Andrii Bidochko

• Updated: July 14, 2026
• 7 min read

From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier

Direct Answer

The paper From Solvers to Research: Large Language Model‑Driven Formal Mathematics at the Research Frontier proposes a paradigm shift: instead of building LLM‑powered theorem provers that solve pre‑specified problems, it outlines a roadmap for “research agents” that can autonomously explore, conjecture, and formally verify new mathematics at the frontier of human knowledge. This matters because it moves AI from a supportive tool to a genuine collaborator capable of extending the boundaries of mathematical discovery.

Background: Why This Problem Is Hard

Mathematics has long been the proving ground for rigorous reasoning. In recent years, AI4Math—AI for Mathematics—has made impressive strides, especially with large language models (LLMs) that can translate informal statements into formal languages used by Interactive Theorem Provers (ITPs) such as Lean, Coq, and Isabelle. Yet three intertwined challenges keep current systems from becoming true research partners:

• Open‑ended problem definition: Frontier research rarely comes with a clean, well‑posed theorem statement. Researchers must first identify a promising conjecture, often after exploring a landscape of definitions, examples, and counter‑examples.
• Multi‑layered abstraction: Advanced mathematics builds on deep hierarchies of concepts (e.g., category theory, homotopy type theory). Existing datasets capture only shallow fragments, making it difficult for models to navigate the full abstraction stack.
• Toolchain fragmentation: Formal proof assistants, symbolic algebra systems, and data repositories each speak different “languages.” Orchestrating them into a coherent workflow demands sophisticated meta‑reasoning that current solvers lack.

Because of these gaps, today’s LLM‑driven provers excel at “solver” tasks—verifying a supplied lemma or filling a missing proof step—but they stumble when asked to generate novel conjectures, design experiments, or iteratively refine definitions. The bottleneck is not model size; it is the absence of an architecture that treats research as a dynamic, exploratory process.

What the Researchers Propose

The authors advocate a new framework called Research‑Oriented Formal Mathematics (ROFM). At a high level, ROFM consists of three interlocking components:

1. Auto‑formalization Engine: An LLM‑based front‑end that ingests informal mathematical text (papers, lecture notes, informal sketches) and produces a structured, machine‑readable representation in an ITP language.
2. Exploratory Reasoning Agent: A goal‑driven planner that can propose conjectures, generate lemmas, and design proof strategies. It treats the formal knowledge base as a searchable graph, allowing it to traverse definitions, theorems, and examples.
3. Tool‑Orchestration Layer: A middleware that dynamically invokes external utilities—symbolic algebra systems, numeric simulators, and even web‑scrapers—to gather evidence, test hypotheses, and feed results back into the formal environment.

Crucially, the framework emphasizes research loops rather than single‑shot solving: the agent iteratively refines its conjectures based on counter‑examples, updates definitions, and re‑formalizes new insights. This mirrors how human mathematicians work, turning the system into a genuine research collaborator rather than a static solver.

How It Works in Practice

The ROFM workflow can be visualized as a cyclical pipeline:

1. Ingestion: The auto‑formalization engine consumes a corpus of informal material (e.g., a recent preprint). Using prompt‑engineering and few‑shot examples, it produces a provisional formal library in Lean.
2. Graph Construction: The formal library is transformed into a relational graph where nodes represent definitions, theorems, and proof objects, while edges capture dependencies and logical flow.
3. Exploration: The reasoning agent queries the graph to identify “gaps” – areas where a definition lacks a natural theorem, or where existing lemmas suggest a pattern. It then generates candidate conjectures expressed in the formal language.
4. Evidence Gathering: The orchestration layer dispatches tasks to auxiliary tools. For instance, a symbolic algebra system may test the conjecture on random instances, while a numeric simulator evaluates its behavior in a specific domain.
5. Feedback Loop: Results from the tools are fed back to the agent. Positive evidence strengthens the conjecture; counter‑examples trigger refinement of definitions or the generation of auxiliary lemmas.
6. Proof Synthesis: Once a conjecture survives empirical scrutiny, the agent attempts a formal proof using existing tactics or by synthesizing new tactics via LLM‑guided search.
7. Publication: Successful proofs are exported as human‑readable LaTeX alongside the formal artifact, ready for peer review.

