Towards Autonomous Mathematics Research

This paper introduces Aletheia, an autonomous AI research agent powered by advanced reasoning models and tool use that successfully generates, verifies, and revises mathematical proofs from Olympiad problems to PhD-level research, achieving milestones such as fully AI-generated papers and the autonomous solution of open problems while proposing new frameworks for quantifying AI autonomy and transparency.

The Gist

Imagine a world where mathematics is like a vast, dark forest. For centuries, human explorers (mathematicians) have been the only ones with the flashlights to find new paths, solve puzzles, and map the terrain. They are brilliant, but they are also tired, they get distracted, and they can only carry so much information in their heads.

This paper, titled "Towards Autonomous Mathematics Research," is a report from a team at Google DeepMind. They are introducing a new kind of explorer: an AI agent named Aletheia.

Here is the story of what they did, explained simply.

1. The Big Leap: From Video Games to Real Exploration

For a while, AI was like a champion video game player. It could solve the "International Mathematical Olympiad" (IMO) problems, which are like the hardest puzzles in a video game. These puzzles are tricky, but they have clear rules and a known ending.

The team asked: "Can this AI stop playing games and start doing real research?"

Real math research is different. It's not a video game level; it's like exploring a jungle where no one knows the map. You have to read thousands of old books, invent new tools, and sometimes you get lost. The AI had to learn to stop just "guessing" and start "thinking" like a human researcher.

2. Meet Aletheia: The AI Researcher

Aletheia isn't just a single brain; it's a team of three AI workers working together in a loop:

• The Generator: The dreamer. It tries to come up with a solution.
• The Verifier: The critic. It checks the dreamer's work and says, "Wait, that doesn't make sense," or "That citation is fake."
• The Reviser: The editor. It fixes the mistakes and tries again.

They keep looping until they find a solution that passes the critic's test. This is like a writer, an editor, and a fact-checker working in a room together, passing a paper back and forth until it's perfect.

3. The "Hallucination" Problem

One of the biggest hurdles was that AI loves to lie (or "hallucinate"). If you ask a human mathematician for a book reference, they might look it up. If you ask an AI, it might invent a book title that sounds real but doesn't exist.

To fix this, Aletheia was given internet access. It learned to use Google Search to check if the books and papers it cited actually exist. It's like giving the AI a library card and a search engine so it can't just make things up.

4. What Did Aletheia Actually Do?

The team tested Aletheia on three types of challenges:

• The "Solo" Mission (Level A): Aletheia solved a complex problem about "eigenweights" (a fancy math term for specific numbers in geometry) entirely on its own. No human touched the solution. It was like the AI writing a whole chapter of a textbook by itself.
• The "Team-Up" Mission (Level C): In other cases, humans and AI worked together. The AI would come up with a big, clever idea (like a map), and the human would fill in the detailed steps to prove it. It was like the AI being the architect and the human being the builder.
• The "Erdős" Challenge: A famous mathematician named Paul Erdős left behind 700 unsolved puzzles. The team let Aletheia loose on them.
  • The Good News: Aletheia solved 4 of them completely on its own.
  • The Reality Check: The team also found that many of the "solutions" the AI found were actually trivial (too simple) or had already been solved by humans decades ago, but nobody knew because the solutions were buried in old papers. This taught them that AI is great at finding things that are "hidden in plain sight," but it still struggles to find things that are truly "new" and "deep."

5. The "FirstProof" Test

To see how good Aletheia really was, they gave it a secret test called FirstProof. These were 10 math problems that real mathematicians had solved but hadn't published yet. It was a blind test.

• The Result: Aletheia solved 6 out of 10 problems correctly.
• The Comparison: Other top AI models (like GPT-5) could only solve 2 of them. Aletheia was the clear winner, proving that this new "team of workers" approach is much better than just a single smart AI.

6. The "Driver's License" for AI Math

The authors realized that people are getting confused. Some say AI has "solved math," while others say it's just a calculator. To fix this, they proposed a new rating system, similar to how we rate self-driving cars:

• Level 0: The AI does nothing useful.
• Level 1: The AI helps with small calculations (like a calculator).
• Level 2: The AI and Human work together as partners (The "Co-Pilot").
• Level 3: The AI does almost everything, and the human just checks the work.
• Level 4: The AI is a fully autonomous genius (We aren't quite here yet).

