AI Science Tracker

Google DeepMind

52 discoveries from 2020 to 2026

Discoveries

Google DeepMind

Agentic cheminformatics: new macrocycles from literature analysis

An agent using Cheminformatics Skills for Antigravity collected thousands of compounds analogous to crizotinib and lorlatinib, analyzed 568, and proposed new rigidified macrocycles with synthesis routes in about an hour. The proposals are computational hypotheses that still need to be synthesized and validated.

Technology: Google Antigravity, Cheminformatics Skills

Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: retrospective or benchmark
  • Tested: The source describes computational evaluation only: benchmarks of several machine learning architectures on four readouts with scaffold splits, and predicted potency; no synthesis or experimental validation is reported.
  • Still unknown: Every proposed design must still be synthesized and validated; the data are sparse and assay noise is high; the designs are testable hypotheses, not finished drugs.
  • AI role: An agent searched literature for crizotinib and lorlatinib analogues, curated 568 compounds, ran matched molecular pair analysis, benchmarked models, and proposed macrocycles and synthesis routes.

Evidence sources: https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

ForTy 2020: global map of forest types at 10 meters

A 10-meter map classifies each land area into six forest or land-use types (including primary forest, plantations, and tree crops), with an initial time series from 2017–2025. Validation across more than 9,000 sites identifies the correct type in 90% of cases and distinguishes forest from non-forest in 95%.

Technology: Satellite-image classification models

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The map was validated at more than 9,000 random sites, where two trained experts checked high-resolution imagery to determine land type: correct forest type in 90% of cases and forest vs non-forest in 95%.
  • Still unknown: Per-class accuracy and regional breakdowns are not reported, the time series is early-release without separate validation, and results depend on the probability threshold chosen.
  • AI role: A custom remote-sensing model trained on about 1.7 million labeled locations, using AlphaEarth embeddings, outputs probabilities for six forest and land-use classes.

Evidence sources: https://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Searching for antimicrobial peptides with Google Antigravity

An agentic workflow searched UniProt for antimicrobial candidates in Streptococcus pseudopneumoniae and found a 97-amino-acid bacteriocin with no homologues among manually reviewed Swiss-Prot proteins. This is a computational candidate now undergoing laboratory validation; no potency or toxicity data are available.

Technology: Google Antigravity (agentic platform)

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: Nothing was tested in the lab: the candidate's novelty was checked by BLAST against Swiss-Prot, its predicted structure was compared with a known bacteriocin fold, and its genomic context was examined; the wet-lab plan is underway with no results.
  • Still unknown: No potency or toxicity results exist; the wet-lab assays outcomes are unknown; the session covered a narrow sequence window in one species.
  • AI role: An agent queried databases, filtered candidates, ran BLAST and structure comparisons, analyzed genomic neighborhoods, synthesized literature and drafted a wet-lab plan under researcher direction.

Evidence sources: https://deepmind.google/blog/accelerating-antimicrobial-discovery/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 1/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

SynthID Bio: watermarking for proteins and synthetic structures

Methods insert a detectable signal into protein sequences and predicted structures. For binders designed against three targets (VEGF-A, SARS-CoV-2 RBD, and PD-L1), watermarked designs matched the others in success rate and affinity. This is a proof of concept that needs to be made more robust against tampering.

Technology: SynthID Bio, AlphaFold 3 (optimized component)

Family AlphaFold 3: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Wet-lab tests on three targets (VEGF-A, SARS-CoV-2 RBD, PD-L1), with Adaptyv Bio helping, showed watermarked designs matched unwatermarked ones in hit rate, affinity and diversity; the phage was tested only in bacterial cultures.
  • Still unknown: The watermark must become more robust against deliberate tampering, the bacteriophage work is early-stage, and the post says no single biosecurity measure is a silver bullet.
  • AI role: SynthID Bio guided amino acid choice in a ProteinMPNN variant, adjusted predicted atomic coordinates, fine-tuned part of AlphaFold 3, and was integrated into Evo 2 to watermark a designed phage genome.

Evidence sources: https://deepmind.google/blog/introducing-synthid-bio/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaGenome Atlas: predictive map of human DNA variation

A precomputed database of predicted molecular effects for about 9 billion possible single-letter changes in the human genome, generated by AlphaGenome. Researchers can access it free through a portal and API. The predictions are a research tool, do not replace experiments, and are not validated for clinical use.

Technology: AlphaGenome, AlphaGenome Variant Impact (AVI)

Family AlphaGenome: profile · models page

Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: retrospective or benchmark
  • Tested: Validation described is the company's own testing of the AlphaGenome Variant Impact score on pathogenicity benchmarks, plus case studies: a DNM1 variant prioritized with GREGoR and 22% more non-coding associations in UK Biobank.
  • Still unknown: The page does not detail how individual predictions were checked; AlphaGenome is not validated or approved for clinical use.
  • AI role: AlphaGenome predicted variant effects across cell types and tissues; these were precomputed for about 9 billion variants and combined with AlphaMissense into a single score.

Evidence sources: https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

WeatherNext 3: a global AI model for weather forecasting

A model that learns from real-time geostationary satellite data and produces hourly forecasts, at resolutions down to 5 km for temperature and humidity (about five times finer than WeatherNext 2). For short-term precipitation, Google reports improvements of up to 60% over NASA's IMERG. These are ‘up to’ figures from Google's evaluations, with an independent ranking by Brightband.

Technology: WeatherNext 3

Family WeatherNext: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Precipitation forecasts were scored with CRPS against baselines: gains of up to 60% over IMERG, 30% over MRMS and 10% over rain gauges for early lead times; independent live evaluations by Brightband ranked the model.
  • Still unknown: The 'up to' figures lack full test setup, longer-range gains are largest where forecasts are weak, and the atmosphere retains unpredictability.
  • AI role: The model ingests live geostationary satellite data and historical analyses, trains on station observations and produces gridded fields, cyclone tracks and station-level forecasts.

Evidence sources: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

WeatherNext Cyclones: forecasting cyclones one day earlier

Three-day forecasts match what earlier models achieved at two days: about 100 km trajectory error and 11 knots intensity error, compared with ECMWF-ENS and HWRF. The work is with the National Hurricane Center, CIRA, and the Met Office. The comparison uses historical 2023–2024 cyclones; why results are good at 28 km remains an open question.

