Scientific claims from AI labs, measured against the knowledge that came before them. Explore the evidence behind the discovery.
Initial research sample. Provisional AI-assisted assessments, not independent expert verdicts. Missing claims are not zero discoveries.
REVIEWED 11 SEPT 2026You’re reading the published research snapshot. Interactive filters become available when the index loads. Claim titles below open their primary sources.
| DISCOVERY / CLAIM | ORGANIZATION | DOMAIN | NOVELTY ⓘ | CLAIM DATE | Details |
|---|---|---|---|---|---|
A smaller algorithm for complex matrix multiplicationEvidence & sourcesA bilinear algorithm multiplies two 4 × 4 complex matrices with 48 scalar multiplications. Provisional assessment: A new construction beyond the documented 49-multiplication baseline. The numerical gain is small, but the construction improves a long-standing bound in this specific setting. This does not establish a new matrix-multiplication exponent or a universal speedup. Prior art: Strassen's recursive construction ↗ · 1969 Applying Strassen's 2 × 2 construction twice gives 49 multiplications for 4 × 4 matrices. AlphaTensor's 47-multiplication binary-field result is not the same setting. Review limits: Provisional assessment from the announcement and disclosed result. We have not independently run symbolic verification or exhaustively searched all bilinear constructions.
| Google DeepMind AlphaEvolve / Gemini | Mathematics | N2New result Medium confidence · provisional | 14 May 2025 | ↗ |
Larger cap-set constructionsEvidence & sourcesFunSearch produces improved cap-set constructions, including a set of size 512 in dimension eight. Provisional assessment: The paper documents new constructions for a well-studied extremal problem. This is new mathematical output within an existing problem and search framework, rather than a new field or a solution of the general cap-set problem. Prior art: Edel and Bierbrauer; generalized product caps ↗ · 2004 Earlier finite cap constructions and product-cap methods provide the comparison framework in the FunSearch paper. A full reconstruction of the record history remains outstanding. Review limits: Prior-art coverage is partial. We rely on the paper's comparison and have not reproduced its programs or its asymptotic construction. | Google DeepMind FunSearch / Codey | Mathematics | N2New result Medium confidence · provisional | 14 Dec 2023 | ↗ |
A 47-multiplication construction over the binary fieldEvidence & sourcesAlphaTensor finds a 4 × 4 matrix-multiplication algorithm requiring 47 scalar multiplications in arithmetic modulo two. Provisional assessment: A new construction for the specified field. It advances the documented baseline within an established tensor-decomposition framework; its novelty should not be generalized to ordinary real or complex arithmetic. Prior art: Strassen's recursive construction ↗ · 1969 Recursive use of the seven-product algorithm requires 49 scalar multiplications. AlphaTensor's improvement here relies on binary-field arithmetic. Review limits: We have not rerun the released constructions. Subsequent algorithms do not change the novelty assessment at the 2022 cutoff.
| Google DeepMind AlphaTensor | Computer science | N2New result Medium confidence · provisional | 05 Oct 2022 | ↗ |
A mixed sodium–lithium solid electrolyteEvidence & sourcesAI-assisted screening leads to a mixed sodium–lithium solid electrolyte, followed by experimental validation with PNNL. Provisional assessment: Provisionally an incremental compositional extension in an established family of solid electrolytes. Screening scale and development speed measure the discovery process, not the material's novelty or superiority. Prior art: Asano et al., solid halide electrolytes ↗ · 14 Sept 2018 High-conductivity halide solid electrolytes for solid-state batteries were already demonstrated. This establishes family-level precedent, not an exact structural match to the new composition. Review limits: Low-confidence rating. Exact composition matching, patents, and a standardized performance comparison could change this classification. No percentage novelty or performance gain is assigned.
| Microsoft Azure Quantum Elements / materials AI | Materials science | N1Incremental Low confidence · provisional | 08 Jan 2024 | ↗ |
Generated inhibitors of the known DDR1 targetEvidence & sourcesGENTRL generates candidate DDR1 inhibitors, with synthesized compounds tested experimentally. Provisional assessment: An incremental medicinal-chemistry result on an established target. Existing selective DDR1 inhibitors and related scaffolds predate the work. Similarity is evidence of close prior art, not proof of copying or an identical molecule. Prior art: Selective DDR1 inhibitors and prior kinase scaffolds ↗ · 10 Apr 2013 Gao and colleagues had reported selective, orally bioavailable DDR1 inhibitors in 2013. Bender's contemporaneous analysis found close chemical neighbors for a generated compound in ChEMBL. Review limits: We have not computed molecular fingerprints or run a patent search. Bender's 75% search threshold is not a measured universal novelty score and is not reported as one. Assay results from different studies are not directly compared.
