AI Trends

ARC Prize Evaluates Gemini 4 Hydrogen's Performance on ARC-AGI

ARC PrizeUpdated Nov 21 01:28

AI summary

ARC Prize measured Gemini 4 Hydrogen at 92.5% on ARC-AGI-2 and 98.5% on ARC-AGI-1, respectively, while using about 52% fewer reasoning tokens than Gemini 3.8 Flash at the high reasoning level. ARC-AGI-3 testing could not be completed due to API issues, including multi-action function calls, MALFORMED_FUNCTION_CALL errors, repeated actions, and RECITATION termination reasons. ARC Prize also invited participants to a NeurIPS rooftop networking event in Atlanta on December 12; space is limited, and registration requires organizer approval.

Why read this

AI score 62

The evaluation also presents scores, token usage, and API errors across different reasoning levels, enabling comparisons of model performance and integration stability.

Topics

Tags

Body

@GoogleDeepMind Join ARC Prize at NeurIPS on December 12, 6-9 PM in Atlanta for a rooftop happy hour across from the convention center.

Space is limited; registration requires host approval.

https://t.co/1cdZppwijd

Source

ARC Prize