A 150M param recurrent model scores 29.5% on ARC-AGI-1 at $0.0007 per task

模型发布

热度趋势

趋势数据积累中

百分比基于当前可用热度信号，而非评论数或独立用户人数。

AI 摘要

一个拥有1.5亿参数的新型循环模型在ARC-AGI-1上取得了29.5%的分数，每个任务的成本为0.0007美元。该模型由Pathway团队开发，采用循环潜在推理设置，在提供答案之前在潜在空间中处理信息。它在ARC-AGI的典型成本/准确性边界之外运行，其紧凑的尺寸使其能够在各种系统上运行。研究人员渴望看到其性能扩展到1-3B参数。

Not a transformer. It's a recurrent latent reasoning setup that keeps "thinking" in latent space before answering. Sits completely outside the published cost/accuracy frontier for ARC-AGI, and something this size runs on basically anything. Paper is from the Pathway team, dropped 4 days ago. I want to see it scaled to 1-3B before getting too excited, but the shape of the result is wild.