| Insilico Medicine GENTRL | Chemistry | N1Incremental Medium confidence · provisional | 02 Sept 2019 | ↗ |
A formalization of Fermat's Last TheoremEvidence & sourcesAnthropic reports a complete Lean formalization of Fermat's Last Theorem, following existing mathematical proofs. Provisional assessment: The theorem is a known result. The potentially new contribution is its formal verification artifact and the automation process. N0 applies only to discovery of the theorem, not to the value or novelty of formalization. Prior art: Wiles; Darmon, Diamond, and Taylor ↗ · 1995 Wiles proved the theorem in work published in 1995. Anthropic explicitly says its formalization follows the later exposition by Darmon, Diamond, and Taylor. Review limits: Method-context record, excluded from discovery rankings. This is acknowledged reuse, not a plagiarism allegation. The formalization's validity and method novelty have not been independently audited here.
| Anthropic Claude / Prove2Me | Mathematics | N0Known result High confidence · provisional | 04 Sept 2026 | ↗ |
A proposed resolution of the Navier–Stokes problemEvidence & sourcesOpenAI announces a proposed solution of the Navier–Stokes Millennium Prize problem and releases a paper and a formal proof artifact. Provisional assessment: A potentially major result requiring independent specialist assessment. The closest proof strategy, theorem assumptions, and priority relative to concurrent work have not been resolved by this initial review. Prior art: Tao's averaged Navier–Stokes blowup construction ↗ · 03 Feb 2014 Tao established finite-time blowup for a modified, averaged equation. That result is relevant background but is not a solution for the original Navier–Stokes equations. Review limits: No proof verification or adjudication of priority was performed. Concurrent results for different equations must not be treated as identical. No claim of copying is made.
| OpenAI Internal research model | Physics | —Unassessed Review incomplete | 08 Sept 2026 | ↗ |
A claimed improvement in the zeta-zero proportionEvidence & sourcesAnthropic reports raising a lower bound for the proportion of Riemann-zeta zeros on the critical line from 41.6% to 67.2%. Provisional assessment: Potentially a new result combining existing mathematical tools. The cited numerical baseline, theorem assumptions, and full proof require specialist review before this leaderboard assigns novelty credit. Prior art: Unconditional pair-correlation results ↗ · 08 Jun 2023 Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh developed an unconditional Montgomery theorem. The announcement also credits Aryan and Bombieri. These are stated ingredients, not a verified latest numerical record. Review limits: We have not checked the Lean artifact, its assumptions, or the complete prior-art literature. This is not a claimed proof of the Riemann hypothesis itself.
| Anthropic Unreleased research Claude | Mathematics | —Unassessed Review incomplete | 10 Aug 2026 | ↗ |
A claimed disproof of the unit-distance conjectureEvidence & sourcesOpenAI reports a construction giving a polynomial improvement over classical grid examples for the planar unit-distance problem. Provisional assessment: A potentially major new result, but this initial review has not compared the proof with the closest constructions or independently established its quantitative improvement. It remains unassessed rather than receiving a score from the announcement alone. Prior art: Classical grid constructions and incidence bounds ↗ · 21 Jul 2025 The existing literature distinguishes lower-bound constructions from upper bounds on the maximum number of unit distances. A 2025 paper provides pre-claim context; an upper bound is not an interchangeable baseline for a new lower bound. Review limits: Independent specialists are reported as checking the result, but this leaderboard has not verified the proof or its exact prior-art distance. No numerical improvement is inferred from unmatched upper and lower bounds.
