AI autoresearch engine
Hard problems need new tools.
Numaro is an AI autoresearch engine powered by agent swarms and an ontological reasoning engine that connects concepts, constraints, evidence, and methods to create the hypotheses, solvers, experiments, and memory a hard problem needs.
hypothesis generationtool-making agentsshared memoryautomatic compute
GHR sum-difference constantA certified base-89 construction sets a new record substantially above DeepMind's AlphaEvolve frontier.1.1873261 certified recordSpherical-code record campaignMore than 11,000 verified table records across dimensions 3–32.11,000+ verified recordsWinograd inference transformsFewer operations in the convolution math used by modern inference stacks.328 -> 289 addsDeepMind AlphaEvolve benchmarkNumaro beats the published n=26 benchmark and eight additional best-known layouts.n=26 benchmark beaten
System
Fully-automated research loop with memory.
Numaro combines multi-agent research, hypothesis generation, tool-making, automatic experiments, and shared memory. Each campaign leaves behind code, traces, and failed routes the next one can use.
01Problem graph
Map papers, constraints, benchmarks, and previous runs.
02Hypotheses
Generate directions before committing compute.
03Tool factory
Write solvers, experiments, checks, and reusable code.
04Memory
Carry traces and working methods into the next campaign.
Additive combinatoricsAnalysis of Boolean functionsDiscrete geometryCoding theoryFast convolutionAlgebraic complexityGPU kernelsGeometry optimizationGeometry optimizationExtremal combinatoricsCombinatorial testingExtremal combinatoricsDiscrete geometryDiscrete geometryCoding theoryQuantum error correctionSequence design