agentIMAGE
RSIAgent paper: GLM-5.3 and Kimi-K3 beat GPT-6 Astra on OSWorld 2.0 without weight changes
The paper reports that GLM-5.3 and Kimi-K3 accumulate and verify experience through autonomous exploration, scoring 78.98% on OSWorld 2.0 versus 72.60% for GPT-6 Astra.
lifcc
@mylifccArchived 2026.09.16Originally posted 2026.09.15
lifcc @mylifccOriginally posted 2026.09.15
刚看到一篇很值得关注的新论文：RSIAgent。 模型权重一行不改。 GLM-5.3 + Kimi-K3 靠自己探索环境、积累经验、验证经验，在 OSWorld 2.0 拿到 78.98%。 论文里 GPT-6 Astra 是 72.60%。 一个方向开始清晰： Scale Model 之外，还有 Scale Experience。
x.com/mylifcc/status/2099785613973115334Archived 2026.09.16 10:20