Technology: WeatherNext Cyclones

Family WeatherNext: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Historical cyclones from 2023-2024 were benchmarked deterministically and probabilistically against ECMWF-ENS (track) and HWRF (intensity): three-day errors matched earlier two-day forecasts; a 2025 storm is cited only as an anecdote.
  • Still unknown: Why the model is accurate at 28 km resolution is an open question; the 'extra day' is an average dependent on the chosen benchmarks, and the evaluation window is slightly unclear.
  • AI role: A single model using Functional Generative Networks produced track, intensity and wind-structure ensembles of up to 1,000 members.

Evidence sources: https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Co-Scientist: repurposed drugs against liver fibrosis

In Gary Peltz's Stanford lab, two of the three existing drugs proposed by Co-Scientist blocked fibrosis and promoted regeneration in human liver cells, while Peltz's own two choices showed no benefit. The anticancer drug vorinostat blocked 91% of a damage response that can lead to scarring. Cell experiments only; published in Advanced Science.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The anti-fibrotic efficacy and toxicity of 25 drugs were assessed serially in human liver organoids grown in microwells; Co-Scientist's candidates were tested in cells alongside the authors' own picks.
  • Still unknown: Not reported.
  • AI role: The AI co-scientist multi-agent system proposed three candidate drugs with reasoning, and helped direct their experimental characterization.

Evidence sources: https://deepmind.google/blog/uncovering-repurposed-medicines-to-fight-liver-fibrosis/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1002/advs.202508751&resultType=core&format=json

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 27 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Co-Scientist and aging: metabolism and the integrated stress response

Calico Life Sciences used Co-Scientist to formulate the hypothesis that metabolism regulates the cellular integrated stress response, refined experiments with the tool, and reports new findings. The post gives no numerical data, and the results have not yet been published or peer reviewed.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: Calico reports that experiments were run with the Co-Scientist hypothesis and produced new findings; the post does not describe the experiments, assays or results.
  • Still unknown: Not reported (the results are not yet published; the team plans to publish them).
  • AI role: Co-Scientist generated the hypothesis that metabolism regulates the integrated stress response, and was used to refine experimental design.

Evidence sources: https://deepmind.google/blog/opening-new-paths-in-aging-research/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Co-Scientist and mechanisms of metabolic steatohepatitis (MASH)

Filippo Menolascina (Edinburgh) used Co-Scientist to search the literature for combinations of MASH therapies. The tool proposed the NLRP3 inflammasome as a link between inflammation and metabolism, and the post says the hypothesis was later tested experimentally. No numbers are given; the drug combinations were suggested for testing, not demonstrated.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: The article says the NLRP3 hypothesis was later experimentally verified but does not describe the experiments, who carried them out, or their results.
  • Still unknown: Not reported (the article does not state limitations or whether the work is published).
  • AI role: Co-Scientist searched the literature, proposed drug combinations for MASH therapy and proposed the NLRP3 inflammasome as the link between inflammation and metabolism.

Evidence sources: https://deepmind.google/blog/accelerating-discovery-of-liver-disease-mechanisms/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Co-Scientist: genetic factors to rejuvenate cells

Omar Abudayyeh and Jonathan Gootenberg used Co-Scientist to search tens of thousands of papers, obtaining over 20 proposed genetic factors for reversing cellular aging, a couple of which were validated in the lab, and reducing analysis of genetic-screening data from six months to a few days. The post gives no sample size or peer review status.

Technology: Co-Scientist

Family AI Co-Scientist: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Lab tests validated a couple of the more than 20 proposed genetic factors, which pushed cells toward a younger state; the experiments and exact number are not described.
  • Still unknown: Not reported (the post gives no sample size or peer-review status and contains only the investigators' quotes).
  • AI role: Co-Scientist scanned tens of thousands of papers, proposed genetic factors for reversing cellular aging and analysed screening data alongside the literature.

Evidence sources: https://deepmind.google/blog/fast-tracking-genetic-leads-to-reverse-cellular-aging/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

WeatherNext and Hurricane Melissa: earlier warning in Jamaica

WeatherNext forecast a Category 5 landfall in Jamaica five days ahead with 80% probability, rising to nearly 100% at three days. According to the post, it was the best single model for track and intensity in the National Hurricane Center's 2025 verification report, which used it alongside its HAFS model. The post gives no numerical comparisons; rapid intensification remains very hard to predict.

Technology: WeatherNext

Family WeatherNext: profile · models page

Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: prospective or independent check
  • Tested: The National Hurricane Center's 2025 annual verification report found WeatherNext the top-performing individual model for track and intensity; the Melissa forecast is presented as a single case.
  • Still unknown: The page notes that predicting rapid intensification is very difficult and that Weather Lab is experimental; no dedicated limitations section.
  • AI role: WeatherNext, an AI weather model with 50-member ensembles, forecast the Hurricane Melissa landfall alongside physics-based HAFS and was used by NHC forecasters.

Evidence sources: https://deepmind.google/blog/how-weathernext-helped-the-national-hurricane-center-better-predict-hurricane-melissas-historic-landfall-in-jamaica/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaEvolve raises nine Ramsey lower bounds

AlphaEvolve-generated search algorithms improved bounds such as R(3,13) from 60 to 61, R(4,16) from 170 to 174 and R(4,19) from 213 to 219. The paper does not describe independent verification; search code is public. arXiv preprint, latest version read.

Technology: AlphaEvolve

Family AlphaEvolve: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Search algorithms evolved by AlphaEvolve improved nine Ramsey lower bounds; the authors report recovering bounds for all known exact Ramsey numbers and matching best known bounds in many other cases. No independent verification is described.
  • Still unknown: Not reported
  • AI role: LLM-based code mutation agent that produced the search algorithms behind all results.

Evidence sources: https://arxiv.org/abs/2603.09172

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Gemini Deep Think on research problems in mathematics, physics, and computer science

The Aletheia mathematical agent scored up to 90% on IMO-ProofBench Advanced and independently solved four open Erdős questions; work with experts covered 18 research problems, including a disproof of a submodular optimization conjecture. Its score on doctoral-level problems falls to about 38–46%; no result is claimed as a breakthrough, and the papers were submitted to journals.

