Back to explore

4B Jev-style Judge Outperforms Up to 27B Models on Robotics Benchmarks

The author shares a 4B Jev-style judge trained with public data and their own recipe, beating multimodal Jev-style models up to 27B on RoboPulse, RoboRewardBench, and real-robot rollouts, while noting specialist reward models still win at home.

3/6 How far can public data + our recipe take a 4B Jev-style judge? It beats every multimodal Jev-style model we tested, up to 27B, on RoboPulse, RoboRewardBench and real-robot rollouts. Specialist reward models still win at home: Robo-Dopamine on progress, RoboReward on

4B Jev-style Judge Outperforms Up to 27B Models on Robotics Benchmarks 1
· 0 likesOpen on X