FormalScience: Using AI Agents to Automatically Formalize Scientific Reasoning into Lean-Verifiable Code

A New Breakthrough in Formal Verification of Scientific Reasoning

Transforming informal mathematical reasoning into formally verified code that computers can rigorously check has long been a major challenge in artificial intelligence. A recently published paper on arXiv (arXiv:2604.23002) introduces a novel framework called "FormalScience," designed to achieve autoformalization of mathematical reasoning in scientific domains through a scalable human-in-the-loop mechanism and AI agent-driven code generation, targeting the interactive theorem prover Lean as the verification language.

Why Is Formalization So Difficult in Scientific Domains?

While autoformalization has made some progress in pure mathematics, the difficulty increases dramatically when extending to natural sciences such as physics. These disciplines feature extensive domain-specific notation systems and computational tools — for example, Dirac notation in quantum mechanics, and vector calculus in electromagnetism and fluid dynamics. These symbolic systems often lack ready-made corresponding libraries in mathematical formalization toolchains, and existing large language models (LLMs) and AI agent methods struggle with cross-domain symbolic translation.

Specifically, autoformalization faces three key challenges:

• Symbolic Gap: There is a vast expressive disparity between informal representations in scientific domains (such as derivations in research papers) and formal languages like Lean.
• Missing Domain Knowledge: General-purpose LLMs lack deep understanding of the axiomatic systems and reasoning rules specific to particular scientific fields.
• Scalability Bottleneck: Traditional methods are often designed for a single domain and are difficult to generalize across multiple disciplines.

Core Design Philosophy of FormalScience

To address these issues, FormalScience proposes a "domain-agnostic" solution architecture. Its core approach encompasses the following layers:

Scalable Human-in-the-Loop Workflow

Rather than pursuing full automation, FormalScience adopts a pragmatic human-in-the-loop strategy. Domain experts can intervene at critical junctures to provide domain-specific guidance, correct semantic deviations during the formalization process, and verify the correctness of the final output. This design ensures formalization quality while significantly reducing the workload on experts.

AI Agent-Driven Code Generation

The framework employs agentic methods for automatic Lean code generation. Unlike simple single-turn LLM calls, the agent system can perform multi-round iterative reasoning: analyzing formalization targets, decomposing proof steps, and leveraging Lean compiler feedback for self-correction, thereby progressively converging on correct formal expressions.

Domain-Agnostic Architecture

Domain agnosticism is a major highlight of FormalScience. The framework is not tied to any specific discipline and can theoretically be adapted to physics, chemistry, biology, and other scientific domains — requiring only the integration of corresponding domain knowledge modules and symbolic mapping rules for extension.

Research Significance and Industry Impact

The value of this research extends far beyond technical innovation. Formal verification is considered the "gold standard" for ensuring the rigor of scientific reasoning, but its high labor costs have long limited large-scale adoption. The emergence of FormalScience promises far-reaching impact in the following areas:

• Accelerating Scientific Verification: Formalizing key derivation processes in fields like physics can help uncover implicit assumptions and potential errors, enhancing the reliability of scientific research.
• Advancing AI for Science: Providing formal verification infrastructure for AI-assisted scientific discovery, enabling rigorous examination of AI-generated scientific hypotheses.
• Bridging Community Gaps: Lowering the barrier for scientists to use formal tools, fostering collaboration between the mathematical formalization community and the natural sciences community.

Notably, this research also reflects an important trend in current LLM research — shifting from pursuing end-to-end full automation toward a more pragmatic human-AI collaboration paradigm. In tasks like formal verification, where correctness requirements are extremely high, combining human expert judgment with AI's efficient generation capabilities may be the optimal path under current technological conditions.

Future Outlook

As the Lean 4 ecosystem continues to mature and LLM reasoning capabilities keep improving, autoformalization in scientific domains is poised to enter a period of rapid development. The scalable human-AI collaboration framework proposed by FormalScience provides a valuable technical blueprint for this direction. Going forward, key topics worthy of continued attention will include how to further reduce the frequency of human intervention, how to build formalized knowledge bases covering more disciplines, and how to seamlessly integrate formal verification into scientific research workflows.