Technology: Gemini Deep Think, Aletheia agent

Family Gemini: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Aletheia was evaluated on IMO-ProofBench Advanced (about 90%, human experts graded), an internal PhD-level benchmark (about 46%), a semi-autonomous run over 700 open Erdős problems (four autonomous solutions), and expert collaborations; Level 2 papers were submitted to journals.
  • Still unknown: The article states that no Level 3 or Level 4 results are claimed; the abstract states no limitations, and the PhD-level score falls to about 38-46%.
  • AI role: Aletheia, a math research agent built on Gemini Deep Think, generated, verified and revised proofs and used search, with varying autonomy across projects.

Evidence sources: https://deepmind.google/blog/accelerating-mathematical-and-scientific-discovery-with-gemini-deep-think/ · https://arxiv.org/abs/2602.03837

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Heat-resistant crops: a hybrid photosynthesis enzyme

At Michigan State University, AlphaFold predicted the 3D structure of glycerate kinase in a crop and a heat-loving alga. Replacing three flexible loops with rigid loops from the alga produced a hybrid that remained stable up to 65 °C. The enzyme has not yet been tested in whole plants.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: Hybrid glycerate kinase enzymes with three loops replaced by rigid loops from a heat-loving alga were tested in the lab, and one stayed stable up to 65 °C; whole-plant testing has not been done.
  • Still unknown: It is not yet known whether the hybrid enzyme improves heat resilience in a whole plant; the page does not say whether the AlphaFold-predicted structures were experimentally validated.
  • AI role: AlphaFold predicted the enzyme's 3D structures, which were then used in molecular simulations to identify loops to modify.

Evidence sources: https://deepmind.google/blog/engineering-more-resilient-crops-for-a-warming-climate/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

The structure of the LDL protein apoB100

At the University of Missouri, Zachary Berndsen and Keith Cassidy used AlphaFold to predict the structure of apoB100, the scaffold of LDL particles, and refined it using cryo-electron microscopy data. The result is a cage-like shell with a ribbon belt. The post gives no figures such as resolution and no clinical validation.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Cryo-EM images were refined with AlphaFold2 predictions; the resulting apoB100 structure was compared with more than 200 intramolecular cross-links and showed close agreement, with subnanometre resolution in most regions.
  • Still unknown: The article states that clinical applications will take time and the structure's implications are not yet realized; it gives no resolution figures or clinical validation, and the abstract states that resolution is subnanometre only in most regions.
  • AI role: AlphaFold2 generated the initial atomic-resolution structure predictions, which were refined against cryo-EM data.

Evidence sources: https://deepmind.google/blog/revealing-a-key-protein-behind-heart-disease/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-024-08467-w&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 48 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Healthier bees: AlphaFold predicts vitellogenin's structure

With the Norwegian University of Life Sciences, predicted structures of honeybee vitellogenin were confirmed by electron microscopy in July 2025; AlphaFold 3 helped study its interactions with other molecules. Early evidence suggests breeding times for resilient bees could fall from months to weeks, a still preliminary claim.

Technology: AlphaFold, AlphaFold 3

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: AlphaFold-predicted vitellogenin structures were confirmed by electron microscopy in July 2025, per the post; the linked paper reports a cryo-EM structure of native honey bee vitellogenin, and the breeding-time claim comes from early trials.
  • Still unknown: The breeding-time benefit (months to weeks) is preliminary and not cited to a publication, and implications for other egg-laying species are possibilities, not established findings.
  • AI role: AlphaFold 3 predicted the vitellogenin structures and helped study how it binds other molecules inside the bee.

Evidence sources: https://deepmind.google/blog/breeding-healthier-and-stronger-honeybees/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41467-025-58575-y&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 12 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaEvolve on 67 math problems with Terence Tao

Tested on 67 problems in analysis, combinatorics, geometry, and number theory, AlphaEvolve matched the best known solutions in most cases and improved them in several; in some cases it generalized finite instances into formulas valid for every input. Preprint not yet peer reviewed; source checked was only the team's article.

Technology: AlphaEvolve

Family AlphaEvolve: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The system was run on 67 problems in analysis, combinatorics, geometry and number theory and compared with the best known solutions. It rediscovered the best known solutions in most cases and improved several.
  • Still unknown: The abstract and page state no explicit limitations, and no verification details or failure cases are given.
  • AI role: An LLM-guided evolutionary coding agent proposed, tested and refined algorithmic constructions, sometimes combined with Deep Think and AlphaProof.

Evidence sources: https://arxiv.org/abs/2511.02864

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

C2S-Scale: a model-generated cancer hypothesis confirmed in cells

The 27-billion-parameter single-cell model, built on Gemma with Yale, predicted that silmitasertib would increase antigen presentation only in an immune-positive context. In human neuroendocrine cell models with low-dose interferon, the increase was about 50%. Validation is in cells only; the authors call this ‘a first step.’

Technology: C2S-Scale 27B (Gemma)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The model's prediction that silmitasertib boosts antigen presentation with low-dose interferon was tested experimentally in human neuroendocrine cell models not seen in training, with a roughly 50% synergistic increase reported.
  • Still unknown: The article calls this an early first step, validation is in cell models only, and clinical use is not shown.
  • AI role: The 27B single-cell language model C2S-Scale ran a dual-context virtual screen of over 4,000 drugs and generated the hypothesis.

Evidence sources: https://blog.google/technology/ai/google-gemma-ai-cancer-therapy-discovery/ · https://www.biorxiv.org/content/10.1101/2025.04.14.648850

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

New families of unstable singularities in fluid equations

Physics-informed neural networks trained to near machine precision revealed a new family of unstable singularities (blow-ups) in three fluid-dynamics equations, with a clear pattern between blow-up speed and instability order for the incompressible porous-medium and Boussinesq equations. Work with Brown, NYU, and Stanford: a step toward computer-assisted proofs, not a proof; the Navier–Stokes problem remains open.

Technology: Physics-informed neural networks

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Physics-informed neural networks trained to near machine precision found unstable singularities, and the accuracy was assessed numerically against the governing equations.
  • Still unknown: The source states the initial conditions must be tuned with extreme precision, the scaling formula is empirical, Navier-Stokes remains open, and supplementary information is not yet available.
  • AI role: Neural networks trained on the equation residuals, with a high-precision optimizer, located new families of unstable singularities, with humans supplying the mathematical insight.

