Search papers, labs, and topics across Lattice.
This paper addresses the bottleneck of skill selection in large-scale LLM agent skill marketplaces by proposing SkillSeek, a two-stage information retrieval system. SkillSeek leverages a standard IR pipeline (bi-encoder followed by cross-encoder) to achieve retrieval parity with expensive LLM-mediated loops at significantly reduced computational cost. The key result demonstrates that this efficient IR approach can match or exceed LLM-based performance on the SkillsBench benchmark, drastically lowering per-trial costs.
Standard information retrieval methods can match LLM-mediated skill selection for agents, slashing per-task costs from $51 to $27.
Anthropic's Agent Skills package reusable procedural know-how for an LLM agent into SKILL.md directories, and open-source aggregations have grown past 230,000 skills, making selection rather than authoring the bottleneck. The standing answer in the literature outsources selection to the agent itself: an LLM-mediated retrieval loop that rewrites queries and refines candidates inside the agent's decision loop, paying LLM tokens on every task. We present SkillSeek, an open-source two-stage skill retriever built from the standard IR recipe (a BGE-base bi-encoder feeding a small cross-encoder, exposed over MCP). Across a $4 \times 11$ grid of pool, backbone, and method on the 89-task SkillsBench benchmark, SkillSeek reaches observed parity with the LLM-mediated loop of Liu et al. at essentially no extra cost: plain bm25 alone records a pass rate at or above their refined loop on three of four settings, and a small cross-encoder covers the remaining difference on the fourth. A first-stage recall ceiling explains the pattern, and total per-trial spend drops from USD 51.30 to USD 27.54 (within fifty cents of the no-skill baseline). Under the SkillsBench tasks and OpenHands harness we tested, this positions the standard IR recipe as a strong default for agent-skill retrieval, with LLM-mediated alternatives a natural fit for cases where deterministic methods fall short.
