Title: Simply Input in Isabelle: AI Agents for Drafting, Formalizing, and Generalizing via Human Guidance

Abstract: Ensuring that printed terms retain their semantic integrity during reparsing and type inference requires the inclusion of type annotations. This paper investigates the challenge of determining complete yet minimal type annotations for rank-one polymorphic $\lambda$-calculus terms, a system central to Isabelle. Extending previous research by Smolka, Blanchette, and colleagues, we provide a rigorous metatheoretical framework for this problem, complete with formal specifications and proofs, all implemented within Isabelle/HOL. Our study comprises a series of experiments comparing formalization workflows driven by humans and AI. In these experiments, both a human expert and an LLM-powered AI agent independently generated pen-and-paper proofs. The AI agent then autoformalized these proofs into Isabelle, with subsequent human-hinted interventions allowing the AI to refine and generalize the resulting development.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC