ARC-AGI-3 Mechanics Clarified: Single Runs, No Shared State, Final Actions Scored
xeophon · x · 2026-08-07
Addressing questions about the ARC-AGI-3 evaluation mechanics, xeophon quoted an official explanation to clarify:
- Single Run: The model only gets one run per game. It cannot "restart" or learn across multiple iterations.
- Compute & Scoring: The model can spend extra compute thinking through a board state, but this incurs an action cost for the RHAE score. While the Kaggle challenge scores all actions taken during evolution, the evolution state is not shared between independent samples.
- Test-Time Training (TTT): This mechanism is similar to the TTT used in ARC-AGI-2, except it learns skills rather than modifying model weights.
More from Research
- NVIDIA Team Builds Data Analysis Agent, Hits #1 on DABStep — JFPuget · 2026-08-07
- Microsoft's Open-Source AI for Beginners Curriculum Hits 63k Stars — bibryam · 2026-08-07
- Gemma 3 27B QAT Fidelity Regression: Why Attention Layers Need More Than Q4_0 — dampflokfreund · 2026-08-07
- Reconstructing Outdoor Scenes from Single 360 Photo: Insta360's G2PS — jiqizhixin · 2026-08-07
- Liquid AI's New Model Uses Multi-Domain Expert Distillation; Revisiting Frontier Techniques — helloiamleonie · 2026-08-07
- Paradigm Shift in Continual Learning: From Parameter-Centric to System-Level Adaptation — CASIA · 2026-08-07