Lisan al Gaib· @scaling01 · X·· 3 小时前AI 评分51
AI 导读
针对 Gemini 4 Argon 在 ARC-AGI-3 上取得 97.5%（provider adapter harness，$4.8K）与 49.8%（standard harness，$5.8K）的成绩，作者表示不确定循环推理在其中扮演了什么角色，但认为靠大量游戏 RL 就能解决，循环推理可能只是让成本大幅降低。他给出的整体判断是约 85% 归因于游戏 RL、15% 归因于循环推理，主要依据是 Opus 5 的数据点和 AI 领域一切都能做 RL 的趋势。
正文
I'm not quite sure what role loopers played in saturating ARC-AGI-3
but I definitely believe it's solvable by just a shit ton of RL on games. loopers probably just made it a lot cheaper
overall I think I am at 85% RL on games and 15% looping for what caused this sudden jump, mostly because of the Opus 5 datapoint and general trends in AI (everything can be RL'd)
Gemini 4 Argon from @Google on ARC-AGI (Verified): - ARC-AGI-3: 97.5%, $4.8K (provider adapter harness), 49.8%, $5.8K (standard harness) - ARC-AGI-2: 91.3%, $0.32/task - ARC-AGI-1: 98.5%, $0.16/task Google DeepMind is the second frontier AI lab to score above 95% on ARC-AGI-3.在 X 查看被引用的帖子
来源：Lisan al Gaib · x.com