Reinforcement Learning for Algorithm Discovery
Using game-playing reinforcement learning agents to find new, provably correct algorithms, such as faster matrix multiplication or sorting.
Tier Grades
Tier A: Frontier AI labs
feasible
Google DeepMind built AlphaTensor and AlphaDev on AlphaZero.
Tier B: Indonesian AI labs
not feasible
Compute is not stated and the actor is frontier only.
Tier C: Institution units such as ITB
stretch
Mathematics groups could study the published algorithms, but compute is not stated.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
not feasible
No delivery use is shown.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeAlphaTensor: compute not stated in the source.
- DataAlphaTensor: matrix multiplication posed as a single-player game, building on AlphaZero.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): Not in the sources read (the blog gives no author count).
- Time horizonMulti-year.
- Shelf-life signalsAlphaTensor addressed a 50-year-old open question since Strassen in 1969.
- Pipeline roleTier A discovers algorithms; Tier C studies and verifies them; Tier B and Tier D have no observed role.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Links
R&D TopicReinforcement Learning for Algorithm Discovery
Models
AlphaTensorGoogle DeepMind, reinforcement learningAlphaDevGoogle DeepMind, reinforcement learningAlphaZeroGoogle DeepMind, reinforcement learning
AlphaTensorGoogle DeepMind, reinforcement learningAlphaDevGoogle DeepMind, reinforcement learningAlphaZeroGoogle DeepMind, reinforcement learningUse Casesnone recorded
Who Runs Itnone recorded
Papers
- Discovering Reinforcement Learning AlgorithmsJunhyuk Oh et al., arXiv preprint, 2020 founding
- Eureka: Human-Level Reward Design via Coding Large Language ModelsYecheng Jason Ma et al., ICLR 2024 (per arXiv comment), 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- deepmind.google/discover/blog/discovering-novel-algorithms-wi... https://deepmind.google/discover/blog/discovering-novel-algorithms-with-alphatensor/
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Reinforcement Learning for Algorithm Discovery", https://aurion.semai.id/topics/reinforcement-learning-for-algorithm-discovery/, records of October 2026.