What distinguishes this approach from prior solvers is the explicit separation of “exploration” and “verification.” Existing systems typically receive a fully formed theorem and focus solely on verification. ROFM, by contrast, endows the agent with the autonomy to pose its own questions, gather data, and iteratively improve its formal knowledge base.

Evaluation & Results

The authors evaluated ROFM on three benchmark tracks designed to stress different aspects of research capability:

• Conjecture Generation: Using a curated set of open problems from the “Open Problem Garden,” the agent was tasked with proposing plausible conjectures. Success was measured by expert judgment on novelty and relevance.
• Proof Synthesis after Exploration: The system attempted to prove the generated conjectures after completing the evidence‑gathering loop.
• Cross‑Domain Transfer: The agent was trained on algebraic topology papers and then asked to explore problems in homological algebra, testing its ability to transfer abstract reasoning.

Key findings include:

1. The agent produced novel, non‑trivial conjectures in 68% of the test cases, a marked improvement over baseline LLMs that merely re‑phrased existing theorems.
2. When equipped with the full exploration loop, proof synthesis succeeded on 54% of the conjectures, compared to 22% when the agent received the conjecture directly.
3. Cross‑domain transfer demonstrated that the relational graph abstraction allowed the agent to reuse patterns from one mathematical subfield to another, achieving a 41% success rate on previously unseen domains.

These results suggest that the research‑oriented loop not only boosts the quantity of provable statements but also improves the quality of the generated mathematics, moving the system closer to genuine discovery.

Why This Matters for AI Systems and Agents

For practitioners building AI agents, ROFM offers a template for embedding rigorous reasoning into broader autonomous workflows. The separation of exploration, evidence gathering, and verification aligns with best practices in Workflow automation studio, where complex pipelines are decomposed into reusable micro‑services. By treating formal mathematics as a graph‑structured knowledge base, developers can apply graph‑neural techniques, retrieval‑augmented generation, and reinforcement learning in a unified setting.

From a product perspective, the ability to generate and formally verify new theorems opens doors for:

• Automated verification of safety‑critical algorithms in aerospace and autonomous driving.
• Dynamic generation of cryptographic primitives that are provably secure.
• Intelligent tutoring systems that can create novel problem sets on the fly, adapting to student progress.

Moreover, the orchestration layer’s plug‑and‑play design dovetails with existing OpenAI ChatGPT integration and Chroma DB integration, enabling agents to store and retrieve formal artifacts at scale. This synergy reduces engineering overhead and accelerates the path from research prototype to production‑grade AI service.

What Comes Next

While ROFM marks a significant step forward, the authors acknowledge several limitations that shape the next research agenda:

• Scalability of Formal Libraries: Current auto‑formalization struggles with large, heterogeneous corpora. Future work must improve chunking strategies and hierarchical abstraction to keep the graph tractable.
• Human‑in‑the‑Loop Interfaces: Effective collaboration requires intuitive UI/UX for mathematicians to intervene, correct, or steer the agent. Integrating with platforms like the UBOS platform overview could provide the necessary collaborative workspace.
• Evaluation Standards: The community lacks benchmark suites that capture the open‑ended nature of mathematical research. Establishing shared datasets and leaderboards will be essential for measuring progress.
• Ethical Guardrails: Autonomous conjecture generation raises questions about attribution, intellectual property, and the potential for AI‑generated “fake” mathematics. Robust provenance tracking must be built into the pipeline.

Potential applications beyond pure mathematics are already emerging. For example, AI marketing agents could leverage formal reasoning to guarantee compliance with advertising regulations, while the Enterprise AI platform by UBOS might embed ROFM modules to audit algorithmic decisions in finance or healthcare.

In summary, the shift from solver to research agent redefines what AI can achieve in formal domains. By embracing iterative exploration, tool orchestration, and graph‑based knowledge representation, the community moves closer to AI systems that not only check proofs but also help write the next chapters of mathematics.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.