RSI-Jev v2.0：AutoScientists 自动跑实验迭代决策模型

Shanghua Gao @GaoShanghua原推发布于 2026.09.26

在 X 上打开原推

~15 hours after RSI-Jev v1.0, our AutoScientists agent team had already run 50 more experiments. 4 made it into v2.0. At 2B, RSI-Jev v2.0 reaches 0.79 vs Jev 1.13’s 0.73 on typed-decisions, and 0.849 vs Qwen3.5-4B’s 0.813 on an unseen benchmark. One interesting result: task-specific gains didn’t transfer much, but uncertainty calibration did. Still early, but this is the RSI loop we want to keep pushing. GitHub: https://github.com/Shanghua-Gao/RSI-Jev v2.0: https://huggingface.co/shgao/rsi-jev-v2.0-qwen3.5-2b

这条是引用着另一条写的：x.com/gaoshanghua/status/2103572751466312150