They classified their own results on this scale. Most of their "solo" work was Level 2 or 3, meaning it was impressive, but not yet a "Landmark Breakthrough" that changes the world like the discovery of gravity.

7. The Bottom Line

The paper concludes with a very important message: AI is not replacing mathematicians; it is becoming their super-powerful tool.

• Humans are still needed for the "big picture," for knowing what questions are worth asking, and for taking responsibility for the final answer.
• AI is amazing at reading millions of books, remembering every formula, and trying thousands of ideas in seconds.

Think of it this way: If mathematics is a giant library, humans are the librarians who decide which books to read and what stories to tell. Aletheia is a robot that can read every book in the library in a minute and find the exact page you need. Together, they can discover things neither could find alone.

In short: The AI is no longer just a student taking a test; it has graduated and is now an intern in the research lab. It's making mistakes, it's sometimes boring, but it's starting to do real work. And that is a huge step forward.

Technical Summary

Here is a detailed technical summary of the paper "Towards Autonomous Mathematics Research" by Tony Feng et al. (Google DeepMind), dated March 9, 2026.

1. Problem Statement

The paper addresses the transition of Artificial Intelligence from competition-level mathematics (e.g., International Mathematical Olympiad, IMO) to professional research-level mathematics. While AI models have recently achieved gold-medal performance on IMO problems, research mathematics presents distinct challenges:

• Scope: Research problems require synthesizing vast, specialized literature rather than relying on standard high-school curricula.
• Complexity: Solutions often span dozens of pages and involve deep, non-standard techniques.
• Hallucination: Foundation models often suffer from "superficial understanding" and hallucinate citations or results when dealing with niche topics.
• Evaluation Gap: There is a lack of standardized metrics to quantify the novelty and autonomy of AI-generated mathematical results, leading to misinformation in public discourse.

The core objective is to develop an autonomous agent capable of generating, verifying, and revising mathematical proofs end-to-end in natural language, and to establish a framework for evaluating these contributions.

2. Methodology: The Aletheia Agent

The authors introduce Aletheia, a math research agent built upon Gemini Deep Think. Unlike previous systems (e.g., AlphaGeometry) that used formal languages, Aletheia operates entirely in natural language.

Core Architecture

Aletheia employs a multi-agent orchestration loop consisting of three sub-agents:

1. Generator: Produces initial solution drafts and proofs.
2. Verifier: Critically analyzes the Generator's output for logical flaws, hallucinations, or citation errors.
3. Reviser: Iteratively refines the solution based on the Verifier's feedback.
   This loop continues until a solution is verified or a compute limit is reached.

Key Technical Innovations

• Inference-Time Scaling Law: The paper demonstrates that increasing compute at inference time (via parallel thinking and iterative refinement) significantly boosts accuracy. This scaling law holds for both Olympiad-level and PhD-level problems, though accuracy drops for the latter without additional interventions.
• Tool Integration: Aletheia heavily utilizes Google Search and web browsing to navigate the literature. This drastically reduces "spurious citation" hallucinations (fabricating papers/authors), though it shifts errors to subtler misinterpretations of existing literature.
• Decoupling Thinking from Output: The system separates the model's internal reasoning trace from its final output. This allows the model to recognize flaws in its own logic that it might otherwise "bluff" through during generation.
• Human-AI Interaction Cards: To ensure transparency, the authors propose a standardized documentation format (similar to "Model Cards") detailing the prompts, AI outputs, and human interventions for every collaborative result.

3. Key Contributions and Results

The paper details four major milestones achieved by Aletheia:

A. Fully Autonomous Research (Level A)

• Paper: Eigenweights for Arithmetic Hirzebruch Proportionality (Feng26).
• Achievement: Aletheia calculated structure constants (eigenweights) in arithmetic geometry without any human intervention. It utilized techniques from algebraic combinatorics unfamiliar to the human authors of the original context, resolving open questions left in a prior paper (FYZ26).
• Significance: This is the first instance of a research-grade paper generated entirely by AI, demonstrating the reliability of the agent for specific, high-complexity tasks.