Evidence sources: https://deepmind.google/blog/discovering-new-solutions-to-century-old-problems-in-fluid-dynamics/ · https://arxiv.org/abs/2509.14185

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Deep Loop Shaping reduces LIGO control noise

The method, described in Science and tested at the Livingston site, reduced noise 30–100-fold in the observatory's most unstable feedback loop, in collaboration with LIGO, Caltech, and GSSI. The result covers one control loop; extending it to all loops and adding hundreds of events per year are projections.

Technology: Deep Loop Shaping (reinforcement-learning control)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The controller was trained in simulation and then tested on the real LIGO Livingston system, where noise in the most unstable feedback loop fell 30 to 100 times and stayed stable over prolonged periods.
  • Still unknown: The source states gains come from one control loop, applying the method to all loops and extra detections per year are described as potential or projected.
  • AI role: A reinforcement learning controller with frequency-domain rewards replaced linear control design for the mirror feedback loop.

Evidence sources: https://deepmind.google/blog/using-ai-to-perceive-the-universe-in-greater-depth/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.adw1291&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 16 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaEarth Foundations: a 10-meter map of Earth

A model combines optical, radar, 3D laser, and climate-simulation data into one representation for 10-meter squares, using 16 times less storage and with 24% lower average error than the compared models. The Google Earth Engine dataset has over 1.4 trillion embeddings per year and has been tried by more than 50 organizations. Benefits with LLM agents remain prospective.

Technology: AlphaEarth Foundations

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Compared against traditional methods and other AI mapping systems on land-use identification and surface-property estimation; average error 24% lower than the models tested, storage 16 times lower. Over 50 organizations tested the dataset; no sample sizes given.
  • Still unknown: The page states no limitations; it says the team is still exploring time-based capabilities and that embeddings may be more useful combined with reasoning LLM agents.
  • AI role: The model turns satellite and Earth observation data into compact embeddings; scientists and partners use them to create maps and classify ecosystems.

Evidence sources: https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/ · https://arxiv.org/abs/2507.22291

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Aeneas: restoration, dating, and provenance of Latin inscriptions

The model restores texts with 73% top-20 accuracy for gaps up to ten characters (58% when length is unknown), attributes provenance with 72% accuracy across 62 Roman provinces, and dates inscriptions within 13 years; it was trained on over 176,000 inscriptions. In a study with 23 historians on 60 inscriptions, results were best when parallels and predictions were used together. Published in Nature.

Technology: Aeneas

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Corpus of 176,861 Latin inscriptions. Aeneas: 72% geographical accuracy over 62 provinces; dating within 13 years on average; restoration character error rate 40.5% with known length. 23 historians on 60 inscriptions, solo versus with Aeneas parallels and predictions.
  • Still unknown: The paper notes a small dataset, survival bias, only about 5% of inscriptions with images, and weaker results where data are scarce. Training data include earlier editors' restorations. Other languages are suggested, not tested.
  • AI role: Aeneas restored text, attributed provinces, estimated dates and retrieved parallels; historians evaluated outputs and combined them with their own knowledge.

Evidence sources: https://doi.org/10.1038/s41586-025-09292-5

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 37 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Gemini Deep Think reaches gold-medal level at the 2025 IMO

An advanced version of Gemini Deep Think solved 5 of 6 problems at the 2025 IMO, scoring 35 out of 42 and working in natural language within the 4.5-hour limit. IMO coordinators assessed and certified the score, while clarifying that they do not validate the system.

Technology: Gemini Deep Think

Family Gemini: profile · models page

Publication: Announcement only · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: direct test
  • Tested: An advanced Gemini Deep Think version solved the official 2025 IMO problems within the time limit, and IMO coordinators graded and certified the 35 of 42 score with the same criteria as for students.
  • Still unknown: The IMO did not validate the system or model, and this version is not yet public and was first given to trusted testers.
  • AI role: Gemini Deep Think worked end-to-end in natural language, exploring multiple solution paths in parallel.

Evidence sources: https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaGenome: predicting the effects of DNA variants

The model reads sequences up to 1 million base pairs and predicts thousands of gene-regulation properties at single-letter resolution. It outperformed the best external models in 22 of 24 sequence evaluations and matched or outperformed the best in 24 of 26 variant evaluations. Stated limits: distant regulatory elements; not validated for personal genomes or clinical use.

Technology: AlphaGenome

Family AlphaGenome: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: AlphaGenome was evaluated on sequence and variant-effect benchmarks against external models, outperforming in 22 of 24 and matching or exceeding in 24 of 26 evaluations, plus a TAL1 case study.
  • Still unknown: The source states it is not validated for personal genomes or clinical use, distant regulatory elements remain a challenge, and cell-specific patterns need improvement.
  • AI role: A model takes up to 1 million DNA base pairs and predicts thousands of genomic tracks and variant effects.

Evidence sources: https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/ · https://www.biorxiv.org/content/10.1101/2025.06.25.661532

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaEvolve: an agent for designing algorithms

An evolutionary agent based on Gemini and automated evaluators. It recovered an average of 0.7% of Google's computing resources, sped up a Gemini training kernel by 23%, and found an algorithm using 48 multiplications for 4×4 complex matrices. Across more than 50 open math problems, it matched the state of the art in about 75% and improved it in about 20%.

Technology: AlphaEvolve (Gemini Flash and Pro)

Family Gemini: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Results were checked by automated evaluators, and the agent was applied to over 50 open math problems and deployed tasks, matching state of the art in about 75% and improving it in about 20%.
  • Still unknown: The source states the approach applies only to problems whose solutions can be checked automatically, and broader access is not yet available.
  • AI role: An ensemble of Gemini models proposed code in an evolutionary loop with automated evaluators.

Evidence sources: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ · https://arxiv.org/abs/2506.13131

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

GenCast: ensemble weather forecasts with a diffusion model

A 0.25° model that outperformed ECMWF's ENS ensemble on 97.2% of 1,320 variable-and-lead-time combinations (99.8% beyond 36 hours) and produces a 15-day ensemble in about 8 minutes on a TPU. Physical models remain essential because they provide training data and initial conditions.

