RSI-Jev v6.0-VL：按置信度自动调节推理深度

Shanghua Gao @GaoShanghua原推发布于 2026.10.06

在 X 上打开原推

System One models can now adjust their inference effort, not by generating more tokens, but by going deeper through the network. AutoScientists introduced RSI-Jev v6.0-VL, where each decision can use low, medium, or high depth, or choose automatically. The key is confidence awareness. System One models know how sure they are, and AutoScientists designed the auto-effort mechanism around that uncertainty. If the model is confident, it can stop early. If not, it looks deeper. We found that different decisions benefit from different depths: intent and routing are often resolved early, judgement benefits from intermediate depth, while knowledge-intensive and reasoning-heavy decisions benefit from going deeper. On the public Decision Index, v6.0-VL improves the 4B score from 38.38 to 46.24, making it the strongest 4B model there. Take-home: for System One models, thinking harder does not have to mean thinking longer. It can mean using more of the network only when needed. And AutoScientists have already found something better. Code: https://github.com/Shanghua-Gao/RSI-Jev Weights: https://huggingface.co/shgao/rsi-jev-v6.0-vl-4b Effort/depth analysis: https://github.com/Shanghua-Gao/RSI-Jev/blob/main/docs/effort-depth.md

这条是引用着另一条写的：「砍掉后 12 层，3B 决策模型 RSI-Jev 超过 4B 模型」