Zhao Song and Xinzhi Zhang prove a universal, dimension-independent bound for the Kannan–Lovász–Simonovits constant of isotropic log-concave probability measures, resolving the KLS conjecture. Their proof repeatedly refines bounds on polynomial variance and curvature; the authors report that AI tools helped explore more than 100 approaches, obtain the earlier log-star bound, and develop and verify the final proof. Independent researchers also posted proofs, and the paper’s streamlined proof is being formalized in Lean. The 1 October preprint gave a slowly dimension-dependent bound; the 4 October version established the dimension-independent theorem described here.
Technology: GPT-6 Astra; GPT-5.6 Sol; Claude Fable 5; Fable (as named in the paper’s AI disclosure)
Family Claude: profile · models page
Publication: Preprint · 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: Preprint
- Test type: prospective or independent check
- Tested: Zhao Song and Xinzhi Zhang prove a universal, dimension-independent bound for the Kannan–Lovász–Simonovits constant of isotropic log-concave probability measures, resolving the KLS conjecture.
- Still unknown: The proof is a recent arXiv preprint, not yet journal peer-reviewed in the source reviewed. Independent proofs were posted, and the Microsoft Research author account reports later Lean formalization of a streamlined proof; the paper version and posted revisions should be checked as the work develops. This entry counts the theorem result once, not the intermediate log-star bound separately.
- AI role: The authors state that they explored candidate approaches collaboratively with AI tools; the tools helped obtain the earlier log-star bound and develop detailed proof steps, while the researchers selected the direction and completed the repeated refinement argument. The authors state they checked the proofs with AI tools and take responsibility for correctness.
Evidence sources: https://arxiv.org/pdf/2610.01447v2 · https://arxiv.org/abs/2610.01447 · https://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/
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 scope and date are not recorded; source checked does not establish full-text review.
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.
Microsoft's Quine, a multimodal biology world model, prioritized thousands of compounds for shifting tumor cell states, and a handful were tested in pancreatic ductal adenocarcinoma cell lines. The top-ranked compounds produced the largest classical-to-basal shifts, as validated by wet-lab experiments, and some moved cells toward a distinct third phenotype. The post gives no quantitative wet-lab figures and links no paper, and it measures cell-state changes rather than clinical efficacy.
Technology: Quine, multimodal biology world model with interactive harness
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: A handful of top-ranked compounds predicted by Quine were tested in patient-derived pancreatic ductal adenocarcinoma cell lines, where the largest classical-to-basal transcriptional shifts were reported.
- Still unknown: The page calls the drug-repurposing findings early evidence and does not report how many compounds were tested or the assay statistics.
- AI role: Quine, a multimodal biology world model with a harness, ranked thousands of compounds by their predicted ability to shift tumor cell states.
Evidence sources: https://microsoft.com/en-us/research/blog/introducing-quine-an-ai-research-system-designed-for-the-complexity-of-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 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.
Published in Nature, the method combines a Transformer model (R-SMILES 2) and a template-based graph model (NeuralLoc) through a learned vote. In blind tests with expert chemists, 90% of complete routes were accepted, compared with 20–50% for the component models or a baseline, and nine of ten difficult targets were solved. Evaluation covered only ten targets; one wrong prediction invalidates an entire route.
Technology: RetroChimera (Transformer and template-based model)
Family RetroChimera: 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: Blind tests by PhD-level chemists rated individual steps and complete routes for ten difficult targets: 90% of RetroChimera routes were accepted, and nine of ten targets were solved, versus 20-50% for the component models or baseline.
- Still unknown: The route-level evaluation covers only ten targets; one incorrect prediction invalidates a route; NeuralLoc is constrained on reactions outside its template library, and R-SMILES 2 may hallucinate.
- AI role: An ensemble of a Transformer (R-SMILES 2) predicting precursors and a graph neural network (NeuralLoc) selecting templates, combined by a learned rank-dependent vote.
Evidence sources: https://www.microsoft.com/en-us/research/blog/improving-synthesis-prediction-of-small-molecules-at-scale-with-retrochimera/
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 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.
A candidate identified with MatterSim-v1, tetragonal tantalum phosphide, was synthesized and showed thermal conductivity of 152 W/m/K in the best samples, close to silicon. MatterSim-v1 inference is 3–5 times faster, and the new multi-task model MatterSim-MT predicts energies, forces, and properties. Battery and other property results are predictions, not experimental confirmations.
Technology: MatterSim-v1 and MatterSim-MT
Family MatterSim: 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: A MatterSim-v1 screen of over 240,000 candidates picked tetragonal tantalum phosphide, which was synthesized at UT Dallas and whose thermal conductivity was measured at UIUC, reaching 152 W/m/K in the best samples; the page does not describe the measurement method or number of samples.
- Still unknown: The page states that practical constraints such as mass density, elemental abundance and environmental stability are needed, but does not say whether TaP meets them.
- AI role: MatterSim-v1 machine-learning interatomic potentials estimated phonon thermal conductivity to screen candidates before experimental validation; MatterSim-MT is a multi-task model.
Evidence sources: https://www.microsoft.com/en-us/research/blog/advancing-ai-for-materials-with-mattersim-experimental-synthesis-faster-simulation-and-multi-task-models/
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.
