Alman and Vassilevska Williams present deterministic algorithms for 3SUM on polynomial-size integers in O(n^1.9992) time and all-pairs shortest paths on directed graphs with polynomially bounded integer weights in O(n^2.9995) time, the first polynomial improvements over the textbook exponents. The single underlying technique computes selected entries of thin matrix products more efficiently, and known reductions yield the main algorithmic consequences. The paper says an internal Anthropic Claude research model discovered the algorithm autonomously; the human authors then understood, simplified, strengthened and extended it. One paper-level theoretical-computer-science result is counted.
Technology: Internal Anthropic Claude research model (specific model version undisclosed)
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: Alman and Vassilevska Williams present deterministic algorithms for 3SUM on polynomial-size integers in O(n^1.9992) time and all-pairs shortest paths on directed graphs with polynomially bounded integer weights in O(n^2.9995) time, the first polynomial improvements over the textbook exponents.
- Still unknown: This is an arXiv preprint as of the checked record. The stated time bounds are for polynomially bounded integer inputs; real-valued variants use Las Vegas randomization and different bounds. Lean certification covers the explicitly enumerated theorems, not every statement or presentation detail in the 76-page paper. The authors take responsibility for the paper and developed important extensions/consequences after the initial AI-discovered algorithm.
- AI role: The paper states that Claude was tasked with checking and improving cryptographic constructions but instead developed the new thin-matrix-product algorithm, first for average-case and then worst-case problems, in a 16-million-output-token session with no human input. The authors subsequently worked to understand, simplify, strengthen and extend the algorithm. Anthropic later used an internal model to Lean-certify specified main results.
Evidence sources: https://arxiv.org/abs/2610.06783
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 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.
In just under four weeks, Claude optimized over 30 models for structure, protein design, and genomics. They were on average four times faster with a small accuracy loss, and nearly twice as fast with identical outputs. It also created a low-memory mode for systems over 10,000 tokens on a single GPU node. Design results are in silico only, and comparisons are internal to Anthropic.
Technology: Claude (in Claude Science)
Family Claude: 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: Accelerated models were checked against their downstream tasks (for example structure prediction) and FlashPairformer kernels were compared with cuEquivariance (2.7-2.9x faster on triangle attention); design results are in silico with internal comparisons.
- Still unknown: Design results have no wet-lab validation, large viral capsid predictions were incorrect, ColabFold kernels were not benchmarked, and the model is not reproducible.
- AI role: Claude wrote custom FlashPairformer kernels, optimized more than 30 models, built a low-memory Big mode, and ran autonomous protein binder design.
Evidence sources: https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling
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 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.
Using the open-source OpenEvolve LLM-driven evolutionary framework with a custom reward, the authors find exact values Z(11,21,3,3)=116, Z(11,22,3,3)=121 and Z(12,22,3,3)=132, and lower bounds for 41 further parameter sets, at under $30 per case. Caveats: arXiv preprint not peer reviewed; a separate NeurIPS 2026 listing by overlapping authors reports different counts.
Technology: OpenEvolve (LLM-guided evolutionary search, AlphaEvolve-style)
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 exact values of three Zarankiewicz numbers and lower bounds for 41 further parameter sets were obtained by OpenEvolve search with a tailored reward.
- Still unknown: The page does not describe how the exact values or bounds were verified.
- AI role: OpenEvolve, an open-source LLM-based evolutionary framework, searched for mathematical constructions at under $30 per parameter combination.
Evidence sources: https://arxiv.org/abs/2605.01120
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.
A transformer trained to predict minimum Hamming distance, paired with a genetic algorithm and exact Brouwer–Zimmermann distance checks, found at least six previously unknown champion generalized toric codes over F₈ at dimensions 18, 22 and 38 (length 49). The article reports more than 500 champion codes in its F₈ search and up to a twofold reduction in exact distance evaluations versus random search. The six records at the three newly improved parameter settings are one paper-level coding-theory result.
Technology: Transformer model adapted from GPT-2, coupled with a genetic algorithm
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: A transformer trained to predict minimum Hamming distance, paired with a genetic algorithm and exact Brouwer–Zimmermann distance checks, found at least six previously unknown champion generalized toric codes over F₈ at dimensions 18, 22 and 38 (length 49).
- Still unknown: The codes are newly discovered mathematical constructions in the generalized toric family, with minimum distances verified computationally using established exact algorithms and compared to known records. The reported >500 total includes known/previous champion codes, while six at dimensions 18, 22 and 38 are identified as novel; these are not six separate catalog entries. The method searches a restricted family and does not establish optimal codes over all linear codes or quantum codes.
- AI role: The transformer predicts generalized toric codes’ minimum Hamming distances and guides genetic search toward promising generator matrices. Candidate champions are then checked with the Brouwer–Zimmermann algorithm and compared with known bounds; the paper reports six previously unknown champion codes at F₈ dimensions 18, 22 and 38.
Evidence sources: https://www.nature.com/articles/s44387-026-00077-3
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 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.
Knuth's note reports that Claude Opus 4.6 found an explicit decomposition of the arcs of the m^3-vertex digraph into three directed Hamiltonian cycles for odd m, an open problem Knuth posed. Knuth's proof of the construction was reportedly formally verified in Lean by Kim Morrison. Caveat: sources disagree on the even case, which an independent arXiv paper (2603.24708) and a GPT-5.4 Pro proof treat differently.
Technology: Claude Opus 4.6
Family Claude: profile · models page
Publication: Not annotated in this catalog · Test context: Not annotated in this catalog. Catalog categories, not an evidence grade.
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: other. 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 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.
A human-LLM pipeline (Co-FunSearch) finds adversarial instances for heuristics, raising the lower bound for Best-Fit bin packing in the random-order model from 1.3 to 1.5. The authors prove the 3/2 ratio analytically and report a search-found instance scoring 1.497 over 10,000 random shuffles against Packingsolver optima. Caveat: arXiv preprint; human analysis supplies the final proofs.
Technology: Co-FunSearch (FunSearch-based LLM program search)
Family FunSearch: 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: Human-refined FunSearch outputs were used to derive improved adversarial lower bounds for standard heuristics; the page gives no details of the bin-packing test setup.
- Still unknown: Not reported on the abstract page; the full paper is needed for the bin-packing evaluation details.
- AI role: LLM-based evolutionary search (FunSearch) supplied initial patterns, which human experts refined and extrapolated into constructions.
Evidence sources: https://arxiv.org/abs/2601.16849
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 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 LLM-driven generate-evaluate-refine loop built on OpenEvolve produced systems algorithms in case studies, with the abstract reporting up to 5.0x runtime improvements or 50% cost reductions over baselines. Solutions are checked with simulators or real systems, which is the validation basis. The results are on specific case studies rather than a general claim, and the work is a preprint.
Technology: OpenEvolve (LLM-driven evolutionary code search, ADRS framework)
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 systems algorithms were run against predefined workloads in real systems or simulators and compared with state-of-the-art human designs across case studies (cloud load balancing, MoE inference, LLM SQL queries, transaction scheduling). The abstract reports up to 5.0x runtime improvement or 50% cost reduction.
- Still unknown: The method assumes a reliable verifier, which the authors present as a precondition; the page has no dedicated limitations section and the work is a preprint.
- AI role: An LLM-driven generate-evaluate-refine loop (penEvolve, an open-source ADRS instance) proposed the algorithms.
Evidence sources: https://arxiv.org/abs/2510.06189
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 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.
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.