AI Agents Formalize Type Annotation Proofs in Isabelle with Human Guidance
A recent study investigates both complete and minimal type annotations for rank-one polymorphic λ-calculus terms within Isabelle, building upon earlier research conducted by Smolka and Blanchette et al. This work delivers a metatheoretical framework, complete with formal specifications and proofs, all formalized in Isabelle/HOL. The research contrasts workflows led by humans with those driven by AI: a human and an AI agent powered by a large language model (LLM) each generate proofs manually, with the AI subsequently autoformalizing them in Isabelle. Additional human-guided AI enhancements aim to refine and generalize the outcomes, concentrating on maintaining term meanings during reparsing and type inference. This study is recorded on arXiv, which also supports community projects through arXivLabs, promoting openness, community engagement, and user data privacy.
Key facts
- The study addresses type annotations for rank-one polymorphic λ-calculus terms in Isabelle.
- It builds on prior work by Smolka and Blanchette et al.
- A metatheoretical account includes full formal specifications and proofs.
- Formalization is conducted in Isabelle/HOL.
- Experiments involve human and LLM-powered AI agents producing pen-and-paper proofs.
- The AI agent autoformalizes proofs in Isabelle with human-hinted interventions.
- The research aims to preserve term meaning under reparsing and type inference.
- The work is published on arXiv with tools like BibTeX and Semantic Scholar.
Entities
Institutions
- arXiv
- arXivLabs
- Semantic Scholar