CaSKG: Counterfactual-Causal Skill Graphs for Scalable Agent Skill Retrieval
CaSKG reports stronger agent skill retrieval by calibrating graph edges before the agent uses them.
The paper builds a directed skill graph offline from semantic, lexical, input/output, and structural signals, then tests skill-pair relations with counterfactual probes. It uses the resulting weighted graph for task-conditioned expansion without changing the downstream agent policy or interface. Across six LLM backbones on ALFWorld ID-140 and ScienceWorld U211, it says CaSKG led every model-benchmark combination tested. The authors report gains over Graph-of-Skills, including ScienceWorld macro-average score rising from 72.62 to 80.50 and ALFWorld success from 80.01% to 86.79%. HF Daily Papers' note
The paper builds a directed skill graph offline from semantic, lexical, input/output, and structural signals, then tests skill-pair relations with counterfactual probes. It uses the resulting weighted graph for task-conditioned expansion without changing the downstream agent policy or interface. Across six LLM backbones on ALFWorld ID-140 and ScienceWorld U211, it says CaSKG led every model-benchmark combination tested. The authors report gains over Graph-of-Skills, including ScienceWorld macro-average score rising from 72.62 to 80.50 and ALFWorld success from 80.01% to 86.79%. HF Daily Papers' note
score 4