Reported evidence
KOPA-Bench study → uses → Combined live validation, trajectory synthesis, and GRPO training
The KOPA-Bench study—not the benchmark itself—uses the combined workflow. S10 distinguishes the 145-task KOPA-Bench benchmark from EDGE, which validates tool connections against live APIs and synthesizes multi-step trajectories, and from the subsequent GRPO model training on the resulting data. The abstract reports that the trained 9B model nearly matches an untuned 27B model from the same family. The original article’s author notes that the abstract does not separately measure each component’s contribution, so the result cannot be attributed to any one component. S10 directly supports the reported workflow and comparison and serves as background for the author’s attribution caution.
Read the evidence
KOPA-Bench研究建立工具输出与另一工具输入之间的连接图,只保留调用实时API后成功的连接,再据此合成多步轨迹;研究随后使用相关数据和GRPO训练模型。[S10]Read the full research