Marvin Eisenberger

Marvin Eisenberger is a machine learning researcher and one of the Google DeepMind developers of AlphaEvolve, a Gemini-powered coding agent that evolves algorithms for mathematics and computing. Much of his earlier work, written with Daniel Cremers and other coauthors, dealt with matching and interpolating 3D shapes: Smooth Shells and Deep Shells appeared in 2020, NeuroMorph in 2021, and the graph-based G-MSM came out of the Technical University of Munich in 2023. In 2022, he also presented a framework for implicit Sinkhorn differentiation, which computes analytical gradients for Sinkhorn layers in neural networks. At Google DeepMind, he helped develop the final version of AlphaEvolve, announced in May 2025, and used it with Emilien Dupont to find new matrix multiplication algorithms, including a 48-multiplication procedure for 4×4 complex-valued matrices. Google DeepMind's May 2026 review of the agent's first year again listed him among its developers.

Now at: Google DeepMind

Timeline

2020

Smooth shells for shape matching

With Zorah Lähner and Daniel Cremers, he presents Smooth Shells at CVPR in 2020, a 3D shape correspondence method that iteratively aligns coarse-to-fine approximations of the shapes. The method fuses classical shape registration with Functional Maps.
2020

Deep Shells at NeurIPS

In 2020, he presents Deep Shells at NeurIPS with Aysim Toker, Laura Leal-Taixé and Daniel Cremers. This unsupervised method builds a multiscale matching pipeline for 3D shapes into a deep neural network, derived from entropy-regularized optimal transport.
2021

NeuroMorph at CVPR

NeuroMorph, which he publishes at CVPR in 2021 with coauthors including Natalia Neverova, Daniel Cremers and Andrea Vedaldi, takes two 3D shapes as input. In a single feed-forward pass, the network produces a smooth interpolation and point-to-point correspondences between them, and it trains without manual correspondence annotations.
2022

Implicit Sinkhorn differentiation

At CVPR in 2022, he presents a unified framework for implicit Sinkhorn differentiation with Aysim Toker, Laura Leal-Taixé, Florian Bernard and Daniel Cremers. The algorithm computes analytical gradients of a Sinkhorn layer and improves gradient stability and accuracy when it replaces automatic differentiation.
June 2023

G-MSM in Vancouver

In June 2023, he presents G-MSM, an unsupervised method for non-rigid shape correspondence, at CVPR in Vancouver. Developed at the Technical University of Munich's Chair of Computer Vision and Artificial Intelligence with coauthors including Daniel Cremers, it builds an affinity graph over training shapes and propagates maps along shortest paths in that graph.
May 2025

Co-developing AlphaEvolve

In May 2025, Google DeepMind announces AlphaEvolve, an evolutionary coding agent that pairs Gemini models with automated evaluators to design algorithms for mathematics and computing, and names him among its developers. The agent's discoveries improve the efficiency of Google's data centers, chip design and AI training.
June 2025

The AlphaEvolve white paper

He is one of nine equal-contribution authors of the June 2025 AlphaEvolve white paper from Google DeepMind, and he develops the agent's final version with Alexander Novikov. With Emilien Dupont, he uses AlphaEvolve to discover new matrix multiplication algorithms; one multiplies two 4×4 complex-valued matrices with 48 scalar multiplications, the first improvement over Strassen's algorithm in this setting in 56 years.
May 2026

AlphaEvolve at one year

In May 2026, Google DeepMind's review of AlphaEvolve's first year again lists him among its developers. The review reports a 30% cut in variant-detection errors for the DNA-sequencing model DeepConsensus and the agent's regular use in designing the next generation of TPUs.