Hierarchical skill KB improves performance of weaker models

Short summary

SkillX automatically constructs hierarchical skill knowledge bases from execution traces, enabling smaller models like Qwen3-32B to gain ~10 points on benchmarks while reducing token budgets. The three-tiered hierarchy allows efficient skill retrieval without redundant exploration. Consider building a pilot skill library from existing logs to measure accuracy and token reduction.

• •SkillX improves smaller models by ~10 benchmark points through hierarchical skill extraction
• •Three-tiered design (strategic plans, functional skills, atomic skills) reduces redundant steps and context length
• •Actionable recommendation: pilot skill library from existing logs to measure accuracy and token reduction

Generated with AI, which can make mistakes.