Large language model (LLM) agents increasingly depend on external skills, but routing user requests across large skill registries remains difficult. According to a new paper posted on Hugging Face, many registered skills are functionally redundant, while complex tasks often require complementary sets of skills rather than a single best match.
Existing skill routers typically rank candidates independently by query relevance. The paper argues this approach can waste the agent's context budget on redundant skills that add little beyond the top result.
To address this, the authors propose Diverse Skill Routing (DSR), a diversity-aware reranking framework built on a Determinantal Point Process (DPP) that balances relevance against non-redundancy. DSR introduces a query-residual diversity kernel that penalizes overlapping, redundant skills while reducing penalties that arise only from shared query relevance.
On the SkillRouter benchmark, the paper reports that DSR improves recall and full coverage over a strong pointwise reranking baseline, with larger gains on multi-skill queries. The authors conclude that skill routing should be treated not merely as relevance ranking but as complementary set selection.
The paper's accompanying discussion raises practical deployment questions, noting that a pairwise kernel over a registry of a few thousand skills runs on every query and represents a real latency cost. It also asks whether the query-residual kernel outperforms a greedy max-min diversity filter in production, given that greedy approaches are parallelizable and cacheable.