Winograd convolution with fewer transform additions
The result is about arithmetic cost, not a GPU speed claim: fewer transform additions under the standard Winograd counting convention.
289 adds for F(4x4,3x3)Research notes
What we found, how we checked it.
Each note starts in plain language, then gives the claim, table, method, checker, limits, and baseline sources.
New best-known additive complexities for fast matrix-multiplication schemes
The improvements come from exact per-side straight-line programs, plus careful checks against the field's own reducers.
31 additive-complexity recordsBit-exact GPU kernels in regimes vendor libraries leave open
The report is deliberately narrow: it names the regimes that win and the regimes that lose.
2-4.5x small batched matmulCircle packing in the unit square: new sum-of-radii layouts
The output is raw circle coordinates, and the checker recomputes every overlap and wall clearance.
9 unit-square records beatenMaximizing total radius in triangle and pentagon containers
The claims handle Friedman's displayed-plus truncation rule explicitly.
17 verified entriesCovering-design records improving the La Jolla Covering Repository
A trusted table had source improvements that had not propagated into the derived cells. Numaro found the gap, built the witnesses, and checked the counts.
6,440 blocks savedCovering-array records improving Colbourn's best-known CAN tables
This is the array version of the covering-design result: a public snapshot lagged construction catalogues, so Numaro built the actual arrays and checked every required interaction.
40 array table improvementsNew lower bounds for Zarankiewicz numbers
The object is simple: put as many 1s as possible in a 0/1 matrix without creating a 3 by 3 all-ones submatrix. Numaro found stronger matrices across the frontier.
31 lower bounds raisedFirst-known no-five-on-sphere values for [n]^3
These are frontier extensions, not beats of AlphaEvolve's published cells.
13 first-known valuesFirst-known no-isosceles grid sets for open n values
The search evolved the construction algorithm itself, but it did not beat AlphaEvolve's published n=64 or n=100 values.
54 first-known grid valuesDNA storage codes with larger codebooks
The result is about larger DNA codebooks under exact constraints, with the convention stated row by row.
16 larger DNA codebooksConstant-weight codes: one propagation beat and exact DCW fills
The main table mostly held. The win is one clean propagation cell, plus an exact fill of a less-policed object.
442 exact DCW valuesqLDPC distances: exact certificates, zero corrections
This is a checking result, not a record beat.
19 exact certificatesLearned mRNA coding-sequence designers are Pareto-suboptimal on their own objective
The result is explicitly about a computable proxy objective, not biological expression.
90/90 dominations reverifiedWalshaw graph partitioning: no record beat, useful calibration
The useful output is the verifier, balance-rule reconstruction, and exact reproduction map.
0 record beats claimed