B. AI-Guided Collaboration (Level C)

• Paper: Lower bounds for multivariate independence polynomials (LeeSeo26).
• Achievement: Aletheia provided the high-level strategic roadmap (e.g., suggesting the use of "dual sets") for a complex proof involving interacting particle systems. Human authors executed the rigorous details.
• Significance: This reverses the typical workflow; instead of humans breaking down problems for AI, the AI provided the "big picture" vision, with humans handling the execution.

C. Systematic Evaluation on Erdős Problems

• Study: A semi-autonomous evaluation on 700 open problems from Bloom's Erdős Conjectures database.
• Results:
  • Out of 700 prompts, Aletheia generated 212 potentially correct responses.
  • After human expert vetting, 13 solutions were deemed "meaningfully correct" (addressing the intended problem).
  • 4 Autonomous Resolutions: Aletheia found the first correct solutions to Erdős-652, 654, 1040, and 1051.
  • Rediscovery: Several "solutions" were found to be rediscoveries of existing literature or trivial interpretations of the problem.
• Insight: Many "Open" Erdős problems are unresolved due to obscurity rather than difficulty. AI excels at scanning vast literature to find these "hidden" solutions.

D. FirstProof Benchmark Performance

• Benchmark: FirstProof, a collection of 10 research-level lemmas curated by mathematicians to test AI (solutions were not public).
• Results: Aletheia successfully solved 6 out of 10 problems (P2, P5, P7, P8, P9, P10).
  • Problem 7: Solved correctly; considered publication-grade.
  • Comparison: Aletheia outperformed other systems (including OpenAI's GPT-5.2 Pro and internal Deep Think baselines) on this specific set, particularly on problems requiring deep literature synthesis.

4. Proposed Taxonomy: Levels of Autonomy and Significance

To address the evaluation gap, the authors propose a two-axis taxonomy analogous to SAE levels for autonomous driving:

Axis 1: Level of Autonomy

• Level H (Primarily Human): AI provides minor assistance (search, calculations).
• Level C (Human-AI Collaboration): Essential contributions from both parties.
• Level A (Essentially Autonomous): Core mathematical content is fully AI-generated.

Axis 2: Mathematical Significance

• Level 0 (Negligible): Olympiad/PhD exercise level.
• Level 1 (Minor Novelty): New result, but not research-paper standard.
• Level 2 (Publication Grade): Publishable in a reputable journal.
• Level 3 (Major Advance): Top-tier journal level.
• Level 4 (Landmark Breakthrough): Generational impact (e.g., Fermat's Last Theorem).

Classification of Results: The authors classify their autonomous paper (Feng26) as Level A2 (Autonomous, Publication Grade), while noting that none of their results yet reach Level 3 or 4.

5. Significance and Limitations

Significance:

• Proof of Concept: Demonstrates that AI can autonomously generate publishable mathematical research, moving beyond solving pre-defined contest problems.
• Efficiency: AI can rapidly resolve "obscure" open problems that have eluded humans due to lack of attention rather than difficulty.
• Transparency: The proposal of "Human-AI Interaction Cards" and a standardized taxonomy sets a precedent for responsible reporting in AI-assisted science.

Limitations & Weaknesses:

• Accuracy: Success rates are low. In the Erdős study, only 6.5% of technically correct responses were "meaningfully correct."
• Hallucination: Despite tool use, models still misinterpret citations or fabricate subtle details.
• Creativity: Current AI excels at technical manipulation and retrieval but lacks the "genuine creativity" or intuition to formulate entirely new research directions or deep conjectures.
• Scope: Current results are relatively brief and elementary compared to the depth of modern human research papers (which often exceed 100 pages).

Conclusion

The paper concludes that AI will serve as a powerful enhancement to human mathematicians rather than a replacement. While AI currently struggles with the deep, creative intuition required for landmark breakthroughs, its ability to synthesize vast knowledge and perform exhaustive searches makes it an invaluable tool for solving specific, high-complexity problems and uncovering overlooked solutions in the mathematical literature. The authors emphasize that human accountability remains essential for the final verification and exposition of mathematical truth.