Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks

Foundation model-driven agents often struggle with long-horizon planning due to the transient nature of purely prompting-based reasoning. While existing skill induction methods mitigate this by distilling experience into state-blind parameterized scripts, they fail to capture the conditional logic required for robust execution in dynamic environments. In this paper, we propose Neuro-Symbolic Skill Induction (NSI), a framework that lifts interaction traces into modular, logic-grounded programs. By synthesizing explicit control flows and dynamic variable binding, NSI empowers agents to discover when and why to act. This paradigm enables the efficient generalization, allowing agents to induce skills from few-shot examples and flexibly adapt to unseen goals. Experiments on a series of agentic tasks demonstrate that NSI consistently outperforms state-of-the-art baselines, empowering agents to self-evolve into architects of logic-grounded skills. Project Page: https://sh-jj.github.io/NSI.

AI agents are increasingly asked to complete multi-step tasks, such as finding items in a home, shopping online, or crafting objects in a game. Today’s agents often fail because they either reason from scratch at every step or repeat fixed scripts that break when the situation changes. We introduce NSI, a method that turns a few successful examples into reusable skills with built-in checks and choices. Instead of simply remembering “do these actions in this order,” the agent learns rules such as when to search, when to pick an object, and what to do if the expected object is missing. When a skill fails, NSI uses the failure feedback to repair and expand the skill for future tasks. Across household, web-shopping, and crafting environments, NSI helps agents complete longer tasks more reliably than prior methods. This work is a step toward AI agents that can learn from experience and adapt to new situations.