ToolLoop is a new method for synthesizing tool-use training data. Instead of generating full samples in one pass and filtering, it breaks synthesis into three stages: sampling function names as ground truth, backward derivation of user queries, and forward derivation of tool calls. At each stage, a dynamic self-feedback loop verifies outputs and feeds corrections back, shifting from generate-then-filter to generate-verify-refine.
The authors trained a 4B parameter model on 11K ToolLoop-generated examples and achieved 86.40% accuracy on the Berkeley Function Calling Leaderboard (BFCL) in non-reasoning mode. Removing BFCL-overlapping candidate functions still yielded 86.07%. On ACEBench, the model reached 72.1% overall accuracy using only 18.3% of baseline training data, showing strong generalization and data efficiency.