Soft Clamp reduces over-calling in multi-teacher on-policy distillation

Researchers identify a behavior leverage imbalance in multi-teacher on-policy distillation where vanilla generalized knowledge distillation causes models to over-call tools despite improved recall. To address this, they propose Soft Clamp, a per-token divergence calibration method that dynamically compresses extreme token-level Jensen-Shannon divergence while preserving nonzero gradients.

• Soft Clamp reduces tool-call over-calling from 13.7% to 9.0% on APIGen-MT relative to vanilla GKD.
• The method matches the decision accuracy of vanilla generalized knowledge distillation.
• In BFCL multi-turn diagnostics, Soft Clamp lowers tool-call loops and repeated calls among GKD variants.

These results suggest that multi-teacher on-policy distillation should monitor where teacher signals act at local token-level positions rather than relying solely on aggregate loss metrics.