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
- One prompt, full Blender animation: dev says he never touched Blender once — iamfakhrealam · 2026-10-26
- Claude Code finds root backdoor in Honda Civic head unit, then debloats it — IndraVahan · 2026-10-26
- Claude Code + Monid MCP builds a $12K/year B2B prospecting engine on live data — aryanXmahajan · 2026-10-26
- Karpathy shares CLAUDE.md rulebook to stop AI coding agents from piling up tech debt — mdancho84 · 2026-10-26
- Rethinking dependencies in the agent era: drop libs that save 20 lines of code — gethackteam · 2026-10-26
- RustGIS hits 2,000 GitHub stars in two weeks with agent-first architecture — giswqs · 2026-10-26