CMU PhD Zhengyang Geng Joins OpenAI After One-Step Generation Breakthrough

Carnegie Mellon University PhD student Zhengyang Geng recently announced on X (formerly Twitter) that he has joined OpenAI . Geng began his doctorate in computer science at CMU in 2022 under the supervision of J. Zico Kolter and has collaborated closely with Kaiming He .

Research Focus: Generative Models, Dynamical Systems, Adaptive Computation

Geng's work centers on three main directions: generative models, dynamical systems, and adaptive computation. His recent publications form a coherent thread exploring how to match computational effort to task difficulty — using minimal computation for simple tasks and scaling up for hard ones.

MeanFlow: One-Step High-Quality Image Generation

In 2025, Geng, He, and colleagues introduced MeanFlow , a method that enables high-quality image generation in a single model evaluation. Unlike distillation-based approaches that rely on a pre-trained multi-step teacher, MeanFlow derives from a mathematical principle linking average velocity and instantaneous velocity in continuous-time flow models, allowing direct training of a single-step generator.

Benchmark: ImageNet 256×256 generation.

Result: FID 3.43 with only one forward pass.

Recognition: Accepted as an Oral presentation at NeurIPS 2025 .

Improved MeanFlow: Better Stability and Quality

The team followed up with Improved MeanFlow , which further enhances training stability and generation quality. On the same ImageNet 256×256 benchmark, it achieves a single-step FID of 1.72 , a substantial improvement over the original MeanFlow.

Recognition: Awarded CVPR 2026 Highlight .

Equilibrium Reasoners: Scalable Reasoning in Latent Space

In a complementary direction, Geng contributed to Equilibrium Reasoners (EqR) , accepted at ICML 2026 . EqR explores how models can perform iterative reasoning internally, in the hidden (latent) space, rather than generating long textual chains of thought.

Mechanism: The model repeatedly updates its internal state until it converges to an attractor representing the answer.

Adaptive computation: Simple questions may need only 1–5 iterations ; harder problems automatically receive more internal steps.

Sudoku-Extreme benchmark: Accuracy rises from 2.6% (feedforward baseline) to over 99% when test-time computation is scaled.

Unifying Principle: Dynamic Compute Allocation

Although MeanFlow (reducing steps) and EqR (increasing steps) appear opposite, they share a common philosophy: computational resources should be allocated dynamically according to task difficulty . Simple tasks get minimal computation; difficult tasks get more. This principle underlies both the push for one-step generation and the design of scalable latent-space reasoners.

With his move to OpenAI, Geng is expected to continue advancing generative modeling and adaptive reasoning at the frontier of AI research.