Towards Autonomous Mathematics Research: the Aletheia agent
For the first time, an AI system has produced a publication-grade mathematics paper with no human intervention — and the same system has resolved four previously open Erdős conjectures. The Aletheia paper (arXiv:2602.10177, Google DeepMind, March 2026) marks a step-change from the competition-level results described in the prior ai-for-mathematics report. The question is no longer whether AI can assist mathematics research, but how to characterise and govern the degree of autonomous contribution.
Source: arXiv:2602.10177v3
Authors: Tony Feng, Trieu H. Trinh et al. (Google DeepMind); with contributing mathematicians Junehyuk Jung, Joonkyung Lee, Carlo Pagano, Sang-hyun Kim, Federico Pasqualotto, Sergei Gukov
Published: March 2026
Code and outputs: https://github.com/google-deepmind/superhuman/tree/main/aletheia
What Aletheia is
Aletheia is a multi-agent system for open-ended mathematical research, powered by Gemini Deep Think. It operates end-to-end in natural language — unlike AlphaProof or AlphaGeometry, which require problems to be formalised in Lean 4 or a domain-specific language first.
Three-agent loop:
| Agent | Role |
|---|---|
| Generator | Proposes candidate solutions or proof strategies in natural language |
| Verifier | Evaluates correctness; explicitly trained to recognise and flag its own errors |
| Reviser | Modifies the solution based on verification feedback |
The loop iterates until the solution is approved or an attempt limit is reached. Decoupling verification from generation is a key design choice: the paper reports that separating these roles substantially improves error recognition in practice.
Tool use is essential. Without internet access, models consistently hallucinate references and computational results. Aletheia uses Google Search and web browsing to retrieve literature, and Python execution for symbolic or numerical verification. The paper notes that internet search training reduces obvious citation hallucinations but shifts failures to subtler inaccuracies — active tool use during inference is necessary, not optional.
Scaling law. The paper introduces an inference-time scaling law that transfers from Olympiad-level to PhD-level problems. The January 2026 Gemini Deep Think model reduced compute requirements by approximately 100× compared to the IMO-Gold model from July 2025, while matching or exceeding its accuracy.
Key results
| Milestone | Classification | What happened |
|---|---|---|
| Feng26 — Eigenweights for Arithmetic Hirzebruch Proportionality | Level A2 (Essentially Autonomous, Publication Grade) | Fully AI-generated paper; resolves a structure-constants problem in arithmetic geometry using algebraic combinatorics techniques unfamiliar to the original human co-authors. No human intervention in core content. |
| LeeSeo26 — Lower Bounds for Multivariate Independence Polynomials | Level C2 (Human-AI Collaboration, Publication Grade) | Aletheia provided the high-level proof strategy; human mathematicians executed rigorous proofs. Links combinatorics with statistical physics. |
| Erdős Problems — four open conjectures resolved | Level A (Essentially Autonomous) | Systematically deployed on 700 open problems from Bloom’s Erdős database. 63 technically correct solutions; 13 meaningfully correct (6.5%); four judged to resolve genuine open problems (Erdős-652, 654, 1040, 1051). One extended into a full paper (BKKKZ26). |
| FirstProof — research-level benchmark | Level varies | 6 of 10 problems solved (best-of-2); one solution (P7) at publication grade. Competing systems (GPT-5.2 Pro) solved 2 of the same 10. |
On Erdős-1051 and specification gaming: The paper honestly reports that many technically correct Erdős solutions were “too elementary” — the problems had remained open for decades not because they were hard, but because they fell outside the attention of working specialists. This is both a strength (AI’s breadth compensates for limited specialist focus) and a weakness (AI tends to solve the easiest interpretation of an ambiguous question, not always the intended one).
The Autonomy Levels taxonomy
The paper proposes a two-axis taxonomy for transparent documentation of AI-assisted mathematical results. This is arguably the paper’s most broadly applicable contribution, independent of Aletheia’s specific performance.
Autonomy axis:
| Level | Label | Definition |
|---|---|---|
| H | Primarily Human | AI contributions minor or auxiliary |
| C | Human-AI Collaboration | Both parties contribute essentially to core content |
| A | Essentially Autonomous | Core mathematical content fully AI-generated |
Mathematical significance axis:
| Level | Label | Definition |
|---|---|---|
| 0 | Negligible | Olympiad or PhD exercise level; not novel |
| 1 | Minor novelty | Some new element but not publication-ready |
| 2 | Publication Grade | Publishable in a good research journal |
| 3 | Major Advance | Top-5 journal quality |
| 4 | Landmark Breakthrough | Generational importance |
Combined: Level A2 = essentially autonomous, publication-grade. This is the classification assigned to Feng26 — the first fully AI-generated mathematics paper.
The paper also proposes Human-AI Interaction Cards: structured templates for documenting the sequence of human and AI contributions to a result, analogous to model cards in machine learning. These have not yet been adopted as a community standard but represent a concrete proposal that journals and institutions could implement.
Limitations and honest assessment
The paper is candid about what Aletheia cannot yet do:
- Results are elementary relative to the field. Even publication-grade outputs are brief and use techniques that are sophisticated for an AI but routine for domain specialists. The system is not producing results in the top-5 journals.
- Success is rare. On Erdős problems, 6.5% of solutions were meaningfully correct. On FirstProof, 10% reached publication grade (1 of 10).
- Specification gaming. The system tends to solve the easiest interpretation of an ambiguous question — a known failure mode of LLMs that is particularly problematic in research, where the intended question is often the hard one.
- Hallucination persists. Even with internet access, models fabricate or misrepresent references. Tool use mitigates but does not eliminate this.
