Diagnosing and Calibrating Tool-Call Boundary Drift in Multi-Teacher On-Policy Distillation
Top-K distillation can hide the token that decides whether a model calls a tool.
The paper audits two-teacher tool-use distillation and finds that high retained probability mass does not guarantee the student gets the decision-critical gradient. In Qwen3.5-9B, restoring the omitted tool-entry coordinate during response supervision cut over-calling from 14.2% to 3.7% across three seeds, while reducing call recall by 12.4 points. The authors separate three fixes: student-aware support, loss shaping, and inference-time entry bias, each with different costs. Llama-3.1-8B showed the same directional support asymmetry under its JSON tool protocol. HF Daily Papers' note
The paper audits two-teacher tool-use distillation and finds that high retained probability mass does not guarantee the student gets the decision-critical gradient. In Qwen3.5-9B, restoring the omitted tool-entry coordinate during response supervision cut over-calling from 14.2% to 3.7% across three seeds, while reducing call recall by 12.4 points. The authors separate three fixes: student-aware support, loss shaping, and inference-time entry bias, each with different costs. Llama-3.1-8B showed the same directional support asymmetry under its JSON tool protocol. HF Daily Papers' note
score 4