Primary evidence
- Discovering faster matrix multiplication algorithms with reinforcement learningS1 · Primary research · Paper · 2022-10-05
- AlphaTensor accompanying repositoryS1 · Primary research · Repository
AlphaTensor discovers improved matrix multiplication algorithms
DeepMind represented matrix multiplication as a tensor-decomposition game and trained an AlphaZero-style reinforcement-learning agent to search the enormous space of factorizations. Released artifacts include discovered factorizations, benchmarking code and checks of algorithmic properties.
What happened?
AlphaTensor discovered provably correct matrix multiplication algorithms that improved the best known multiplication counts for multiple matrix sizes, including a 4-by-4 construction over Z2 using 47 scalar multiplications instead of 49 from two-level Strassen.
Why it matters
The work moved AI4Math beyond solving prepared exercises by demonstrating machine search that produced new exact algorithms with mathematical and computational significance.
Significance rationale: AlphaTensor provided a prominent example of reinforcement learning discovering exact algorithms that improved long-standing matrix-multiplication records, including a finite-field 4-by-4 result reported as the first improvement over Strassen's two-level construction in that setting.
Historical context
Earlier developments
—
Evidence & status at a glance
- Significance
- H1 · Historical milestone
- Verification
- Paper released
- Evidence level
- E3 · Inspectable artifact
- AI role
- AI primary, human verified
- Formal assurance
- Artifact available
Contribution & verification
- Mathematical novelty
- New algorithm
- Interface
- Symbolic, Numeric, Hybrid
- AI contribution
- AlphaTensor searched for tensor decompositions and discovered candidate algorithms; the resulting decompositions admit exact mathematical checking and were analyzed by the research team.
- Human contribution
- Researchers formulated TensorGame, designed the reinforcement-learning system and reward structures, chose target problems, verified outputs and interpreted their algorithmic consequences.
- Verification
- Paper released · E3 · Inspectable artifact · Artifact available
- Last reviewed
- 2026-09-11
Technical details
- Systems
- alphatensor
- Organizations
- deepmind
- Problems
- matrix-multiplication
- Methods
- reinforcement-learning, monte-carlo-tree-search, tensor-decomposition, algorithm-search
Sources & artifacts
- Discovering faster matrix multiplication algorithms with reinforcement learningS1 · Primary research · Paper · accessed 2026-09-10
- AlphaTensor accompanying repositoryS1 · Primary research · Repository · accessed 2026-09-10
- AlphaTensor algorithms and verification materialsArtifact · Code
Verification history
- 2022-10-05 · Paper released · E2 · Research record
Nature published the discovery results and their mathematical formulation.
- 2022-10-05 · Paper released · E3 · Inspectable artifact
DeepMind released discovered algorithms and supporting verification and benchmarking materials.
Corrections
No corrections recorded.