Google DeepMind releases AI mathematician: refreshing the most difficult benchmark SOTA 48%

Core conclusion

Google DeepMind released AI Co-Mathematician on May 11, 2026, a multi-agent collaboration system designed specifically for mathematical research, aiming to change the status quo of "AI is just an auxiliary reasoning tool in the field of mathematics". The system achieved 48% accuracy on FrontierMath Tier 4 (the most difficult level), setting a new record for all AI systems, and has helped Oxford University mathematicians solve problem No. 21.10, which has been unsolved for decades in the field of group theory.

Key Points

• Event time: 2026-05-11 (paper release)
• Core breakthrough: no longer a single inference tool, but a multi-Agent workbench for long-term collaboration
• Benchmark: 48% accuracy on FrontierMath Tier 4 (only 19% on base model Gemini 3.1 Pro)
• Real impact: 3 mathematicians have been assisted to advance or solve open-ended mathematical problems

Background: The paradigm shift from "tool" to "collaborator

In the past, improvements in AI mathematical capabilities focused on local capabilities: stronger reasoning, more mature formal proofs, and more convenient tool calls. But these capabilities have not been integrated into a sustainable research process.

The core problem that AI Co-Mathematician tries to solve is exactly this - it is no longer just answering a certain step of reasoning or supplementing a proof, but building a multi-agent workbench that can collaborate for a long time.

According to the paper description (arXiv:2605.06651), the system is managed uniformly by a Project Coordination Agent: first clarifying the boundaries of the problem, confirming the research goals, and then splitting the tasks into different workflows. Each workflow can be advanced in parallel, calling sub-agents such as literature retrieval, code experiment, proof attempt, result review, etc., and writing intermediate results back to the shared file system.

The final delivery is no longer a context-rich conversation, but a continuously updated working document that retains marginalia, source notes, internal links, and traces of review.

Key Impact (by Dimension)

Detailed explanation of multi-Agent collaboration architecture

The architectural design of AI Co-Mathematician deserves the attention of all AI Agent practitioners. Its core design principles include:

1. Project coordination Agent mode Users mainly interact with the top-level coordination agent, which first clarifies the goal and then breaks down the task. This design avoids the confusion of users facing multiple sub-Agents directly.

2. Parallel Workflow Multiple research paths can be advanced simultaneously and a reviewed report will be produced after each is completed. If a certain path fails to complete the task, the system will directly give an eye-catching warning.

3. Uncertainty Management

• The code has not passed the test and cannot be considered completed.
• The report has not passed review and cannot be finalized directly
• Chronic stuck issues must be clearly exposed to users

4. Failed Assumption Tracking All failed assumptions, dead ends, and holes exposed in the review are retained as formal research context rather than simply discarded.

Implications for AI Agent automation practitioners

The design concept of AI Co-Mathematician has direct reference value for AI automation workflow:

• Project coordination layer: In the automated process driven by n8n or Claude Code, similar coordination agents are introduced to uniformly manage complex tasks.
• Failure Logging: Don't throw away errors and failure paths - they are valuable in debugging and process optimization
• Parallel Workflow: Multiple Agents advance simultaneously and complement each other, which coincides with WayToClawEarn's content flywheel strategy

If you want to apply a similar pattern in automated production, you can refer to the multi-Agent orchestration method in the tutorial.

Related extended information

-Original text of the paper: https://arxiv.org/abs/2605.06651

• 36 Chinese report: https://36kr.com/p/3804373708775174
• Google DeepMind official blog (to be released)

Tool entry

Tool entries involved in the text: Google DeepMind, Gemini, Claude Code, n8n, LangGraph

Internal link guidance

• Want to start building a multi-Agent system? See: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
• Real case: 18-Year-Old Built a $5,000/mo SaaS With AI Agents — Zero Hand-Written Code
• Recommended local AI practices: How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial