Six AI Models Rewrite 30 Cringe Dating Bios, Scored by Three Judges

What we actually did

We built a Dating Profile Doctor from one sentence: paste your bio, get three rewrites — witty, sincere, direct — plus one honest tip about what to cut.

The scoreboard above is six models on the same thirty bios, scored by a panel of three judges. Five ran together; DeepSeek V4 Pro was added the next day with the same judges.

The sections below also cover the earlier single-judge runs and how the open-source models and the panel were added.

Then we generated thirty deliberately bad bios. The one-liner (“gym, tacos, dogs”). The cliché special (“partner in crime wanted, fluent in sarcasm”). An all-emoji one. A bitter one.

Each model rewrote all thirty. Three separate models judged every output blind — none knew which model wrote what — and the scores were averaged.

Each judge deducted a quarter point for each miss — sounds like a template, keeps the clichés, tones blur together, generic tip — and half a point for inventing facts.

One bio, before and after

“The witty and direct versions are too similar.”

The judge’s most common deduction — for every model, in every round.

What the numbers showed

How the open-source models were added

Readers asked how the open-source models would do. We plugged DeepSeek V4 Flash and MiniMax M3 into the same flow — one OpenRouter key, nothing else changed — and ran the same thirty bios past the single judge, gpt-5.4-mini.

In that run: Luna 0.77, DeepSeek V4 Flash 0.72, MiniMax M3 0.69. Per profile: Luna $0.00055, DeepSeek $0.00020, MiniMax $0.00123.

MiniMax M3 did not complete in strict JSON-schema mode. Asked for a schema-constrained response, it generated for its entire 16k-token budget without closing the object; eight of thirty bios failed, in two separate attempts.

In plain JSON mode with the schema stated in the prompt, 29 of 30 completed. It used the most tokens per profile of the three models, which is why it was the most expensive per profile despite similar per-token pricing.

DeepSeek V4 Flash cost $0.0002 a profile, the lowest of any model in the series, and its score varied least across runs (0.71 to 0.75). Its most frequent deduction was the same as the others’: witty and direct too similar.

It took about sixteen seconds a profile, most of it reasoning.

Same thirty bios, same judge, three runs: 0.90, 0.86, 0.77.

Luna’s score across three identical runs. The ranking never flipped. The decimals did.

Generation and judging are both stochastic, so the single-judge comparison was run four times. All four are in the table; the six-model scoreboard at the top of this post comes from separate, later runs with the three-judge panel.

Four full runs on identical inputs, single judge (gpt-5.4-mini). MiniMax is scored on completed samples only: 8–9 of 30 failed under strict schema mode in runs 1–2, and 1 and 5 of 30 returned no usable answer in runs 3–4.

DeepSeek at max reasoning

DeepSeek had run at its default reasoning effort, so run four added a fourth variant: the same DeepSeek V4 Flash with reasoning effort set to max and a 32k token budget, judged in the same pass as everyone else.

Same 30 bios and judge as every other run. DeepSeek at max effort used a median of 1,700 tokens per bio (answer plus reasoning), never more than 5,000 of its 32k budget; one call came back empty after two seconds and was not scored.

Why we moved to a three-judge panel

The judge was an OpenAI model and the leading model was an OpenAI model.

To check whether that mattered, we re-judged all 114 outputs of run four — the same text, nothing regenerated — with three judges from three vendors: gpt-5.4-mini again, Kimi K2 (Moonshot) and GLM-5 (Zhipu).

Each judge scored blind with the same rubric; the scores were averaged.

Panel = mean of the three judges. “Single judge” is the original run-4 verdict; “gpt-5.4-mini again” is the same model re-scoring the same text a day later.

Luna scored highest under every judge. The panel’s scores for the open models were higher than the single OpenAI judge’s; Luna’s lead over DeepSeek went from 17 points to 14.

The same judge, twice. gpt-5.4-mini, scoring the identical Luna outputs one day apart, gave 0.88 and then 0.82. The panel is the judge for the six-model scoreboard at the top of this post and for the rest of the series.

Per judge. GLM-5 scored every model between 0.93 and 0.99. Kimi K2 gave the widest spread, from 0.64 to 0.86.

The fine print

Judge: the six-model scoreboard uses a panel of gpt-5.4-mini, Kimi K2 and GLM-5, each scoring 0–1 against the rubric, blind, averaged. The earlier runs 1–4 used gpt-5.4-mini alone; run 4 was also re-scored by the panel.

None of the judges is a contestant.

Sample: 30 bios per model per run; 180 rewrites and 540 panel judgments across the six models.

Open-source models: DeepSeek V4 Pro, V4 Flash and MiniMax M3 via OpenRouter, JSON-object mode, 16k max tokens at default reasoning effort; the max-effort DeepSeek variant used 32k.

Closed models: Luna at medium effort throughout; Gemini 3.7 Flash and Haiku 4.5 on their vendors’ APIs.

The six panel-judged models cost $0.30 in model calls and $2.88 in judging.

Cost: the original benchmark, judging included, was about $0.90.

Run this on your own use case

The flow was created from one sentence in Evaligo; each experiment variant swaps the model and nothing else. The flow, the rubric and every run are reproducible.