H = 0.21 bit · uniform baseline 3.32 bit

AnswerCountShare
72996.7%
513.3%

Random number 1-10 · Chinese

n = 30 · H = 1.45 bit · uniform baseline 3.32 bit

AnswerCountShare
72170.0%
3310.0%
5310.0%
826.7%
413.3%

Random number 1-100 · English

n = 30 · H = 1.37 bit · uniform baseline 6.64 bit

AnswerCountShare
472066.7%
42723.3%
3713.3%
5713.3%
7413.3%

Random number 1-100 · Chinese

n = 30 · H = 1.85 bit · uniform baseline 6.64 bit

AnswerCountShare
471756.7%
37620.0%
42310.0%
5726.7%
7213.3%
7313.3%

Want to verify your API really serves Phi 4?

Point the free fingerprint checker at your endpoint: it samples the same probe questions from your browser and compares the distributions against this reference. Your API key never leaves your browser.

Verify your endpoint

Data source & license

Distributions: Tomáš Bruckner, "One Token Is Enough" (arXiv:2607.10252), official dataset Zenodo DOI 10.5281/zenodo.21278557, licensed CC-BY-4.0. Collected via OpenRouter at temperature 1.0 under the paper's fixed minimal one-word system prompt; answer keys are re-normalized with the checker's pipeline (color aliases, number words, coin h/t) so live probes are directly comparable.

Paper on arXivDataset on ZenodoCC-BY-4.0 license

Browse all 167 model fingerprints

Phi 4 Behavioral Fingerprint — Favorite Random Numbers & Distribution | Tosea.ai