Technology: GenCast (diffusion model)

Family GenCast: profile · models page

Publication: Publication status unspecified in annotation · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Not reported
  • Test type: retrospective or benchmark
  • Tested: Forecasts were compared with ECMWF ENS on held-out 2019 data (trained up to 2018) across 1,320 variable and lead-time combinations; GenCast was more accurate on 97.2% of targets.
  • Still unknown: Forecasts are especially uncertain beyond a few days, the wind-power analysis is only a proof of principle, and physics-based models remain needed for training data and initial conditions.
  • AI role: A diffusion model generates an ensemble of 50 or more 15-day weather trajectories from the current weather state.

Evidence sources: https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1072 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaQubit: a neural decoder for quantum error correction

A Transformer-based decoder trained on data from a 49-qubit Sycamore processor made 6% fewer errors than tensor-network methods and 30% fewer than correlated matching in the largest experiments. It is too slow for real-time error correction on fast superconducting processors.

Technology: AlphaQubit (Transformer)

Family AlphaQubit: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: On Sycamore data, logical error per round was 2.901e-2 (distance 3) and 2.748e-2 (distance 5), versus 3.028e-2 and 2.915e-2 for a tensor-network decoder. Simulations to distance 11 (241 qubits); correlated matching was 1.25 to 1.4 times worse.
  • Still unknown: Throughput is slower than the 1 microsecond per round target. Training at distance 11 needs more data. Accuracy beyond distance 11 and logical computations remain future work. Larger-distance results are simulated.
  • AI role: A recurrent transformer neural network decodes surface-code syndromes to predict logical errors; it is the method under test.

Evidence sources: https://doi.org/10.1038/s41586-024-08148-8

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 195 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaProteo: AI-designed protein binders

A system for designing high-affinity protein binders, tested against seven targets, achieved 88% laboratory success on BHRF1 and bindings 3 to 300 times stronger than the best existing methods. It produced no valid binders for TNFα; strong binding is only a first step.

Technology: AlphaProteo

Family AlphaProteo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Binders were designed against seven target proteins and tested in the lab, with 3 to 300 times stronger binding and higher experimental success rates than the best existing methods.
  • Still unknown: Per-target success rates are not given on the abstract page and no limitations are listed.
  • AI role: A family of machine learning models for protein design that generates the binders.

Evidence sources: https://arxiv.org/abs/2409.08022

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaProof and AlphaGeometry 2: silver-medal level at the Math Olympiad

The two systems solved 4 of 6 problems at IMO 2024, scoring 28 of 42 points, one point below the gold threshold. Problems were manually translated into formal language, and some required up to three days of computation instead of the official 4.5 hours.

Technology: AlphaProof (Lean) and AlphaGeometry 2

Family AlphaGeometry: profile · models page

Publication: Peer-reviewed paper · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: prospective or independent check
  • Tested: The systems solved official IMO 2024 problems, graded under IMO rules by Prof. Timothy Gowers and Dr Joseph Myers, reaching 28 of 42 points; problems were manually translated into formal language first.
  • Still unknown: Two combinatorics problems remained unsolved, translation was manual, and some solutions took up to three days.
  • AI role: AlphaProof solved three problems with formal proofs, and AlphaGeometry 2 proved the geometry problem within 19 seconds of formalization.

Evidence sources: https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Original institutional announcement checked for the specific results, AI contribution and limitations cited in the antibiotic dossier or mathematical follow-up pilot. Full-paper reanalysis, external replication and an exhaustive follow-up search were not performed..

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 146 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships
  • publication · 12 Nov 2025: The updated original announcement dates publication of the AlphaProof methodology in Nature to 12 November 2025. This journal transition is not an independent replication and does not apply to every component of the combined system. Source. Recorded 16 Nov 2026.

Follow-up checked 16 Nov 2026 within scope: Bounded check of the original announcements cited by the events recorded here, including explicit links to later program work. This is not an exhaustive literature search or ongoing monitoring commitment..

Directed relationships from this entry; they do not merge catalog counts.

  • 2024-01-alphageometry · extends: AlphaGeometry 2 is explicitly described as an improved AlphaGeometry version; this link concerns the geometry component, not attribution of all four IMO solutions. Source

AlphaFold 3: structure and interactions of life’s molecules

It predicts joint structures of proteins, nucleic acids, small molecules and ions. DeepMind and Isomorphic Labs report at least 50% greater accuracy than existing methods for protein interactions with other molecules. Free for non-commercial research through the AlphaFold Server.

Technology: AlphaFold 3

Family AlphaFold 3: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Higher accuracy than docking tools for protein–ligand interactions, than nucleic-acid-specific predictors for protein–nucleic acid interactions, and than AlphaFold-Multimer v2.3 for antibody–antigen prediction.
  • Still unknown: Accuracy on new experimental targets beyond the benchmarks is not described in the abstract.
  • AI role: A diffusion-based architecture predicting joint structures of proteins, nucleic acids, small molecules, ions and modified residues.

Evidence sources: https://doi.org/10.1038/s41586-024-07487-w

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 17949 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaGeometry: Olympiad geometry without human demonstrations

A language model proposing constructions and a symbolic engine deriving proofs solved 25 of 30 problems in the IMO-AG-30 benchmark (the previous method solved 10; the average gold medalist score was 25.9), trained on 100 million synthetic examples. It covers geometry only and a 30-problem benchmark.

Technology: AlphaGeometry (language model and symbolic engine)

Family AlphaGeometry: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: On the IMO-AG-30 benchmark of 30 Olympiad geometry problems, the system solved 25 versus 10 for the previous method, with each solution checked by computer.
  • Still unknown: It covers geometry only, roughly a third of Olympiad problems, and approaches but does not match gold-medalist performance.
  • AI role: A language model proposes auxiliary constructions and a symbolic engine derives proofs, looping until a proof is found.

Evidence sources: https://deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review date: 16 Nov 2026. Review scope: Original institutional announcement checked for the specific results, AI contribution and limitations cited in the antibiotic dossier or mathematical follow-up pilot. Full-paper reanalysis, external replication and an exhaustive follow-up search were not performed..

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 494 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships
  • same-team extension · 25 Jul 2024: DeepMind describes AlphaGeometry 2 as an improved version used for the IMO 2024 geometry problem. The combined result also used AlphaProof; it does not belong to AlphaGeometry alone. Source. Recorded 16 Nov 2026.

Follow-up checked 16 Nov 2026 within scope: Bounded check of the original announcements cited by the events recorded here, including explicit links to later program work. This is not an exhaustive literature search or ongoing monitoring commitment..

