SkillPyramid organizes agent skills into a hierarchy

SkillPyramid raised average reward 38% and cut execution steps 27.7% by organizing an agent's learned skills into tiers instead of one growing list, per the paper posted to arXiv. That's a real gap for agents that pile up routines over long sessions. Flat skill libraries only get slower to search as they grow.

Agent frameworks that build skill libraries, the approach Voyager introduced, typically just append every successful trajectory as a new entry. SkillPyramid replaces that with a hierarchy of composable tiers, plus a self-evolution loop that consolidates related skills as the agent keeps learning.

The numbers come from three text-based environments: ALFWorld, WebShop, and ScienceWorld. Composable tiers generalized across all three, according to the paper, while the flat approach didn't.

Rewriting how an agent stores skills is a bigger lift than tuning how it searches them. That's likely why adoption lags behind the numbers. The real test is whether these gains hold once a library holds hundreds of skills, something the paper doesn't cover.

It's timely because it's new, an arXiv paper testing hierarchy against a flat list, posted this week.