Terence Tao’s use of LLMs reveals the limits—and real value—of AI in mathematical research
Terence Tao’s approach to large language models demonstrates why expecting AI to solve open problems like the Riemann Hypothesis is unrealistic. The technology’s strength lies not in generating proofs but in transforming how researchers navigate the sheer volume of modern mathematical literature.
Tao avoids treating AI as a substitute for human reasoning. Instead, he leverages it to process information at scale. At the frontier of mathematics, the accumulation of papers, theorems, and auxiliary results creates a bottleneck. His workflow relies on LLMs to distill dense technical passages, effectively accelerating his reading speed without replacing his own critical analysis.
A structured AI workflow for advanced research
To integrate LLMs meaningfully, researchers must rethink their prompts. Tao’s methods reflect this shift:
- Summarization over generation: Instead of demanding proofs, he directs models to extract key logical chains—such as "Outline the essential steps in this argument" or "Map the relationships between these two lemmas." The goal is to offload the labor of parsing papers, not to offload the work of proof itself.
- Code for verification: For computational mathematics, AI-generated scripts in Python or Lean serve as a sanity check. The utility here stems from the model’s ability to translate conjectures into executable code, not from its understanding of the underlying mathematics.
- Clarifying dense notation: LLMs excel as interactive "rubber ducks" for opaque definitions. When a paper’s phrasing obscures meaning, a prompt like "Rephrase this concept in plain terms" can reveal structure before revisiting formal notation.
Why LLMs struggle with pure mathematical reasoning
The technology’s probabilistic nature makes it unreliable for high-stakes logic. Attempting to use an LLM without formal verification risks hallucinations—outputs that appear plausible but lack rigor. Specific pitfalls include:
- Logical gaps: Models often omit critical steps, assuming connections that do not hold under scrutiny.
- Notational misalignment: In specialized fields, LLMs default to familiar but incorrect interpretations of symbols, guided by training data rather than context.
- Lack of formal certainty: An LLM does not validate proofs; it predicts the next likely token, offering no guarantee of correctness.
The practical takeaway is clear: AI should not replace human judgment but should automate the repetitive, time-consuming aspects of research. Use it to organize references, extract insights, and generate testable code—but never delegate the drafting of a proof. The most effective deployment treats LLMs as tools for efficiency, not as collaborators in discovery.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
LaTeX debugging is a lifesaver. Which tool handles complex mathematical formatting the best? I've been observing how the world's top mathematicians are weaving AI into their everyday routines, and a recent deep dive into Terence Tao's methods serves as a valuable reality check for anyone expecting an LLM to crack the Riemann Hypothesis overnight. The gap between using AI to produce a proof and employing it as a sophisticated, high-speed assistant for managing the overwhelming scale of modern mathematical literature is enormous. Tao doesn't treat AI as an oracle. Instead, he views it as a means to handle information density. When you're operating at the cutting edge of mathematics, the volume of existing papers, theorems, and lemmas becomes staggering. He relies on these models to sift through complex documentation and condense specific technical sections, effectively multiplying his reading efficiency. The practical AI workflow for high-level research To move past basic chat prompts and harness an LLM agent for serious research, you need to shift your mindset. Here's how a top-tier researcher approaches the technology: 1. Information Synthesis over Generation: Instead of asking "Prove this theorem," the focus in prompt engineering should be on "Summarize the core logical steps of this specific argument" or "Identify the connections between these two disparate lemmas." The aim is to lighten the cognitive burden of scanning a paper. 2. Code-Assisted Verification: For mathematicians tackling computational aspects, having AI write small scripts to test conjectures proves highly effective. It's less about the AI grasping the math itself and more about its ability to automate tedious calculations or generate code for verification, such as "Write a Python script to compute the first 100 terms of this sequence and look for patterns." This concrete step, writing scripts for verification, is a practical way to leverage AI without overestimating its theoretical capabilities.
I’m curious whether he’s leveraging Lean for formal proofs or using it more as a tool to extract and synthesize key logical structures—like how top researchers often ask AI to distill the core arguments from dense papers rather than expecting it to produce full proofs outright. The distinction between AI as a proof engine and a high-speed assistant for parsing existing work is critical, especially in fields where the sheer volume of prior literature demands efficiency over direct generation.
He's likely using it for Python scripts too. Anyone else use LLMs for rapid prototyping? I've noticed how top mathematicians are integrating AI into their workflows, and Terence Tao's approach offers a great example. Instead of asking "Prove this theorem," he focuses on prompts like "Summarize the core logical steps of this specific argument" to manage the overwhelming volume of mathematical literature. This approach helps lighten the cognitive load when scanning papers.