No source-supported relationships recorded.

Google DeepMind

FunSearch: new mathematical results with a language model

A language model (PaLM 2) paired with an automatic evaluator found the largest cap sets on record in some dimensions, the biggest increase in 20 years, and bin-packing heuristics using fewer bins than established methods. The problem must first be expressed as code with an evaluation procedure.

Technology: FunSearch (PaLM 2 and automatic evaluator)

Family FunSearch: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Programs were scored automatically by an evaluator; on the cap set problem they found larger cap sets than the best known, and on online bin packing their heuristics beat established baselines.
  • Still unknown: The abstract does not explicitly state limitations; the blog says the method needs an automatic evaluation procedure and that the bin-packing heuristics were tailored to specific data.
  • AI role: A pretrained large language model (PaLM 2) generated candidate programs in an evolutionary search guided by an automatic evaluator.

Evidence sources: https://deepmind.google/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06924-6&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 990 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

GNoME: 2.2 million new predicted crystals

A graph neural network predicted 2.2 million new crystals, including 380,000 of the most stable, raising the number of known stable materials from about 48,000 to 421,000. External laboratories synthesized 736 predicted structures. Stability is calculated and does not guarantee a material can be produced or will be useful.

Technology: GNoME (Graph Networks for Materials Exploration)

Family GNoME: profile · models page

Publication: Peer-reviewed paper · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: prospective or independent check
  • Tested: Predicted stability was checked computationally with DFT, and external laboratories independently synthesized 736 of the predicted structures.
  • Still unknown: The blog states that most of the 380,000 stable predictions had not been made in a lab and that stability is judged by current scientific standards; manufacturing and usefulness remain open.
  • AI role: A graph neural network trained on about 48,000 stable crystals predicted 2.2 million new crystals, including 380,000 stable ones.

Evidence sources: https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-023-06735-9&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1876 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 4/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

GraphCast: global 10-day weather forecasts

It outperforms ECMWF’s HRES system on over 90% of 1,380 variables and test forecast lead times, and produces a 10-day forecast in under a minute on a single TPU. Comparison is limited to deterministic HRES; training uses ERA5 data.

Technology: GraphCast (graph neural network)

Family GraphCast: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: 10-day forecasts of hundreds of variables at 0.25° in under a minute; outperforms the most accurate operational deterministic systems on 90% of 1,380 verification targets; supports better tropical-cyclone, atmospheric-river and extreme-temperature prediction.
  • Still unknown: Comparison with ensemble systems is not described in the abstract.
  • AI role: A machine-learning model trained directly on reanalysis data.

Evidence sources: https://doi.org/10.1126/science.adi2336

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2507 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaZero finds improved finite extremal-graph bounds

A neural network-guided AlphaZero search with a curriculum generated triangle- and 4-cycle-free graphs that improve known lower bounds for the extremal function ex(n,{C3,C4}) over finite sizes; the proceedings report AlphaZero improvements for n=64–136 and release the graph constructions. The paper also finds improvements using a separate tabu-search method, which often matches or outperforms AlphaZero. This does not solve the asymptotic Erdős conjecture. The result was present in the arXiv v1 posted 2023-11-06; the 2024 IJCAI proceedings paper is a later version.

Technology: AlphaZero reinforcement learning with neural Pairformer representation and curriculum

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: The paper reports released graphs with improved lower bounds for ex(n,{C3,C4}) for node counts 64–134. Incremental AlphaZero (neural network-guided Monte Carlo tree search) is explicitly reported to improve bounds for sizes 64–136 in the proceedings version; incremental tabu search also improves bounds and the authors report it can match or exceed AlphaZero. This is a finite-graph construction result, not a proof of the asymptotic Erdős conjecture.
  • Still unknown: The result consists of explicit finite constructions/lower bounds; the asymptotic conjecture remains open. The work does not establish that learned search outperforms strong local search.
  • AI role: AlphaZero with a neural Pairformer representation and a curriculum generated extremal graphs improving the established lower bounds for n=64–136; tabu search is a separate, non-neural method and its results are not attributed to AI.

Evidence sources: https://arxiv.org/html/2311.03583v1 · https://ijcai.org/proceedings/2024/0772 · https://ijcai.org/proceedings/2024/0772.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 17 Nov 2026. Review scope: Primary arXiv v1 full text and metadata plus IJCAI 2024 proceedings page/PDF for AlphaZero versus tabu-search graph construction results, theorem limits and author affiliations..

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaMissense: cataloging missense genetic mutations

It classified 89% of the 71 million possible missense variants: 32% as likely pathogenic and 57% as likely benign, with 90% precision on a database of known variants. Predictions are not intended for direct clinical use.

Technology: AlphaMissense

Family AlphaMissense: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: The model was validated on genetic and experimental benchmarks, and Genomics England cross-referenced its predictions with previously aggregated human variant pathogenicity data and found them consistent. The reported comparison uses previously collected pathogenicity and variant data; an independent data source is distinct from a new independent experimental replication.
  • Still unknown: The model does not predict structural or stability effects, the blog says predictions are not designed for direct clinical use, and the abstract describes no clinical validation.
  • AI role: AlphaMissense, an AlphaFold adaptation fine-tuned on human and primate variant population frequency data, classified missense variants and released predictions for all possible human substitutions.

Evidence sources: https://deepmind.google/blog/a-catalogue-of-genetic-mutations-to-help-pinpoint-the-cause-of-diseases/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1126/science.adg7492&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 2381 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaDev: reinforcement learning discovers faster sorting routines

Working at the assembly-instruction level, the system found sorting routines up to 70% faster for sequences of 3–5 elements and about 1.7% faster for sequences over 250,000 elements; they were adopted in the LLVM libc++ standard library. A hashing algorithm 30% faster was added to Abseil. The largest gains are for short sequences.

Technology: AlphaDev (reinforcement learning)

Family AlphaDev: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: AlphaDev found sort 3 and sort 5 routines with fewer instructions than human benchmarks, matched sort 4, and had lower latency on variable sorts. Libc++ versions were up to 70% faster for five elements and about 1.7% faster above 250,000 elements.
  • Still unknown: Hashing result (30% faster) not seen in the text read. Search is at assembly level; the paper reports optimality proof only for sort 3 (17 instructions). Results cover tested CPUs: ARMv8, Intel Skylake, AMD Zen 2.
  • AI role: AlphaDev, an AlphaZero-based reinforcement learning agent with a transformer encoder, built assembly programs instruction by instruction, rewarded for correctness and latency; the team translated results to C++.