- Data contamination is difficult to assess. The paper acknowledges it cannot fully rule out that some solutions are reproductions of material seen during pretraining, reframed as novel.
- Creativity remains a limitation. Success is driven by broad knowledge retrieval and technical manipulation, not by the generative creative leaps characteristic of major mathematical advances.
Relevance to Sheffield
Who benefits now
Aletheia itself is not publicly available — it is an internal Google DeepMind research system. There is no API, no early access programme, and no pricing. Sheffield researchers cannot use Aletheia directly.
However, the paper’s significance for Sheffield is not about accessing the specific system. It is about what the results establish:
| Audience | Relevance |
|---|---|
| Pure mathematicians (algebra, number theory, combinatorics) | The existence of Level A2 results validates heavy investment in AI-assisted research as a strategic direction, not a curiosity |
| Formal methods / CS verification | The contrast with AlphaProof is instructive: Aletheia uses natural language throughout; formal verification tools like harmonic-aristotle remain the more accessible path for rigorously verified results |
| Research leadership and governance | The Autonomy Levels taxonomy is immediately applicable to Sheffield’s own GenAI principles — it provides a vocabulary for describing AI contribution that is more precise than “AI-assisted” |
| Mathematics educators | The discussion of AI’s different intelligence (breadth over depth, unlimited time) is useful framing for departmental conversations about AI’s role in research training |
Governance considerations
- All Aletheia computation runs on Google DeepMind infrastructure (US-hosted). Any future equivalent system would require a DPA review before use with unpublished research.
- The paper’s introduction of Autonomy Levels and Human-AI Interaction Cards is directly relevant to Sheffield’s existing GenAI Principles for Research, which do not yet specify how AI contribution to publications should be characterised. This paper provides a community-developed vocabulary that Sheffield could adopt or reference.
- The Erdős problem result illustrates a governance risk: AI systems may publish solutions to open problems derived from unpublished work shared in prompts. Researchers should not share unpublished problem statements with externally hosted AI systems.
What to monitor (12–24 months)
- Gemini Deep Think public access — the underlying model powering Aletheia is already partially accessible via Google AI Studio and Gemini API. As the January 2026 model becomes available, the scaling gains described in this paper will be accessible to Sheffield researchers without needing Aletheia itself.
- Community adoption of Autonomy Levels — if journals or preprint servers begin requesting AI contribution taxonomy on submissions, Sheffield researchers will benefit from familiarity with the H/C/A × 0–4 framework.
- Formal proof integration — the paper notes that Aletheia operates entirely in natural language, trading hallucination risk for accessibility. A hybrid approach integrating Lean 4 verification with LLM reasoning is a plausible near-term development; harmonic-aristotle is the most accessible current step in that direction.
Relationship to the prior landscape
This paper supersedes the state-of-play described in ai-for-mathematics (drafted March 2026). The prior report identified three paradigms (evolutionary search, formal proof, LLM reasoning) and characterised LLM reasoning as the least reliable of the three. Aletheia represents a fourth paradigm — autonomous natural-language research agents — that had not yet been demonstrated at publication grade when the prior report was written.
Updated milestone table (extending the table in ai-for-mathematics):
| Milestone | Date | Significance |
|---|---|---|
| AlphaProof IMO 2024 silver medal | July 2024 | First AI result at IMO level |
| AlphaEvolve solves 67 open problems | November 2024 | Constructive search at scale |
| Harmonic Aristotle IMO 2025 gold | July 2025 | Formal proof surpasses IMO gold on all six problems |
| Gemini Deep Think IMO-Gold scale | July 2025 | LLM reasoning reaches IMO gold in natural language |
| Aletheia FirstProof: 6/10 research-level problems | February 2026 | Natural-language agent solves research-level (not competition-level) problems |
| Aletheia Feng26: first fully AI-generated publication | March 2026 | Level A2 result — autonomous, publication-grade mathematics |
| Aletheia resolves four open Erdős conjectures | March 2026 | Autonomous resolution of problems open for decades |
Recommended actions
Immediate (now)
-
Share this report with the Mathematics department and the formal methods group in Computer Science. The Feng26 result and the FirstProof benchmark results are concrete evidence that autonomous mathematics research has crossed a threshold. Faculty engaged with AI tools will want to understand what Aletheia represents.
-
Reference the Autonomy Levels taxonomy in any update to the GenAI Principles for Research. The H/C/A × 0–4 framework provides exactly the kind of precise vocabulary for AI contribution in publications that the current principles documents lack. This is low-cost to adopt and directly addresses an emerging governance gap.
Near-term (Q2–Q3 2026)
-
Commission a DPA review that covers research-agent use cases. The existing DPA landscape covers document writing and literature review tools. It does not adequately address AI systems that generate novel content that could be submitted as research output. The Aletheia paper makes this gap concrete.
-
Monitor Gemini Deep Think API availability. The inference-time scaling gains described in this paper will become practically accessible to Sheffield researchers as the January 2026 and later model versions reach the Gemini API. Research IT should track this and be ready to advise on access.
Sources
- arXiv:2602.10177v3 — Feng, Trinh et al.: “Towards Autonomous Mathematics Research” (Google DeepMind, March 2026)
- GitHub: google-deepmind/superhuman/tree/main/aletheia — prompts, outputs, and Human-AI Interaction Card templates
- ErdosProblems.com — Bloom’s database of Erdős conjectures (700 problems used in evaluation)
- FirstProof benchmark — research-level problem set, released February 2026
Drafted: 2026-04-01 · Branch: 8-add-unesco-survey-on-hei-ai-guidance-to-knowledge-base · Requires human review before promotion to main