Existing methods for evolving LLM agent skill libraries follow a linear path, sequentially updating the latest version. This often leads to local optima and prevents exploring alternative improvements. SkillVine addresses this by treating as a graph search problem, using a branching exploration strategy that maintains a version graph.
SkillVine employs a trunk-branch collaborative mechanism: the trunk performs incremental updates on each training batch, while branches sprout from historical versions to explore alternative paths. It uses a complementarity-aware node selector to pick promising starting points for branches and an adaptive-granularity update rule to mitigate harmful updates.
Evaluated on five benchmarks with two LLMs, SkillVine outperformed six baselines in nine out of ten model-benchmark combinations. Crucially, every best-performing skill library was discovered via a branch, demonstrating the value of branching exploration over linear evolution.