Search papers, labs, and topics across Lattice.
This paper investigates the effectiveness of multi-teacher on-policy distillation (OPD) in training agentic language models to balance tool usage and direct responses. The authors introduce the concept of behavior leverage imbalance, revealing that local token-level signals can disproportionately influence global generation modes, leading to issues like over-calling tools. By implementing a novel calibration method called Soft Clamp, they achieve a significant reduction in over-calling from 13.7% to 9.0% while maintaining decision accuracy, highlighting the importance of monitoring teacher signal locations in training.
Tool over-calling can be reduced by nearly 4% through a novel calibration method that reveals the hidden dynamics of token-level influences 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.
