ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generation and Dynamic Self-Feedback
ToolLoop reports BFCL gains from 11K synthetic tool-use examples built with iterative self-feedback.
The paper proposes a three-stage synthesis pipeline: sample function-name combinations, derive user queries backward, then derive tool calls forward. Its feedback loop shifts generation from static filtering to repeated generate, verify, and refine steps. A 4B model trained on the data reaches 86.40% BFCL accuracy in non-reasoning mode, or 86.07% when overlapping candidate functions are removed. On ACEBench, it reports 72.1% overall accuracy using 18.3% of the baseline training data. ArXiv · AI/CL/LG's note
The paper proposes a three-stage synthesis pipeline: sample function-name combinations, derive user queries backward, then derive tool calls forward. Its feedback loop shifts generation from static filtering to repeated generate, verify, and refine steps. A 4B model trained on the data reaches 86.40% BFCL accuracy in non-reasoning mode, or 86.07% when overlapping candidate functions are removed. On ACEBench, it reports 72.1% overall accuracy using 18.3% of the baseline training data. ArXiv · AI/CL/LG's note
score 5