DeepMind releases AI math research assistant: its multi-agent architecture outperforms GPT-5.5Pro and solves previously unsolvable problems.

According to Beating, Google DeepMind released AI co-mathematician, a multi-agent interactive research workbench for mathematicians. The system achieved a 47.9% accuracy rate (solving 23 out of 48 problems) on FrontierMath Tier 4, currently the most difficult research-grade mathematics benchmark, directly surpassing the previous record of 39.6% held by GPT-5.5 Pro. This system doesn't use a new-generation platform; it directly uses Gemini 3.1 Pro. The model itself only achieved 19% accuracy on Tier 4 without the framework, but its performance more than doubled after adding the agent framework. DeepMind built a multi-layered architecture: a top-level "project coordinator" breaks down research tasks into multiple workflows, distributing them to sub-agents responsible for literature retrieval, coding, and inference. The written proofs must pass a review by multiple "reviewing agents" before submission. This robust scaffolding demonstrates that in top-tier mathematical inference, the incremental capabilities that orchestration can extract can be greater than that of replacing the entire model. The blind testing was conducted by Epoch AI. To prevent cheating, the DeepMind team did not see the questions throughout the process, and each question was allowed 48 hours to run. The result was not only a top-ranking achievement, but the system also solved three questions that all previous models had failed to solve. Although named "Deputy," it's more like a colleague with innovative thinking. Group theory expert Marc Lackenby used it in actual research to solve a public conjecture in Kourovka's notebook. Interestingly, the system's initial strategy was marked as "flawed" by its own review agent, but Lackenby saw the clever idea hidden in the rejected solution, filled the gaps himself, and ultimately completed the proof. Currently, the AI co-mathematician is only open for internal testing to a small number of mathematicians.