CleaveNet combines a generative model conditioned on protease cleavage profiles with a predictor that ranks candidate sequences. In an in-vitro screen of 95 peptide substrates against 12 matrix metalloproteinases, the authors report that designed substrates were cleaved by MMP13 and identify substrates with different efficiency/selectivity trade-offs. The panel includes training controls and baseline candidates; it is not 95 novel AI discoveries. These are laboratory substrates, not validated diagnostic or therapeutic products.
Technology: CleaveNet generator and predictor; deep 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: The paper reports the specific in-vitro protein assay summarized here.
- Still unknown: Clinical or diagnostic deployment and independent replication are not established.
- AI role: A generator proposes peptide sequences conditioned on cleavage profiles; a model ranks candidates and researchers test synthetic substrates against recombinant enzymes.
Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12775482/
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 date: 16 Nov 2026. Review scope: Full primary journal article text, results and experimental assays reviewed. The preprint version-specific claim correspondence was not checked..
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.
It comes close to the 1 kcal/mol threshold on the W4-17 benchmark (0.85 kcal/mol on the single-reference subset) and is competitive with leading hybrid functionals on GMTKN55 at lower cost. It was trained on about 150,000 high-accuracy energy differences, roughly 100 times the previous dataset size. Accuracy is shown only within the training data's chemical space; it was not yet public at the time.
Technology: Skala (neural network for DFT)
Family Skala: 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: Skala was evaluated on the W4-17 atomization benchmark, reaching 0.85 kcal/mol on the single-reference subset, and on GMTKN55, where it was competitive with hybrid functionals.
- Still unknown: Accuracy is claimed only within the training data chemical space, expansion to broader chemistry is at an early stage, and the functional was not yet public.
- AI role: A deep-learning exchange-correlation functional learned from about 150,000 high-accuracy energy differences.
Evidence sources: https://www.microsoft.com/en-us/research/blog/breaking-bonds-breaking-ground-advancing-the-accuracy-of-computational-chemistry-with-deep-learning/ · https://arxiv.org/abs/2506.14665
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.
An open-source model that generates thousands of protein structures per hour on a GPU and reproduces the equilibrium distribution of a molecular-dynamics simulation using 10,000 to 100,000 times fewer GPU-hours. Predicted folding free energies correlate well with experimental values. The authors call it ‘a first step.’
Technology: BioEmu-1
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: Generated ensembles were compared with a molecular dynamics simulation of protein G using 10,000 to 100,000 times fewer GPU hours, and predicted folding free energies were correlated with experimental values.
- Still unknown: The authors state they are aware of limitations but the post does not list them, and it is presented as an early step.
- AI role: A generative deep learning model produces thousands of protein structures per hour on a GPU.
Evidence sources: https://www.microsoft.com/en-us/research/blog/exploring-the-structural-changes-driving-protein-function-with-bioemu-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.
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.
A diffusion model trained on 608,000 stable materials. Compound TaCr2O6, generated with a constraint of 200 GPa bulk modulus, was synthesized and measured at 169 GPa (relative error under 20%), but with Ta–Cr disorder differing from the proposed structure. One experimental example.
Technology: MatterGen (diffusion model)
Family MatterGen: 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: The generated TaCr2O6 structure was synthesized and its bulk modulus measured at 169 GPa against the 200 GPa target, with relative error under 20%.
- Still unknown: The source reports one experimental example, with Ta-Cr disorder differing from the proposed structure, and cites no separate publication for this compound.
- AI role: MatterGen, a diffusion model conditioned on a 200 GPa bulk modulus, generated the structure.
Evidence sources: https://www.microsoft.com/en-us/research/blog/mattergen-a-new-paradigm-of-materials-design-with-generative-ai/
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, 933 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.
A 1.3-billion-parameter model for high-resolution weather and air-pollution forecasting: it matches or exceeds CAMS on 74% of targets and matches or beats GraphCast on 94%. The roughly 5,000-fold speedup over the IFS system is an estimate; GraphCast remains slightly better up to five days.
Technology: Aurora (foundation model)
Family Aurora: 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 CAMS for air pollution (Jun-Nov 2022, matched or beat 74% of targets) and with GraphCast (matched or beat 94% of targets) over 2022 evaluation periods.
- Still unknown: GraphCast was slightly better on specific humidity up to five days, and the page has no dedicated limitations section.
- AI role: A 1.3-billion-parameter foundation model, pretrained on over a million hours of simulations, forecasts weather and air pollution at 0.1 degrees.
Evidence sources: https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/
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.
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.
A diffusion model generates stable, novel materials directly from desired properties: 2.9 times more stable, novel and unique structures than CDVAE, and over 250 materials with bulk modulus above 400 GPa versus 2 in the reference dataset. Results were checked with DFT calculations; experimental verification remains the ultimate test.
Technology: MatterGen (diffusion model)
Family MatterGen: 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: Generated structures were checked with DFT calculations and compared with CDVAE and screening baselines (2.9 times more stable, novel and unique structures than CDVAE, per the blog); experimental verification is stated as not yet done.
- Still unknown: Limitations stated: DFT has known limitations and experimental verification remains the ultimate test; the arXiv abstract states no limitations.
- AI role: A diffusion model generated new crystal structures conditioned on desired properties.
Evidence sources: https://arxiv.org/abs/2312.03687 · https://www.microsoft.com/en-us/research/blog/mattergen-property-guided-materials-design/
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.