Google’s Aletheia AI Solves 6 of 10 Groundbreaking Math Problems in FirstProof Challenge

Google’s Aletheia AI Solves 6 of 10 Groundbreaking Math Problems in FirstProof Challenge

In a landmark development for artificial intelligence in pure mathematics, Google DeepMind’s Aletheia agent has successfully solved six out of ten research-level problems in the FirstProof challenge—a benchmark designed to test AI’s capacity for original mathematical reasoning. The results, published on arXiv on February 25, 2026, demonstrate that AI systems are no longer merely assisting mathematicians but can independently generate novel, peer-review-worthy solutions to open problems.

According to Reuters, Aletheia, powered by the Gemini 3 Deep Think architecture, autonomously tackled the problems within the 8-day window of the challenge, submitting its answers prior to the public release of official solutions. The problems, curated by a team of professional mathematicians and released on February 5, 2026, were drawn from active research areas in algebraic geometry, number theory, and topology, making them particularly resistant to pattern-matching or pre-trained heuristics.

The evaluation process was overseen by a panel of 12 experts from institutions including MIT, Caltech, and the Institute for Advanced Study. Each solution was assessed for correctness, logical coherence, and originality. Aletheia received full consensus on problems 2, 5, 7, 9, and 10. Problem 8, involving a novel conjecture in derived algebraic geometry, received majority approval (8 out of 12 experts) but sparked debate over the interpretation of a key lemma. The authors note this as evidence of Aletheia’s ability to operate at the frontier of mathematical research, where even human experts may disagree.

Transparency was central to the project. Google DeepMind published all raw prompts, intermediate reasoning traces, and final outputs on GitHub under an open license, enabling independent verification. This level of disclosure is unprecedented in AI research, particularly in fields traditionally guarded by academic secrecy. “We wanted to avoid the ‘black box’ critique,” said Tony Feng, lead author of the paper. “If an AI solves a problem no human has solved, the community deserves to see how it got there.”

The FirstProof challenge was conceived by a group of mathematicians led by Mihai Abouzaid as a direct response to the growing claims of AI capabilities in STEM. Unlike previous benchmarks such as MATH or GSM8K, which rely on textbook-style problems, FirstProof problems were selected from recent preprints and conference discussions—problems that had stumped even seasoned researchers for months. The fact that Aletheia solved more than half of them suggests a qualitative leap in AI reasoning, not just quantitative improvement.

Some experts caution against overinterpretation. “Solving six problems doesn’t mean AI can do mathematics,” said Dr. Elena Vasquez, a professor at Princeton. “It means it can solve six specific problems, in a constrained environment, with massive computational resources. The real test is whether it can identify the next problem to solve.”

Nonetheless, the implications are profound. Aletheia’s success may accelerate the adoption of AI as a co-researcher in academic labs. Several institutions have already begun integrating similar agents into their workflows. The paper’s authors suggest future iterations could be trained to propose new conjectures, not just solve existing ones.

For now, the mathematical community is digesting the results. The GitHub repository has attracted over 12,000 stars in its first week, and independent teams are attempting to replicate Aletheia’s approach using open-source models. Whether this marks the dawn of AI-driven mathematical discovery or a temporary anomaly remains to be seen—but one thing is clear: the role of the mathematician is evolving.