DGCR: Learning from Local Corrections for Super-Tiny Function Calling Models

摘要

A partially correct tool call can contain a useful learning opportunity even when its complete output fails. Two limitations make that opportunity difficult to exploit: a response-level reward does not specify a local correction, and equal rewards within a rollout group eliminate the reward-relative update. We introduce Degenerate-Group Counterfactual Repair (\method{}), which addresses both limitations with verified local supervision for function calling in super-tiny language models. From eligible groups rejected by reinforcement learning, DGCR constructs bounded structural edits using training references, verifies their gains under controlled call assignments, and applies gain-weighted cross-entropy to the repaired spans under prefixes generated by the student. The same repair supplies an explicit local target and recovers supervision from a group that receives no direct Sim-RL update. It requires no additional language-model query for target construction and adds no inference-time component. For Qwen3-0.6B, the reported BFCL v3 Overall Acc is 53.45%, versus 51.76% for the CKD+Sim-RL reference; this comparison remains provisional pending verification of the original multi-turn records.