[Google DeepMind] Aletheia: The Dawn of Autonomous Mathematical Research

TL;DR

Google DeepMind's Aletheia, powered by Gemini 3 Deep Think, has achieved a milestone in automated reasoning by solving 60% of the FirstProof challenge—a set of ten problems designed by professional mathematicians to be "AI-hard." Moving beyond simple prompt-engineering, Aletheia uses a sophisticated agentic scaffolding to produce publication-quality LaTeX proofs with zero human intervention during the generation phase.

The Challenge of FirstProof

In early 2026, a group of mathematicians released FirstProof, a collection of 10 research-level questions intended to test whether AI could act as an autonomous researcher. The criteria were strict:

1. Strict Autonomy: No human-in-the-loop for mathematical ideas.
2. Professional Rigor: Proofs must conform to the standards of peer-reviewed journals.
3. Precise Citations: References must include specific statement numbers from arXiv or journals.

Existing models often "hallucinate" logical steps or struggle with the sheer scale of the inference required for research-level proofs. Aletheia was designed specifically to tackle this "reliability bottleneck."

Methodology: The Aletheia Scaffolding

Aletheia's secret sauce isn't just a bigger model; it's the orchestration. The system uses a "Best of 2" approach across various versions of Gemini 3 Deep Think.

The core of the process involves a Verification and Extraction Pipeline:

• Generation: The base model produces a candidate proof.
• Critique: A reviewer agent (also Deep Think) line-by-line verifies the logic, actively searching for fallacies or "leaps in logic."
• Resolution: Depending on the verdict ([CORRECT], [FIXABLE], or [WRONG]), the agent either accepts, autonomously revises, or discards the solution.

Figure: The autonomous pipeline—from raw problem LaTeX to expert-verified solution.

Key Performance Metrics

Aletheia managed to solve 6 problems, including open problems like Problem 7 (a variation on a theme of Borel regarding semi-simple groups).

The "Inference Cost" (Figure 1) provides a fascinating insight: the model spent significantly more "thinking time" on Problem 7—an order of magnitude higher than previous benchmarks. This correlates directly with the problem's status as a formerly open question in mathematics literature.

Figure: Inference-time computation as a proxy for problem difficulty.

Deep Dive into Problem 10: Outsmarting the Humans

Perhaps the most impressive feat occurred in Problem 10, involving CP decomposition in an RKHS (Reproducing Kernel Hilbert Space).

• The Baseline: Standard direct solvers cost $O(n^3{r}^3)$.
• Human SOTA: Most expert solutions used an iterative PCG (Preconditioned Conjugate Gradient) approach that scaled with $O(q)$ (number of observed entries).
• Aletheia's Discovery: The agent autonomously derived a precomputation step involving per-row Gram accumulators. This allowed the iterative loop to run in $O(n^{2}r+n{r}^2)$, making it significantly faster for dense datasets where $q\gg n$.

Critical Analysis & Conclusion

Aletheia represents a transition from "Chatbot" to "Agent." While it couldn't solve all 10 problems (failing on P1, P3, P4, and P6 due to its own "self-filtering" reliability mechanism), the 6 it did solve were handled with a level of rigor that fooled expert reviewers.

Takeaway: We are entering an era where Inference Scaling—letting the model think longer and verify its own work—is the primary driver of capability in formal sciences.

Limitations: Aletheia still lacks a "global" intuition for which problems are worth solving, and its self-filtering sometimes discards valid but "unsure" paths. However, its ability to generate $O(N)$-free algorithms for complex tensor problems suggests that AI will soon be a co-author, not just an assistant, in high-level mathematics.

For those interested in the raw proofs, DeepMind has released the full prompts and PDF outputs on their public GitHub repository.