CaSKG is a retrieval framework that calibrates the reliability of a skill graph’s edges before using it to retrieve procedural knowledge for LLM agents, using counterfactual probes that remove, substitute, and reorder skill pairs to score which relations are trustworthy. Across six LLM backbones on the ALFWorld and ScienceWorld benchmarks, CaSKG achieves the highest task score in all twelve model-benchmark combinations tested, improving ALFWorld success from 80.0% to 86.8% over the prior Graph-of-Skills method. The authors have released the implementation as open source.
