Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents
Wang Wei, Tiankai Yang, Samyadeep Basu, Hongjie Chen, Yue Zhao, Zhengzhong Tu, Xiyang Hu, Franck Dernoncourt and 2 others
Virginia Tech · University of Southern California · Adobe Research · Dolby Labs · Texas A&M University · Arizona State University
Published September 5, 2026 · on the feed September 14, 2026 · ▲ 7 · arXiv 2609.05824
The abstract
Large language model (LLM) agents increasingly rely on external skills, but routing user requests over large skill registries is difficult because many skills are functionally redundant while complex tasks often require complementary skill sets. Existing skill routers typically rank candidates independently by query relevance, which can waste context budget on redundant skills. We propose Diverse Skill Routing (DSR), a diversity-aware reranking framework that uses a Determinantal Point Process to balance relevance and non-redundancy. DSR introduces a query-residual diversity kernel that penalizes redundant skill overlap while reducing penalties caused only by shared query relevance. On the SkillRouter benchmark, DSR improves recall and full coverage over a strong pointwise reranking baseline, with larger gains on multi-skill queries. These results suggest that skill routing should be treated not only as relevance ranking, but also as complementary set selection.
The Brief
1:49A short podcast summary of the paper. Written and voiced by AI.
Transcript
Beyond Top-k Skill Retrieval. Diversity-Aware Skill Routing for LLM Agents. By Wang Wei and nine others, from Virginia Tech, the University of Southern California, and four other institutions. Published on the fifth of September. No code repository is linked to it.
Agents built on large language models increasingly call on external skills, and the registries holding those skills keep growing. Routing a user request across such a registry is hard for two reasons that pull against each other. Many skills are functionally redundant. And complex tasks often need a complementary set of skills rather than a single one. Existing skill routers typically rank each candidate independently by its relevance to the query. That can fill the context budget with several near-duplicates that all score well, while a skill the task also needs gets crowded out.
Diverse Skill Routing, or DSR, is a reranking framework that is aware of diversity. It uses a Determinantal Point Process to balance relevance against non-redundancy, so what it picks is a set of skills rather than a list of individually good ones.
The piece DSR adds is a query-residual diversity kernel. Two skills can look alike partly because both are relevant to the same query, and penalizing that kind of similarity would punish exactly the skills a router should keep. The kernel penalizes redundant overlap between skills while reducing the penalties that come only from their shared relevance to the query.
On the SkillRouter benchmark, DSR improves both recall and full coverage over a strong pointwise reranking baseline. The gains grow on multi-skill queries, which is where choosing complements rather than duplicates has the most room to matter.
The authors draw a conclusion about framing. Skill routing is not only relevance ranking. It is also the selection of a complementary set, where the value of adding a skill depends on what has already been chosen.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A shortlist of five parsers is not a toolkit. Start with that failure.
Your coding assistant can load skills from a registry of tens of thousands: parsers, cleaners, chart makers. Only a few fit in its context, so you must pick a shortlist for each request.
The road ahead
- 1 · The problem ~7 min When a router scores each skill on its own and keeps the top few, a request that needs several different skills gets a shortlist of near-duplicates.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Pick skills as a set with a determinantal point process, measuring redundancy only after removing what each skill shares with the request.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Retrieve candidates cheaply, score their relevance, strip query-aligned parts before measuring similarity, then greedily pick a relevant, non-redundant set.
- 6 · The experiment ~6 min You’ll guess: On the SkillRouter benchmark, DSR was compared with a strong pointwise reranking baseline. Where were its gains in recall and full coverage larger?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.