Evidence sources: https://doi.org/10.1038/s41586-023-06004-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 238 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaTensor: new algorithms for matrix multiplication

A reinforcement-learning agent found matrix-multiplication algorithms better than known ones for many sizes (for example, 76 multiplications instead of 80 for 4×5 by 5×5). Algorithms optimized for hardware were 10–20% faster on V100 GPUs and TPU v2. The improvement over Strassen applies in a finite field.

Technology: AlphaTensor (based on AlphaZero)

Family AlphaTensor: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Algorithms are provably correct by construction and their multiplication counts are compared with known methods; hardware-targeted algorithms were benchmarked on V100 GPU and TPU v2.
  • Still unknown: The page lists no formal limitations; it notes the search space is very large and the work is specific to matrix multiplication.
  • AI role: Reinforcement-learning agent searched for tensor decompositions giving new matrix-multiplication algorithms.

Evidence sources: https://deepmind.google/blog/discovering-novel-algorithms-with-alphatensor/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-022-05172-4&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 613 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold Database: over 200 million structures

The predicted-protein-structure archive, maintained with EMBL-EBI, grew from nearly 1 million to over 200 million structures, covering practically all cataloged proteins, with free access and bulk download from Google Cloud. More than 500,000 researchers in 190 countries had already used it. These are predictions, not experimental measurements.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: Test context unspecified in annotation. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: not reported
  • Tested: The post gives no accuracy metrics for the new structures; it cites two Nature AlphaFold papers and uses in experiments as evidence.
  • Still unknown: The post states no explicit limitations or caveats on prediction accuracy.
  • AI role: AlphaFold predicted protein structures added to the AlphaFold Database.

Evidence sources: https://deepmind.google/discover/blog/alphafold-reveals-the-structure-of-the-protein-universe/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 5/5 · Validation 1/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Ithaca: restoring and dating ancient Greek inscriptions

A neural network reconstructs missing text in damaged Greek inscriptions, identifies their provenance and dates them: 62% accuracy in restoration, 71% in location, and dating within 30 years of reference ranges. With Ithaca, historians improved from 25% to 72% correct restorations. It is intended to assist historians.

Technology: Ithaca (deep neural network)

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Restoration (62%), location (71%) and dating were checked on held-out dated inscriptions, and a historian study rose from 25% to 72% correct restorations with the tool.
  • Still unknown: Accuracy is far from perfect, dating is a range within about 30 years, and the model was trained only on Greek inscriptions.
  • AI role: Neural network restored missing text, identified provenance and dated Greek inscriptions.

Evidence sources: https://deepmind.google/discover/blog/predicting-the-past-with-ithaca/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-022-04448-z&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 268 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Tokamak plasma control with reinforcement learning

A single neural network controls the magnetic coils of Lausanne’s TCV tokamak and forms plasmas in different configurations (snowflake, ITER-like shape, double droplet). Trained mainly in simulation and then tested on the real machine. Stated limitations: plasma lasts about three seconds and the machine changes from day to day.

Technology: Deep reinforcement learning

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: On the TCV tokamak the controller produced and held elongated, conventional, negative-triangularity and snowflake configurations, with accurate tracking of location, current and shape, and sustained two separate 'droplet' plasmas.
  • Still unknown: Transfer to larger or reactor-scale devices is not reported in the abstract.
  • AI role: A deep reinforcement-learning architecture learned to command the full set of control coils from high-level objectives.

Evidence sources: https://doi.org/10.1038/s41586-021-04301-9

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 907 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

DM21: a neural network as a density-functional theory functional

A neural network represents the functional used in DFT and addresses two longstanding errors of traditional functionals: excessive electron delocalization and spin-symmetry breaking during bond breaking. Published in Science with open code; the post gives no benchmark figures and DFT remains an approximation.

Technology: DM21 (neural network for the exchange-correlation functional)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: DM21 is compared with traditional functionals on benchmarks for main-group atoms and molecules and on bond-breaking and charge-localisation examples.
  • Still unknown: The page gives no numerical benchmark figures and says DFT remains an approximation.
  • AI role: Neural network represents the exchange-correlation functional in density functional theory.

Evidence sources: https://deepmind.google/discover/blog/simulating-matter-on-the-quantum-scale-with-ai/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1126/science.abj6511&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 416 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AI-guided pure mathematics: knots and representation theory

With mathematicians Williamson (Sydney) and Lackenby and Juhász (Oxford), supervised-learning techniques revealed patterns leading to an unexpected link between a knot’s algebraic signature and its geometry (the ‘natural slope’), and to a new algorithm for the combinatorial invariance conjecture. The conjectures and proofs are the mathematicians’; the algorithm was checked on over 3 million examples, not proved in general.

Technology: Supervised learning and attribution techniques

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Conjectures and proofs were established by mathematicians, and the new algorithm was computationally verified on over 3 million examples.
  • Still unknown: Certain patterns still elude ML, and the representation-theory structure appears only at 100s or 1000s of vertices.
  • AI role: Supervised learning revealed patterns that led to the knot and representation-theory results and a new algorithm.

Evidence sources: https://deepmind.google/discover/blog/exploring-the-beauty-of-pure-mathematics-in-novel-ways/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-04086-x&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 551 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Enformer: gene expression predicted from DNA sequence

A Transformer-based model predicts gene expression from 200,000 base pairs of input DNA, over five times more than earlier methods, and is much more accurate at predicting variant effects. Enhancer attribution scores are comparable to methods using experimental data. The post gives no numerical gain and acknowledges that linking enhancers to the genes they regulate remains an open problem.

Technology: Enformer (Transformer)

Family Enformer: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Enformer is compared with prior models on variant-effect prediction and enhancer contribution scores against methods using experimental data, with a rs11644125 case study.
  • Still unknown: Linking enhancers to the genes they regulate is described as a major unsolved problem, and many disease variants are spuriously correlated.
  • AI role: Transformer model predicted gene expression from 200,000 base pairs of DNA.

