AI for Science Quarterly Review (2026 Q4 Kickoff): From Discovery to Verification
This is the quarterly tracking report of the AI-HPC expert team, based on public reports and papers from late September to mid-October 2026. Core judgment: agentic science has moved from "proposing hypotheses" to "being verified and challenged" — formal verification is becoming the reliability main battleground of science AI, the lightweight small-model route and the giant foundation-model route run in parallel, and AI as new research infrastructure is entering the "inner court."
I. Industry Progress: AI Enters Science's "Inner Court," Compute Extends to Space
1.1 New-Infrastructure Consensus: From "Visitor" to "Inner-Court Partner"
On October 10, 2026, Xinhua published a systematic account of "AI entering the inner court of research": the "fifth research paradigm" after experimental, theoretical, computational and data science has been formally established. Zhou Bowen, director of the Shanghai AI Laboratory, described early AI for Science as a "visitor to scientists" — helping retrieve literature, analyze data and run computations, yet always "a guest outside the house"; today AI has entered the core research workflow ["http://www.xinhuanet.com/liangzi/20261010/be01c3243b374e2ebb0376af1cc60e8a/c.html"].
Engineering implication: as AI shifts from "tool" to "first-class citizen of the research process," research teams' compute demand shifts from "occasionally borrowing compute" to "continuous, mixed, schedulable supply" — exactly the industry backdrop of the AI-HPC platform positioning: AI-compute (training/inference) and HPC (numerical simulation/verification) need unified scheduling rather than two separate silos.
1.2 Giants Open Up and Race: From "Selling Models" to "Compute + Models"
- Google DeepMind × Google Cloud RAISE (Oct 14): announced direct frontier-model access for up to 200 European researchers, plus $3M in compute credits, linked to the EU's RAISE initiative — covering weather forecasting, biology, mathematical problem solving and Earth mapping ["https://www.neoteo.com/en/google-announces-ai-access-pledge-for-up-to-200-european-researchers"] ["https://lapaasvoice.com/google-deepmind-opens-frontier-ai-models-to-european-researchers/"].
- Microsoft Research Quine (early Oct): an experimental AI research system aiming to build a biology multimodal world model — covering genomics, proteins, chemistry and more, with an interactive harness framework connecting models, scientific tools, literature and researchers, pushing AI from assisted analysis toward scientific discovery ["https://finance.sina.com.cn/roll/2026-10-03/doc-initxwvx3877521.shtml.md"].
Observation: the top players' AI4S competition is shifting from "releasing a model" to a full-stack supply of "model + compute + toolchain + researcher ecosystem" — echoing the Software Factory idea of "continuously manufacturing reproducible software tools for real scientific problems."
1.3 Compute Reaches Space: Compute Constellation Goes Engineering
On Oct 14, Beijing News reported that the solar-wing battery arrays of 6 compute satellites were delivered in product-grade batches and launched (single-star wing area over 4 m², 6 stars over 24 m² total), marking low-cost flexible silicon-solar-wing battery technology moving from on-orbit verification into batch engineering application for compute constellations ["http://m.toutiao.com/group/7696540402539512383/"].
Engineering implication: compute forms are diversifying (ground clusters → edge → space), but the core constraints for space compute returning to ground remain "power + cooling + networking" — for ground AI/HPC engineers, understanding the boundaries of heterogeneous compute forms matters more than chasing new forms.
II. Academic Frontier: From "Discovery" to "Verification"
2.1 Mathematics Becomes the "Touchstone" for Agentic Science: Peking University Team Refutes Two Conjectures in 5 Hours
On Oct 14, STAR Market Daily reported progress from the team of Dong Bin (Boyi Distinguished Professor, Beijing International Center for Mathematical Research, Peking University): the team built formal retrieval engines, natural-language math search engines and other underlying infrastructure, and developed automated reasoning agents and multi-agent collaboration systems — agents can autonomously prove and refute the Andersson conjecture, refuting two frontier math conjectures within 5 hours; and completed large-scale formalization of the Poincaré conjecture within two weeks at low cost ["http://m.toutiao.com/group/7696320607918506531/"].
Why it matters: mathematics is the only discipline where "verification equals truth" — agents breaking through in math amounts to the strictest proving ground for the reliability of science AI.
