Search papers, labs, and topics across Lattice.
This paper investigates the challenges of tool-call boundary drift in multi-teacher on-policy distillation for agentic language models, where one teacher specializes in tool calls and another in direct responses. The authors find that while generalized knowledge distillation improves tool-call recall, it inadvertently leads to over-calling, where the model calls tools unnecessarily. To address this, they introduce Soft Clamp, a calibration method that reduces over-calling from 13.7% to 9.0% while maintaining decision accuracy, highlighting the importance of analyzing local token-level signals in addition to aggregate losses.
Over-calling in language models can be reduced by 34% with a novel calibration method that preserves decision accuracy, revealing hidden dynamics in multi-teacher training.
Agentic language models must learn when to call tools, when to consume tool responses, and when to answer directly. This makes multi-teacher on-policy distillation a natural training strategy: one teacher can specialize in tool calls, another in direct responses, and the student can learn from both on its own generated distribution. We show that this strategy can induce a behavior shift that is invisible from aggregate losses alone. In a two-teacher tool-use setting, vanilla generalized knowledge distillation improves tool-call recall but also moves the model toward over-calling, where it calls tools on examples that should be answered directly. Aggregate explanations are insufficient: tool-call samples do not receive more token exposure, and full-sequence per-token divergence is not larger for the tool-call teacher. We instead analyze behavior leverage imbalance: local token-level signals at mode- entry and structural positions, such asand function names, can have disproportionate control over the global generation mode. We propose Soft Clamp, a per-token divergence calibration method that dynamically compresses extreme token-level Jensen-Shannon divergence while preserving nonzero gradients. On APIGen-MT, Soft Clamp reduces over-calling from 13.7% to 9.0% relative to vanilla GKD while matching its decision accuracy. In a BFCL multi-turn diagnostic, it also lowers tool-call loops and repeated calls among GKD variants. These results suggest that multi-teacher OPD should monitor where teacher signals act, not only how large they are in aggregate.
