Binxu Wang

Binxu Wang is a neuroscientist and AI researcher whose work connects generative image models with visual neuroscience. Raised in Beijing, she studied physics at Peking University, earned a neuroscience PhD with Carlos Ponce after following his lab from Washington University to Harvard Medical School, and joined Harvard's Kempner Institute as a Research Fellow in 2023. Her research mapped GAN image manifolds, closed-loop neuron-guided synthesis in visual cortex, and the score geometry, spectral bias and random-matrix consistency of diffusion models. In July 2026, she and her collaborators received an ICML Outstanding Paper Honorable Mention for their random-matrix analysis of diffusion consistency, and she completed her Kempner fellowship in September 2026. She is an incoming Group Leader at HHMI's Janelia Research Campus, with a December 2026 start.

Now at: Harvard Medical School and Howard Hughes Medical Institute

Timeline

By 2018

Physics at Peking University

Raised in Beijing, she studies physics at Peking University's Yuanpei College, taking courses in mathematics, philosophy and psychology before beginning neuroscience graduate study in 2018.
2018–2021

GAN image geometry

In 2018, she starts a neuroscience PhD at Washington University in St. Louis and works in Carlos R. Ponce's lab. She and Ponce publish The Geometry of Deep Generative Image Models and its Applications in 2021, using a Riemannian metric to map GAN image manifolds and expose interpretable directions for latent-space optimization.
2021–2023

Move to Harvard Medical School

After the Ponce lab relocates to Harvard Medical School in fall 2021, she moves to Boston and enrolls at Harvard Griffin GSAS. Earning a neuroscience PhD in fall 2023, she joins the Kempner Institute's inaugural Research Fellow cohort at Harvard. With John Vastola, she publishes The Hidden Linear Structure in Score-Based Models and its Application in November 2023, showing that high-noise diffusion scores are well approximated by a linear Gaussian score to accelerate image sampling by 15–30%.
2024

A Gaussian score for diffusion

In December 2024, she and John Vastola publish The Unreasonable Effectiveness of Gaussian Score Approximation for Diffusion Models and its Applications. They compare learned scores with Gaussian and Gaussian-mixture scores and find that a linear Gaussian approximation dominates learned scores at moderate to high noise scales. This result underpins analytical teleportation, a hybrid sampling method that accelerates samplers such as DPM-Solver-v3 and UniPC.
2025

Spectral bias in diffusion

At NeurIPS 2025, she and Cengiz Pehlevan present An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models, deriving an inverse power-law relation between data-mode variance and diffusion-model convergence time. Experiments on Gaussian and natural-image data test how data covariance governs the order and rate at which different modes emerge.
March 2026

Neuron-guided synthesis in Nature Neuroscience

Collaborating with Carlos R. Ponce in March 2026, she publishes a study in Nature Neuroscience on how neuronal tuning aligns with object and texture manifolds across the visual hierarchy. Closed-loop optimization in their experiments allows neurons in V1, V4 and the posterior inferotemporal cortex to guide image synthesis.
July 2026

An ICML honorable mention

With Jacob Zavatone-Veth and Cengiz Pehlevan in February 2026, she posts A Random Matrix Theory Perspective on the Consistency of Diffusion Models, giving a random-matrix account of why diffusion models trained on separate data splits produce similar outputs from the same noise seed. In July 2026, ICML awards the paper an Outstanding Paper Honorable Mention.
September–December 2026

Incoming Group Leader at Janelia

She completes her Kempner Research Fellowship in September 2026 and is an incoming Group Leader at HHMI's Janelia Research Campus in Ashburn, Virginia, with her appointment due to start in December 2026.