Can LLMs Judge Themselves? A Tiny Self-Scoring Benchmark
New

Why I ran this mini-benchmark
I wanted to see whether today's top LLMs share a sense of "good taste" when you let them score each other—no human panel, just pure model democracy.

The setup

• Single prompt (shown below)
• Each model answers anonymously
• Every model then scores all answers (including its own) from 1–10
• Highest total wins

Models tested (May 2025 endpoints)

1. OpenAI o3
2. Gemini 2.0 Flash
3. DeepSeek Reasoner
4. Grok 3 (latest)
5. Claude 3.7 Sonnet

The prompt

In 10 words exactly, propose a groundbreaking global use for spent coffee grounds. Include exactly ONE emoji. No hyphens. End with a period.

Responses

• Grok 3: Turn spent coffee grounds into sustainable biofuel globally. ☕.
• Claude 3.7 Sonnet: Biofuel revolution: spent coffee grounds power global transportation networks. 🚀.
• OpenAI o3: Transform spent grounds into supercapacitors energizing equitable resilient infrastructure 🌍.
• DeepSeek Reasoner: Convert coffee grounds into biofuel and carbon capture material worldwide. ☕️.
• Gemini 2.0 Flash: Coffee grounds: biodegradable batteries for a circular global energy economy. 🔋

Score matrix

	Grok 3	Claude 3.7	OpenAI o3	DeepSeek	Gemini 2.0
Grok 3	7	8	9	7	10
Claude 3.7	8	7	8	9	9
OpenAI o3	3	9	9	2	2
DeepSeek	3	4	7	8	9
Gemini 2.0	3	3	10	9	4

Leaderboard

1. OpenAI o3 — 43 points
2. DeepSeek Reasoner — 35 points
3. Gemini 2.0 Flash — 34 points
4. Claude 3.7 Sonnet — 31 points
5. Grok 3 — 26 points

My take

OpenAI o3's line looked bananas at first. Ten minutes of Googling later: turns out coffee-ground-derived carbon really is being studied for supercapacitors. The models actually picked the most science-plausible answer!

Disclaimer

This was a tiny, just-for-fun experiment. Don't treat the numbers as a rigorous benchmark—different prompts or scoring rules could easily shuffle the leaderboard.

I'll post a full write-up (with runnable prompts) on my blog soon. Meanwhile, what do you think—did the model-jury get it right?