Spaces:

---

FINAL-Bench

/

Leaderboard

like 22

Running

Just curious

What if Turing met ChatGPT?
The first 5 minutes, he'd be thrilled. "Good heavens, it actually holds a conversation!"
The next 5 minutes, he'd be confused. "Wait — why is it so confident about something it just got completely wrong?"
That's exactly what we found building FINAL Bench.
Today's LLMs ace exams. The problem? They wear the same poker face whether they're right or wrong. In our data, 94.8% of metacognitive failure came from models declaring "I totally know this!" — and then getting it wrong.
The Turing Test asks: "Can it pass as human?" → Pass.
FINAL Bench asks: "Does it know what it doesn't know?" → Fail.
We think Turing would sum it up like this:
"Brilliant speaker. Terrible self-critic."