Xinyu Zhao, Zhen Tan, Vaishnav Tadiparthi, Nakul Agarwal, Kwonjoon Lee, Ehsan Moradi Pari, Hossein Nourkhiz Mahjoub, Tianlong Chen
This work frames the skill selection problem for LLM agents as a structured composition task and proposes SkillComposer, a generative model that jointly predicts the subset, count, and order of skills to activate.
As skill libraries for LLM agents grow, selecting the optimal composition of skills—which skills to use, how many, and in what order—becomes a central bottleneck. Existing approaches, which either expose the entire library or use retrieval, fail to model the joint, structural nature of this decision.
The authors formalize the problem as task-conditioned skill sequence prediction. They propose SkillComposer, which uses a constrained autoregressive decoder over skill identifiers. This allows the model to jointly determine the activated subset, count, and execution order in a single decoding pass, naturally capturing dependencies between skills.
SkillComposer was trained on task-composition pairs from a human-curated skill library and evaluated on SkillsBench. On GPT-5.2-Codex and Gemini-3-Pro-Preview, it improved pass rates by +23.1pp and +18.2pp over a no-skill baseline, surpassing top-3 retrieval methods and matching the gold-skill retrieval upper bound at a lower prompt-token cost.