A library of relevant instructions still leaves an agent to work out which step depends on which. Ran Zhou and colleagues’ SPEM stores successful procedures alongside individual skills, preserving their order and prerequisites.
It tries a matching procedure first. When that is unavailable or insufficient, a graph neural network ranks a small set of skills using the current task and state. Successful executions become compact reusable procedures; failures update their reliability records.
The authors test 140 ALFWorld household episodes and 67 valid SkillsBench technical tasks. The technical tests share a 200-skill library and consistently exclude unresolved infrastructure failures.
- With GPT-5.2-Codex, technical-task reward rises from graph retrieval’s 37.50% to 38.38%. Average input falls from 1,162,375 to 980,029 tokens; execution time falls from 705.2 to 618.7 seconds.
- With MiniMax-M2.7, reward is 25.14%, below SkillDAG’s 27.09%. SPEM uses 26.3% fewer input tokens and 24.3% less execution time in that comparison.
- Household-task success reaches 87.1% with GPT-5.2-Codex and 70.7% with MiniMax. Template reuse is mainly examined in this more repetitive household setting.
These are author-reported results. The main tables supply no uncertainty estimates, making the stronger model’s small reward gain harder to interpret. The manuscript also refers to supplementary implementation details absent from the supplied PDF and HTML.
Caveat: time savings cover agent execution, including online skill selection. They exclude setup, dependency installation, container building and verification, so they are not total job-time savings.