| OpenAI Unreleased reasoning model | Mathematics | —Unassessed Review incomplete | 20 May 2026 | ↗ |
A generated study of compositional regularizationEvidence & sourcesAn AI-generated manuscript reports negative results for compositional regularization and receives workshop review scores above the stated acceptance threshold. Provisional assessment: Review scores assess a manuscript, not its distance from prior art. The specific negative result needs comparison with earlier regularization and compositional-generalization experiments before a novelty rating can be assigned. Prior art: Earlier compositional-generalization and regularization research ↗ · Unresolved The closest experiment has not been identified in this initial review. The earlier AI Scientist framework is process context and does not establish priority for this scientific result. Review limits: The paper was withdrawn under the experiment protocol; the workshop did not perform a final meta-review. This record does not describe an accepted main-conference publication.
| Sakana AI The AI Scientist-v2 | Computer science | —Unassessed Review incomplete | 12 Mar 2025 | ↗ |
A large catalogue of predicted stable crystalsEvidence & sourcesGNoME reports 381,000 new entries on an updated computational convex hull, within a larger set of predicted crystal structures. Provisional assessment: A substantial catalogue expansion is documented. A blanket novelty rating would conceal differences between new compositions, new structures, and variants of known prototypes. Entry-level matching against announcement-date databases is required. Prior art: Materials Project and OQMD catalogues ↗ · 2021-03 The study starts from March 2021 database snapshots. These are useful baselines, but leave a gap before the November 2023 public claim. Review limits: No structure-by-structure deduplication, patent search, or synthesis validation was performed here. Computational stability is distinct from experimental synthesizability.
| Google DeepMind GNoME | Materials science | —Unassessed Review incomplete | 29 Nov 2023 | ↗ |
Learning to control tokamak plasma shapesEvidence & sourcesA learned controller produces and maintains multiple plasma configurations on EPFL's TCV tokamak. Provisional assessment: An experimentally demonstrated control-method advance. It is tracked as enabling research, not counted as a new physical law or a separate discovery for each plasma configuration. Prior art: Conventional TCV magnetic feedback control ↗ · 2021 Model-based control and the study of shaped plasmas predate this work. The Nature paper describes the conventional controller architecture it replaces and cites the preceding control literature. Review limits: Research-method context record; excluded from discovery rankings. A separate controller-novelty review would require comparison with earlier control architectures.
| Google DeepMind Deep reinforcement learning | Physics | —Unassessed Review incomplete | 16 Feb 2022 | ↗ |
Predicted structures across the human proteomeEvidence & sourcesAlphaFold expands structural coverage across the human proteome, reporting confident predictions for 58% of residues. Provisional assessment: Expanded predictive coverage can enable discoveries, but does not establish that every prediction is a new biological finding. Each structure needs its own comparison to experimental structures, templates, and earlier predictions. Prior art: Experimental structures and earlier AlphaFold predictions ↗ · 2020 The study reports experimental coverage of 17% of human-protein residues. Earlier structure-prediction methods already existed, including the CASP13 AlphaFold system. Review limits: This record aggregates a catalogue and receives no blanket novelty score. Predictions and experimental observations are explicitly distinguished.
| Google DeepMind AlphaFold 2 | Biology | —Unassessed Review incomplete | 22 Jul 2021 | ↗ |
An AI-designed antifibrotic drug candidateEvidence & sourcesInsilico reports an AI-selected antifibrotic target and AI-designed candidate, later disclosed as the TNIK inhibitor rentosertib. Provisional assessment: TNIK inhibition was known before the claim. The possible novelty lies in the particular compound and its antifibrotic use. Neither target novelty nor chemical novelty can be resolved from the early undisclosed-target announcement alone. Prior art: NCB-0846 and earlier TNIK inhibition ↗ · 26 Aug 2016 Masuda and colleagues published a small-molecule TNIK inhibitor in cancer research in 2016. A 2018 study also examined TNIK inhibition in an epithelial–mesenchymal transition context. Review limits: A dated patent and chemical-structure review is outstanding. Later clinical studies strengthen evidence for the candidate but cannot retroactively establish novelty at the 2021 cutoff. Clinical efficacy is not rated here.
| Insilico Medicine PandaOmics / Chemistry42 | Medicine | —Unassessed Review incomplete | 24 Feb 2021 | ↗ |
Try another search or broaden your filters.
