# Numaro

> Numaro is an AI autoresearch engine: multi-agent research software that generates hypotheses, writes tools, runs experiments, scales compute, and remembers what works across campaigns.

Numaro focuses on hard technical problems where progress can be checked against exact constraints, public baselines, or reproducible measurements.

Contact: contact@numaro.tech

## Core pages
- [Home](https://numaro.tech/): AI autoresearch engine overview.
- [Research](https://numaro.tech/research/): plain-language reports, result tables, visual explanations, methods, verification notes, and baseline sources.
- [Sitemap](https://numaro.tech/sitemap/): human-readable site index.
- [Full LLM context](https://numaro.tech/llms-full.txt): expanded machine-readable summary of every research report.

## Research reports
- [Winograd convolution with fewer transform additions](https://numaro.tech/research/winograd-transform-additions-2026/): The result is about arithmetic cost, not a GPU speed claim: fewer transform additions under the standard Winograd counting convention.
- [New best-known additive complexities for fast matrix-multiplication schemes](https://numaro.tech/research/matrix-multiplication-additive-complexity-2026/): The improvements come from exact per-side straight-line programs, plus careful checks against the field's own reducers.
- [Bit-exact GPU kernels in regimes vendor libraries leave open](https://numaro.tech/research/faster-ml-kernels-2026/): The report is deliberately narrow: it names the regimes that win and the regimes that lose.
- [Circle packing in the unit square: new sum-of-radii layouts](https://numaro.tech/research/circle-packing-unit-square-2026/): The output is raw circle coordinates, and the checker recomputes every overlap and wall clearance.
- [Maximizing total radius in triangle and pentagon containers](https://numaro.tech/research/friedman-packing-center-2026/): The claims handle Friedman's displayed-plus truncation rule explicitly.
- [Covering-design records improving the La Jolla Covering Repository](https://numaro.tech/research/covering-designs-la-jolla-2026/): 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.
- [Covering-array records improving Colbourn's best-known CAN tables](https://numaro.tech/research/covering-arrays-colbourn-2026/): 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.
- [New lower bounds for Zarankiewicz numbers](https://numaro.tech/research/zarankiewicz-2026/): 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.
- [First-known no-five-on-sphere values for [n]^3](https://numaro.tech/research/no-5-on-sphere-2026/): These are frontier extensions, not beats of AlphaEvolve's published cells.
- [First-known no-isosceles grid sets for open n values](https://numaro.tech/research/no-isosceles-grid-2026/): The search evolved the construction algorithm itself, but it did not beat AlphaEvolve's published n=64 or n=100 values.
- [DNA storage codes with larger codebooks](https://numaro.tech/research/dna-storage-codes-2026/): The result is about larger DNA codebooks under exact constraints, with the convention stated row by row.
- [Constant-weight codes: one propagation beat and exact DCW fills](https://numaro.tech/research/constant-weight-codes-2026/): The main table mostly held. The win is one clean propagation cell, plus an exact fill of a less-policed object.
- [qLDPC distances: exact certificates, zero corrections](https://numaro.tech/research/qldpc-distance-certificates-2026/): This is a checking result, not a record beat.
- [Learned mRNA coding-sequence designers are Pareto-suboptimal on their own objective](https://numaro.tech/research/mrna-codon-structure-design-2026/): The result is explicitly about a computable proxy objective, not biological expression.
- [Walshaw graph partitioning: no record beat, useful calibration](https://numaro.tech/research/walshaw-graph-partitioning-2026/): The useful output is the verifier, balance-rule reconstruction, and exact reproduction map.

## Suggested queries this file answers
- What is Numaro?
- What is an AI autoresearch engine?
- What research results has Numaro published?
- How does Numaro use multi-agent research, hypothesis generation, tool-making agents, experiments, and memory?