Evidence sources: https://deepmind.google/blog/predicting-gene-expression-with-ai/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41592-021-01252-x&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1533 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Short-term rainfall prediction with a generative model

Developed with the UK Met Office, the model forecasts precipitation up to two hours ahead; over 50 expert meteorologists chose it as their first choice in 89% of cases compared with the most widely used methods. The authors cite long-range accuracy and rare, intense events as limitations.

Technology: Deep generative radar model (DGMR)

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Over 50 Met Office meteorologists chose the model first in 89% of cases, and it was compared with PySTEPS and UNet on UK and US radar data.
  • Still unknown: No method is without limitations; long-range accuracy and rare intense events need improvement.
  • AI role: Generative model forecast precipitation up to two hours ahead.

Evidence sources: https://deepmind.google/blog/nowcasting-the-next-hour-of-rain/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI%3A10.1038/s41586-021-03854-z&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 1232 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

AlphaFold Protein Structure Database: first 350,000 models

Predicted structures for the human proteome (about 20,000 proteins) and 20 other organisms were published, over 350,000 models in total, hosted free by EMBL-EBI. DeepMind announced an expansion to over 100 million structures. These are predictions that complement experimental methods. Note: the page header says July 22, 2022, but the text calls CASP14 “last December” and the 2022 post refers to it as a 2021 release; date corrected to 2021.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Predictions cover 98.5% of human proteins; 58% of residues have a confident prediction and 36% very high confidence. Case studies illustrate possible hypotheses.
  • Still unknown: The structures are predictions; the abstract does not report experimental confirmation of new structures.
  • AI role: AlphaFold 2 applied at proteome scale, with confidence metrics built on the model.

Evidence sources: https://doi.org/10.1038/s41586-021-03828-1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 3058 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 5/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

AlphaFold wins CASP14 for protein folding

AlphaFold achieved a median GDT score of 92.4 across the target set (about 1.6 Å mean error) and 87.0 on the hardest targets, which organizers judged a solution to the problem of single-protein structures. DeepMind notes that protein complexes and interactions with other molecules remain open challenges.

Technology: AlphaFold 2

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: prospective or independent check
  • Tested: Blind assessment at CASP14 on recently determined protein structures: median GDT of 92.4 across targets and 87.0 on the hardest free-modelling targets, judged by the organizers as a solution to single-protein structure prediction.
  • Still unknown: Protein complexes, interactions with DNA, RNA or small molecules, and full side-chain placement remain open; not every predicted structure will be perfect.
  • AI role: Attention-based neural network using evolutionary sequence alignments and residue-pair representations to predict 3D protein structure from sequence.

Evidence sources: https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 5/5 · Validation 4/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

FermiNet: quantum physics of molecules from first principles

FermiNet approximates molecular ground-state electron wavefunctions from first principles. Against the coupled-cluster reference, it recovered 99.8% of the correlation energy beyond a single Slater determinant for the smallest molecules and at least 97% for bicyclobutane (30 electrons). It also modeled stretched nitrogen and hydrogen chains. These are computational results from the original work announced on 19 October 2020. The source page was updated on 22 August 2024 with a separate excited-state study; that later study's carbon-dimer error is not a 2020 result.

Technology: FermiNet

Publication: Publication status unspecified in annotation · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Not reported
  • Test type: retrospective or benchmark
  • Tested: Energies compared with coupled-cluster and variational Monte Carlo reference calculations: 99.8% of the correlation energy beyond a single determinant recovered for the smallest molecules and at least 97% for bicyclobutane.
  • Still unknown: The coupled-cluster baseline scales poorly, so results apply to medium-sized molecules; the Monte Carlo method is approximate; excited-state results are within about 0.1 eV of the most demanding calculations rather than matching them.
  • AI role: Neural network wavefunction, antisymmetric in its inputs, trained by minimizing energy with Monte Carlo sampling using only nuclear positions.

Evidence sources: https://deepmind.google/blog/ferminet-quantum-physics-and-chemistry-from-first-principles/

The legacy label “Not reported” does not distinguish information absent from reviewed material from information not assessed. It is not a scientific finding of absence.

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 664 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Google DeepMind

Predicted structures of SARS-CoV-2 proteins

DeepMind openly shared AlphaFold predictions, without waiting for review, for six poorly studied viral proteins (membrane, ORF3a, Nsp2, Nsp4, Nsp6, and the C-terminal domain of papain-like protease). They were not experimentally verified; an ORF3a experimental structure deposited in the PDB in June 2020 matched the second predicted model well.

Technology: AlphaFold

Family AlphaFold: profile · models page

Publication: Announcement only · Test context: prospective or independent check. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Announcement only
  • Test type: prospective or independent check
  • Tested: Predictions compared with experimental structures: the spike prediction matched the PDB structure, and the second-ranked ORF3a model agreed well with later experimental work.
  • Still unknown: Predictions were not experimentally verified overall; the system was still in development and accuracy is not certain; released before peer review; the top ORF3a model misplaced some transmembrane helices.
  • AI role: AlphaFold free-modelling predicted structures for six understudied SARS-CoV-2 proteins, with per-residue confidence scores.

Evidence sources: https://deepmind.google/blog/computational-predictions-of-protein-structures-associated-with-covid-19/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

An AI breast-cancer screening system evaluated in two countries

Among 25,856 women in the UK and 3,097 in the US, the system reduced false positives by 5.7% (US) and 1.2% (UK), and false negatives by 9.4% and 2.7% in absolute terms compared with radiologists. Trained only on UK data, it also outperformed specialists on US data. Retrospective evaluation (Nature): prospective studies are needed, and clinical deployment is unclear.

Technology: Neural network for mammograms

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: retrospective or benchmark
  • Tested: Retrospective evaluation on UK (25,856 women) and US (3,097 women) screening mammograms, compared with individual radiologists' decisions; a reader study and a simulated double-reading workflow were also run.
  • Still unknown: The authors state prospective studies are required and that how to deploy the system clinically is not yet determined; the US data were enriched; the second-reader scenario was only simulated.
  • AI role: Predicted from screening mammograms alone whether a biopsy would later show cancer.

Evidence sources: https://deepmind.google/blog/international-evaluation-of-an-ai-system-for-breast-cancer-screening/ · https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:10.1038/s41586-019-1799-6&resultType=core&format=json

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: blog. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 3668 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Open the interactive profile for the timeline, or browse all organizations and model families.