AlphaEvolve

AlphaEvolve is an autonomous coding agent developed by Google DeepMind that uses large language models to improve computer programs through an evolutionary search loop. Announced in May 2025, it pairs an ensemble of Gemini models with automated evaluators that run and score every proposed program, allowing the system to discover new algorithms and optimise critical pieces of production infrastructure without human intervention.^[1]

History

AlphaEvolve was announced on 14 May 2025 as a generalisation of DeepMind's earlier FunSearch system, which in 2023 had shown for the first time that language models could generate code that discovers provably correct knowledge on an open mathematical problem.^[1] Where FunSearch evolved a single Python function of roughly 10–20 lines, AlphaEvolve evolves whole code files of hundreds of lines, in any programming language, using frontier models and rich natural-language feedback.^[2] A comprehensive technical white paper describing the system was published as an arXiv preprint in June 2025.^[2]

Key Concepts and Technology

At its core, AlphaEvolve runs an evolutionary loop. A prompt sampler selects a parent program from a database of previously tried solutions, a Gemini model proposes changes as structured diffs, the diff is applied to produce a candidate program, and evaluators — deterministic functions supplied by the user — execute the candidate and score it. Only high-scoring programs are retained and resampled in later generations.^[2]

Three design choices distinguish it from prior work. First, the evolution blocks API: users annotate regions of an existing codebase with # EVOLVE-BLOCK-START and # EVOLVE-BLOCK-END comments, so the agent can be pointed at a live codebase with minimal edits.^[2] Second, multi-objective optimisation: a single run can balance several evaluation metrics (speed, accuracy, memory) rather than one scalar score as FunSearch did.^[2] Third, grounding through execution: because every suggestion is run and measured, incorrect hallucinations from the underlying model are filtered out automatically.^[2]

AlphaEvolve uses an ensemble of Gemini 2.0 Flash for high-throughput candidate generation and Gemini 2.0 Pro for higher-quality suggestions on difficult steps. The program database combines the MAP-Elites algorithm with island-based population models to balance exploration and exploitation.^[2]

Applications and Impact

AlphaEvolve's results fall into two categories: engineering optimisation at Google scale, and novel scientific results.^[1][2]

• Data centre scheduling: a simple evolved heuristic for Google's Borg scheduler has run in production for over a year, recovering on average 0.7% of Google's worldwide compute resources.^[1]
• Hardware design: the agent proposed a Verilog rewrite removing unnecessary bits from a highly optimised arithmetic circuit in Google's TPU accelerators, verified for functional correctness.^[1]
• AI training speed: by re-tiling a matrix-multiplication kernel used to train Gemini, AlphaEvolve delivered an average 23% kernel speedup and a 1% reduction in Gemini's overall training time, compressing months of expert engineering into days of automated experiments. It also achieved up to a 32.5% speedup on a FlashAttention kernel.^[1][2]
• Mathematics: the system found an algorithm multiplying two 4×4 complex-valued matrices with 48 scalar multiplications — the first improvement over Strassen's 1969 algorithm in that setting after 56 years — and was later applied to a suite of 67 problems across analysis, combinatorics, geometry and number theory, rediscovering most best-known solutions and improving several.^[2][3]

>See Also

• Google DeepMind
• Agentic AI
• Code Generation
• AI in Mathematics
• Gemini