ToolCompass post-training groups tool calls by function so agents explore unseen tools without wasteful trials
arXiv·low signal
arXiv 2609.25678 models each tool-function class as a von Mises-Fisher distribution during post-training. Experience with seen tools then transfers to functionally similar unseen ones, and the agent skips unrelated alternatives. It needs no ground-truth call traces and adds no inference overhead. The authors report consistent out-of-distribution gains on AppWorld and FTRL under GRPO, RFT and DMPO.