Asymmetric Clipping Boosts Policy Gradient RL Accuracy By 10.1%

End result: 84.7% accuracy, beating every frontier model tested, at roughly 1/14th the inference cost.

The trained model made 29.8% fewer mistakes than the best frontier model.

CISPO loss with asymmetric clipping, replacing standard importance sampling. In policy gradient RL, you reweight updates by the ratio of the new and old policy's probabilities for an action, then clip that ratio so a single update can't swing too far. Standard clipping is symmetric, with the same bound in both directions.

On-policy distillation with a moving teacher. The student is regularized toward a teacher model's output distribution, penalized for drifting away from it. But the teacher isn't fixed. Every 20 steps the current student checkpoint can be promoted to teacher, only if it has just hit a new validation high. +3.1% accuracy.

The bigger claim here is that general-purpose frontier models may plateau on narrow, task-heavy enterprise tasks, and small fine-tuned models trained on high-quality expert data can beat them on both accuracy and cost.

Very cool.