Constraint-Consistent Skill Composition for Reliable Zero-Shot Task Generalization in LLM Agents

Addressing the frequent emergence of new tasks and long-tail demands in open environments for large language model agents, this paper proposes a task transfer and zero-shot generalization method for composable skill libraries. This transforms the agent’s capability organization from an end-to-end strategy into a problem of structured construction and combinatorial scheduling of reusable skill units. First, skills are uniformly encapsulated in a structured manner, explicitly describing interface parameter patterns, preconditions, posteffects, and semantic representations, enabling skills to possess searchable and verifiable invocation semantics across different tasks and execution contexts. Second, a shared semantic space for tasks and skills is constructed. Candidate skill retrieval and selection are achieved through joint scoring of semantic relevance and constraint consistency, reducing misselection and combinatorial instability caused by relying solely on text similarity. Third, task solving is modeled as a joint optimization process of skill sequence planning and state evolution. Precondition satisfiability constraints and state transition consistency checks are introduced during the planning process to improve the executability and controllability of the generated sequences. Finally, cross-task shared priors are introduced to constrain policy distribution deviations, enhancing stable transfer capabilities under changes in task descriptions and environmental differences. Comparative experiments based on public datasets verify the comprehensive advantages of the proposed method in terms of task success rate, tool invocation accuracy, constraint satisfaction, and average steps, demonstrating that the structured skill library and constraint consistency mechanism can effectively support zero-shot task adaptation and reliable execution of agents.