Tag
This paper introduces ParamBench, a difficulty-graded benchmark for LLM tool-call parameter generation, and proposes probe-guided training methods (PBT and PGR) that improve exact-match accuracy from 19.7% to 59.6%.