Ranked by the number of discovery claims provisionally assessed N2 or N3 in this sample. Unassessed claims earn no novelty credit. Research-method records are shown separately.
| RANK | ORGANIZATION | N2 + N3 | INCREMENTAL | KNOWN | UNASSESSED | CLAIMS | METHODS |
|---|---|---|---|---|---|---|---|
| 01 | 5 domains | 3 | 0 | 0 | 2 | 5 | 1 |
| — | 1 domain | 0 | 0 | 0 | 1 | 1 | 1 |
| 02 | 2 domains | 0 | 1 | 0 | 1 | 2 | 0 |
| 02 | 1 domain | 0 | 1 | 0 | 0 | 1 | 0 |
| — | 2 domains | 0 | 0 | 0 | 2 | 2 | 0 |
| — | 1 domain | 0 | 0 | 0 | 1 | 1 | 0 |
The full dataset is used here, regardless of index filters. Ties share a rank; organizations with no assessed discovery claims are unranked. Counts do not measure overall research quality, resources, or discovery rate.
We evaluate the novelty of a precisely scoped result relative to what was publicly known at its claim date. Novelty, correctness, importance, and the contribution of AI are separate questions.
Record the result, organization, model, collaborators, and earliest public disclosure we can establish. Use preprint v1 over a later journal date when available. Dates based on later accounts are marked explicitly. The public claim date is a proxy, not proof of the internal discovery time.
Search named results, synonyms, cited predecessors, and competing methods. Record the nearest documented match and whether it matches the exact task, just the family, or only the broader context. Prior art must predate the cutoff; later critiques can identify earlier work but do not become earlier discoveries.
Separate a new instance, a new construction, a new theorem, and a new principle. When results share assumptions and units, show the raw before-and-after values and the calculation. Otherwise state what differs and leave the numeric distance unassigned.
Attach source links, search queries, confidence, evidence type, and review gaps. All current judgments are provisional AI-assisted editorial assessments. “Unassessed” means insufficient review; it does not mean a result lacks novelty.
The same scientific result is documented before the cutoff. Rediscovery and formalization may still be valuable. This label does not allege copying.
A localized extension, parameter improvement, or new variant within a closely established result family.
A distinct construction, finding, or solution beyond the identified prior art in an established problem. The conceptual framework may be familiar.
A supported new principle or framework with a substantial conceptual departure. Requires broad prior-art review. No initial record meets that review bar.
For compatible numerical quantities, absolute change = result − baseline. Relative change = (result − baseline) / |baseline| × 100%, where the baseline is nonzero. A percentage-point change is distinct from a relative percentage change. Fewer multiplications do not automatically mean faster software.
We do not compute a weighted average of ordinal levels or treat them as equally spaced. Molecular similarity, proof structure, performance, and conceptual novelty need different measures. No embedding-based “novelty percentage” is presented as scientific fact.
Count one record per scoped claim, assign one primary domain, and attribute it to one lead organization. Academic collaborators remain credited in the record. A catalogue of predicted materials counts as one catalogue claim, not thousands of proven discoveries. A later trial or publication updates evidence for the existing claim instead of adding a discovery.
The leaderboard counts only discovery records. Research methods and formalizations are context records. Rank by the count of N2/N3 claims; use competition ranks for ties. An organization with assessed claims but zero N2/N3 can rank below those with credited claims; an organization with only unassessed claims has no rank.
High: the narrow novelty judgment has a direct, explicit match or strong documentary support. Medium: the comparison is supported, but reproduction or wider search remains incomplete. Low: nearest prior art, disclosure dates, exact comparability, or proof review is unresolved. These are editorial labels, not calibrated probabilities.
Curated initial sample of 14 records across 6 organizations. This is not an exhaustive census or a systematic literature review. Missing organizations and missing claims are not evidence of no discoveries. Rankings describe only this sample.
Initial research uses targeted public web searches and primary publications where accessible. Search queries are preserved per record; original reference checks are incomplete where stated. There is no continuous monitoring, exhaustive patent search, complete citation-graph crawl, or independent expert review in this release.
Source pages may change; links are not immutable archival snapshots. Versioned arXiv links are used where possible. Date precision is retained rather than inventing exact dates. New evidence can revise a rating in either direction.
Supply a precise claim, a dated primary source, and the closest earlier result. To challenge an assessment, explain the difference in assumptions, structures, or conclusions. We require specific evidence before attributing copying or misconduct.
Email the evidence ↗Know an earlier result, a missing claim, or a sharper comparison?