AI Navigate

TaoBench: Do Automated Theorem Prover LLMs Generalize Beyond MathLib?

arXiv cs.AI / 3/16/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

Key Points

  • TaoBench is introduced as an undergraduate-level benchmark derived from Terence Tao's Analysis I that formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, and includes both from-scratch and MathLib constructions.
  • The authors build an agentic pipeline to automatically extract a compilable, self-contained local environment for each problem and translate every problem into Mathlib to create paired TaoBench–Mathlib statements for direct comparison.
  • On standard MathLib problems, ATP models perform capably, but there is an average ~26% performance drop on the definitionally distinct TaoBench formulation, indicating that limited generalization across definitional frameworks—not task difficulty—is the main bottleneck.
  • TaoBench highlights a gap between benchmark performance and real-world applicability in research mathematics and provides a concrete foundation for developing provers better aligned with exploratory mathematical work.

Abstract

Automated theorem proving (ATP) benchmarks largely consist of problems formalized in MathLib, so current ATP training and evaluation are heavily biased toward MathLib's definitional framework. However, frontier mathematics is often exploratory and prototype-heavy, relying on bespoke constructions that deviate from standard libraries. In this work, we evaluate the robustness of current ATP systems when applied to a novel definitional framework, specifically examining the performance gap between standard library problems and bespoke mathematical constructions. We introduce TaoBench, an undergraduate-level benchmark derived from Terence Tao's Analysis I, which formalizes analysis by constructing core mathematical concepts from scratch, without relying on standard Mathlib definitions, as well as by mixing from-scratch and MathLib constructions. For fair evaluation, we build an agentic pipeline that automatically extracts a compilable, self-contained local environment for each problem. To isolate the effect of definitional frameworks, we additionally translate every problem into a mathematically equivalent Mathlib formulation, yielding paired TaoBench-Mathlib statements for direct comparison. While state-of-the-art ATP models perform capably within the MathLib framework, performance drops by an average of roughly 26% on the definitionally equivalent Tao formulation. This indicates that the main bottleneck is limited generalization across definitional frameworks rather than task difficulty. TaoBench thus highlights a gap between benchmark performance and applicability, and provides a concrete foundation for developing and testing provers better aligned with research mathematics.