How AI Agents Solve World-Class Math Problems? DeepMind Aletheia Revealed

While most are still amazed by AI's text generation capabilities, Google DeepMind has quietly propelled AI Agents to a new height. On March 1st, its latest creation, the Aletheia AI Agent, independently solved six out of ten world-class mathematical problems in the “FirstProof Challenge,” an achievement that directly pushes AI to the forefront of PhD-level research. This is not just a technological leap but a strong validation of my steadfast vision of “one person, one Agent team,” heralding boundless possibilities for individuals to expand business boundaries with AI.

Aletheia's Breakthrough: PhD-Level Reasoning from Code to Concept

The achievement of Google DeepMind's Aletheia AI Agent in the “FirstProof Challenge” held on March 1st marks an unprecedented level of AI in solving world-class mathematical problems. This challenge aims to evaluate AI systems' ability to solve research-level mathematical problems, with questions derived from mathematicians' actual research processes and answers being “zero-day math” — not searchable online. Aletheia independently solved six out of ten difficult problems (specifically problems 2, 5, 7, 8, 9, 10), and its generated proofs were independently verified by most academic mathematicians. This achievement is not just about solving difficult problems; it demonstrates AI's complete reasoning chain, from problem identification and solution conceptualization to logic verification, reaching a complexity previously only achievable by human PhD researchers.

Notably, when faced with the four unsolved problems, Aletheia chose not to output any results rather than “guessing,” which reflects its design's high priority on mathematical accuracy and effective suppression of “hallucinations.” As Thang Luong, Google DeepMind's Lead Scientist and Head of the Super Reasoning team, stated, this achievement is even more valuable than AI reaching International Mathematical Olympiad (IMO) gold medal level last year.

AI Agents are no longer mere executors; they are becoming explorers with deep thinking capabilities.

Gemini 3 DeepThink: Aletheia's Thinking Engine

The core of Aletheia's groundbreaking achievement lies in its powerful Gemini 3 DeepThink model. So, what exactly is Gemini 3 DeepThink?

Gemini 3 DeepThink is an advanced reasoning capability built into the Google Gemini 3 Pro model, enabling extended, multi-step logical processing to solve complex problems. Unlike standard reasoning that provides instant responses, DeepThink activates a more deliberate reasoning process. It breaks down challenges into structured steps, evaluates multiple solution paths, and validates its logic before drawing conclusions. This capability is designed specifically for tasks requiring deep analysis rather than quick answers. DeepThink utilizes Gemini 3 Pro's full 1 million token context window, equivalent to 700,000 to 800,000 words, or approximately 1500 to 2000 pages of text. It excels in broad domains such as mathematics, science, and engineering, and has surpassed Gemini 3 Pro's performance in benchmarks like “The Last Exam for Humans” and GPQA Diamond.

“Self-filtering” Mechanism: A Panacea for Suppressing AI Hallucinations?

Another key capability demonstrated by Aletheia in the “FirstProof Challenge” is its “Self-filtering” mechanism, which effectively suppresses the occurrence of AI hallucinations. In the field of AI, “hallucination” refers to a model generating information that appears plausible but is actually inaccurate or fabricated. This is fatal for mathematical research requiring high precision and credibility.

The working principle of the Self-filtering mechanism can be understood as the AI Agent continuously performing internal verification, cross-referencing, and multi-step logical validation during the solution generation process. When Aletheia chooses not to provide an answer for certain problems, it is precisely this mechanism at play – it does not blindly “guess” but rather recognizes that reliable conclusions cannot currently be drawn, thereby avoiding the generation of incorrect “hallucinations.” This is remarkably similar to human scientists maintaining caution regarding uncertainty in research, avoiding unverified conclusions. For example, the Google DeepMind team previously reported that earlier versions of Aletheia sometimes “hallucinated” non-existent constraints to make a proof valid. However, through continuous iterative optimization with DeepThink and the self-filtering mechanism, Aletheia has become more reliable in avoiding such errors. This Reflect-Correct Loop is a crucial capability for agents to achieve self-optimization.

The introduction of the 'self-filtering' mechanism marks a crucial step for AI on the path to pursuing truth and reducing hallucinations, an essential route to true intelligence.

A New Paradigm for AI Agents: From Tools to Collaborators

The Aletheia case is not merely a technical victory; it further signals that AI Agents are transforming from simple tools into partners capable of deep intellectual collaboration with humans. In the past, we might have viewed AI as an automated tool for executing specific tasks. But Aletheia's ability to solve world-class mathematical problems indicates it can now engage in high-level abstract thinking, logical reasoning, and problem-solving – capabilities once considered exclusively human.