2.2 "Future Mathematics Initiative": Making AI-Generated Content Verifiable, Accumulable, Reusable
On Oct 14, the Institute for Advanced Study in Mathematics at Zhejiang University and Zhejiang Lab jointly launched the "Future Mathematics Initiative," covering four directions around two-way AI-math empowerment: developing mathematical foundation models (making math-reasoning ability the underlying capability of scientific foundation models), conducting formal-verification research (making AI-generated mathematical content verifiable, accumulable and reusable), and studying the mathematical foundations behind AI ["http://m.toutiao.com/group/7696521233546756623/"].
2.3 OpenAI's Math Claim Stirs Academic Controversy: Science AI's "Trust Problem" Surfaces
On Oct 14, CCTV reported that OpenAI claims its models can crack hard math problems, drawing academic controversy ["https://ydyl.cctv.com/2026/10/14/ARTISYcV2STtBU0Wq6xJg7bX261014.shtml"].
Our reading: the controversy is not about "can it solve," but "how do we prove the solution is correct" — which precisely confirms this review's theme: AI4S's bottleneck has moved from "generation capability" to "verification capability." Without a verification loop, a "discovery" is merely a "guess" in scientific terms.
2.4 The Lightweight Route Rises: MatBrain Dual-Model Collaboration
On Sept 10, the Shenzhen Institute of Advanced Technology, CAS released "MatBrain," the first lightweight dual-model collaborative agent in materials (published in Nature Machine Intelligence): without a "super-large model," two small models collaborating can perform complex materials research ["http://www.cas.cn/cm/202609/t20260911_5120171.shtml"].
Engineering implication: this directly echoes the "tool packaging + reliability defenses" route of 13 · Scientific Agent Deployment — small models + tool collaboration + clear boundaries is a replicable agent route for research groups with limited compute budgets; "bigger is always better" is not the only answer.
2.5 A General-Purpose Biomedical Agent: Biomni
Stanford researchers released Biomni — a general-purpose biomedical AI agent: its action-discovery agent mines tools, databases and protocols from thousands of papers across 25 domains, building a unified agentic environment; the architecture integrates LLM reasoning, retrieval-augmented planning and code-based execution, dynamically composing workflows without predefined templates ["https://profiles.stanford.edu/186687"].
III. The Verification Bottleneck: Science AI's "Trust Problem"
Viewed together, one thread runs through October's events: from "can it generate" to "can it verify."
| Layer | Verification means | October progress/signal |
|---|---|---|
| Mathematics | formal verification | PKU Poincaré formalization, ZJU Future Math Initiative, OpenAI controversy |
| Materials/Bio | experiment/simulation re-check | MatBrain lightweight route, Biomni unified benchmarking |
| Physics/Engineering | solver re-check + physics constraints | see 14 · Physics Consistency methodology |
| Grand facilities | observational verification | 2026 Nobel Prize in Physics awarded to the IceCube Neutrino Observatory (Oct 6); grand facilities + compute remain the hardest link of "verification" ["http://m.toutiao.com/group/7694091799849927204/"] |
Our judgment: over the next 2–4 quarters, AI4S competition will shift from "what the model can propose" to "whether the system can prove the conclusion trustworthy" — formal verification, evaluation benchmarks, and closed-loop re-checks against solvers/experiments will become the core competitiveness of science-AI platforms.
IV. Engineering Recommendations (Mapped to This Platform's Knowledge Base)
For research groups: prefer the "lightweight + tool collaboration + verification defenses" route for agentic science (see the MatBrain model and 13 · Agent Deployment) — the four defenses (tool-output validation, retry, numerical result verification, human confirmation) are the lowest-cost "trustworthiness" measures.
For cluster engineers: AI entering science's inner court means mixed workloads become the norm — training, inference and simulation verification interleave on the same cluster. Heterogeneous scheduling (07) and GPU resource partitioning (MIG practice) upgrade from "optional optimization" to "standard capability."
For platform builders: make the "verification loop" a first-class citizen — formalization tools, evaluation task sets and re-check channels to simulation solvers should be platform capabilities on par with model training (echoing the AI4Science Engine and Software Factory positioning).
V. Next Review Watchlist
- Whether mathematical foundation models' reasoning ability "probes upward" into scientific foundation models' base (progress of the ZJU initiative);
- Impact of the giants' open ecosystems (RAISE, Quine) on research paradigms;
- Whether a "unified scientific benchmark" emerges in verification and evaluation;
- Progress of compute-constellation engineering and its synergy with ground mixed compute.
Note: this review is compiled from public reporting; facts and figures follow the cited sources; opinions are this team's engineering-perspective analysis and do not constitute investment or technology-selection conclusions.