RSI-Jev v6.1-VL 27B tops the public Decision Index 0.3
AutoScientists' multimodal System One model, built entirely by automated research, returns calibrated probabilities in one forward pass and scores 65.64 on Decision Index 0.3, ahead of Jev 1.13 at 57.96.
Our AutoScientists just released RSI-Jev v6.1-VL 27B! It is now #1 on the public Decision Index 0.3, with a score of 65.64, ahead of Jev 1.13 at 57.96 and other strong decision models. It is a multimodal System One model: it reads images, makes decisions in one forward pass, uses no reasoning tokens, and returns calibrated probabilities. With effort set to auto, it chooses how deep to go for each question. About 72% are settled at three-quarters depth, while only one in five needs the full model. Because it knows how confident it is, auto effort is both faster and more accurate than always using full depth. No one hand-built it. AutoScientists proposed it, trained it, tested it, and shipped it. Every experiment is public. When I launched this project, I didn’t expect AutoScientists to get this far this quickly. It’s making me more confident that RSI could become a real way to build the next generation of AI models. GitHub: https://github.com/Shanghua-Gao/RSI-Jev Weights: https://huggingface.co/shgao/rsi-jev-v6.1-vl-27b Release: https://github.com/Shanghua-Gao/RSI-Jev/blob/main/versions/v6.1-vl-27b.md
This post quotes another post: x.com/gaoshanghua/status/2103572751466312150