This transformation from “tools” to “collaborators” has profound implications for XB Lab's commitment to the “one person, one Agent team” philosophy. This means individuals are no longer limited by their personal knowledge and capability boundaries but can leverage AI Agents with “PhD-level” intelligence, like Aletheia, to explore broader commercial and innovative domains.

Rotifer Protocol and the Evolutionary Vision for AI Agents

XB Lab's Rotifer Protocol (AI Agent) is precisely dedicated to exploring how AI agents can become smarter through autonomous evolution. Rotifer Protocol enables AI agents to autonomously evolve and become smarter through a genetic competition framework. Aletheia's achievements, especially its DeepThink and “self-filtering” mechanism, provide valuable reference and inspiration for Rotifer Protocol. We believe that AI Agents' self-learning and self-optimization capabilities are crucial for their path to true intelligence. Rotifer Protocol is currently in TRL 5 internal testing, and we are actively exploring how to combine this evolutionary mechanism with cutting-edge reasoning capabilities to empower every individual to create and manage their own intelligent Agent teams, thereby realizing “one person, an army” in the business world.

Digital Asset Security and Agent Inheritance: Future Considerations for 404 Protocol

As AI Agents' capabilities grow stronger, their participation in the digital economy will also increase, inevitably leading to new demands for digital asset security and inheritance. Imagine, when an AI Agent team manages a large volume of digital assets, how can the security of these assets be ensured in extreme situations? XB Lab's 404 Protocol (DeFi) precisely protects your digital assets through an on-chain proof-of-life protocol, ensuring seamless inheritance. When AI Agents become crucial managers of our digital wealth, a robust, decentralized protocol capable of proving “life” status will be key infrastructure to guarantee digital asset ownership and secure inheritance. 404 Protocol is also currently in TRL 5 internal testing, and in the future, we will continue to explore the intersection of AI Agents and DeFi, providing more robust underlying safeguards for the AI-driven digital economy.

GownCloud: When Traditional Industries Meet Digital Transformation

While pursuing cutting-edge AI and Web3 protocols, XB Lab is also committed to bringing the power of digitalization and intelligence into traditional industries. GownCloud (Other) is a high-end gown rental and cloud-based digital management platform. Although it seems distant from the theoretical AI breakthroughs represented by Aletheia, its core philosophy – enhancing efficiency, optimizing experience, and expanding business boundaries through technology – is consistent with XB Lab's overall vision. In an era of “one person, one Agent team,” even traditional industries can achieve unprecedented business value through efficient digital management and potential future AI-assisted decision-making.

In an era where 'people are nodes,' AI Agent teams will be the ultimate weapon to amplify individual potential and cross business boundaries.

Frequently Asked Questions

Q1: What is Aletheia?

Aletheia is an AI Agent developed by Google DeepMind, specialized in solving research-level mathematical problems. Powered by the Gemini 3 DeepThink model, it independently solved six out of ten world-class mathematical problems in the “FirstProof Challenge” on March 1, 2026.

Q2: What is the “FirstProof Challenge”?

The “FirstProof Challenge” is a challenge set by a team of mathematicians, designed to evaluate current AI systems' ability to solve research-level mathematical problems. The problems in the challenge originate from the mathematical research process, and their solutions have not been publicly released previously, thus preventing AI from solving them by memorizing training data.

Q3: How does the “self-filtering” mechanism reduce AI hallucinations?

The “self-filtering” mechanism is a key capability of Aletheia, allowing the AI Agent to perform internal verification, cross-referencing, and multi-step logical validation when generating solutions. When it cannot reach a reliable conclusion, it chooses not to output a result rather than generating inaccurate or fabricated information, thereby effectively suppressing the occurrence of AI hallucinations.

Q4: What is the connection between Rotifer Protocol and Aletheia?

Rotifer Protocol is XB Lab's AI Agent product, focused on enabling autonomous evolution of AI agents through a genetic competition framework. Aletheia's breakthroughs in mathematical problem-solving, especially its DeepThink and “self-filtering” mechanism, provide important technical reference and visionary inspiration for Rotifer Protocol's exploration of autonomous learning and self-optimization paths for AI Agents.

Q5: What impact will the development of AI Agents have on the future?

The development of AI Agents, especially Aletheia with its PhD-level reasoning capabilities, signals that AI will transform from simple tools into partners capable of deep intellectual collaboration with humans. This will greatly expand the business boundaries for individuals and enterprises, empowering the “one person, one Agent team” model to achieve unprecedented innovation and efficiency improvements across various domains.

Does the advent of Aletheia mean we are on the cusp of a new business era driven by a few elite AI Agent teams? What are your thoughts? We look forward to your comments!