Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0%
antgroup · hf · 2026-07-21
What the paper says
Multi-teacher on-policy distillation for agentic language models can improve tool-call recall, but the paper finds it also shifts behavior toward over-calling: the student starts invoking tools in cases that should be answered directly.
- The authors argue that aggregate losses can miss this boundary drift.
- They analyze the problem as behavior leverage imbalance: token-level signals near mode-entry points such as <toolcall> and function names can dominate the global generation mode.
- To address this, they introduce Soft Clamp, a per-token divergence calibration method that compresses extreme Jensen-Shannon divergence while keeping gradients nonzero.
Results
- On APIGen-MT, Soft Clamp reduces over-calling from 13.7% to 9.0% relative to vanilla GKD while matching decision accuracy.
- On a BFCL multi-turn diagnostic, it also reduces tool-call loops and repeated calls among GKD variants.
The main takeaway is that multi-teacher OPD should monitor not just how much teacher signal there is, but where it acts during generation.
More from coding & agent
- TWSE MCP Server brings Taiwan stock data into natural-language workflows — modelcontextprotocol · 2026-07-21
- The hardest part of agent execution may be permissions, approvals and failure handling — marcelk231 · 2026-07-21
- Business chatbots should add RAG only when answers depend on company-specific documents — recro69 · 2026-07-21
- A privacy email provider is building email for AI agents, with JMAP, PoW signups and 2,000 accounts — Atomic_Ke · 2026-07-21
- Open-source PenEcho AI whiteboard now renders interactive animated explanations — Civil-Direction-6981 · 2026-07-21
- AI agents are making full UI specs obsolete, so teams should build a “baby” app first — haltakov · 2026-07-21