Unified Operational Formalism for LLM-based Theorem-proving Systems

Published: 02 Mar 2026, Last Modified: 11 Mar 2026ICLR 2026 Workshop VerifAI-2EveryoneRevisionsBibTeXCC BY 4.0
Track: tiny / short paper (up to 4 pages)
Keywords: theorem proving, llm, formalism, framework
Abstract: LLM-based theorem provers differ widely in how they organise interaction between language models and interactive theorem provers, ranging from whole-proof generation to tactic-level and multi-stage pipelines. We propose a unified operational framework that makes this interaction structure explicit by modelling proof search as transitions over a joint formal and informal state and by introducing orchestration as a first-class abstraction that controls how tools such as language models, retrieval components, and provers are scheduled and coordinated. Within this framework, existing systems such as Baldur, COPRA, DSP, POETRY, etc, can be expressed uniformly as different orchestration strategies, enabling principled comparison of interaction patterns, rapid prototyping of new strategies, and backend-agnostic evaluation and reuse across provers, libraries, and models.
Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.
Submission Number: 58
Loading