Multi-Step Tool-Calling over Korean Open Public APIs: KOPA-Bench and EDGE
Original title:Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe
Stricter data-sovereignty mandates often require public institutions to run on-premise agents against live government APIs, where open-weight models frequently struggle with multi-step reasoning. To address this shortfall, researchers introduced KOPA-Bench—a suite of 145 real-world tasks—and EDGE, a data synthesis framework that maps API input-output compatibilities and retains only chains verified through live execution. Trained on these grounded trajectories using GRPO, a 9B model achieved tool-calling accuracy comparable to an untuned 27B counterpart across both local public endpoints and the standard BFCL benchmark.
Why it's worth reading
It provides a practical, execution-grounded data synthesis recipe that enables smaller open models to reliably navigate complex, real-world public APIs under data sovereignty constraints.