Doctoral Thesis Proposal - Zhengyang Geng
September 4, 2026 1:00PM—2:30PM
Location:
Newell-Simon 3002
Speaker:
ZHENGYANG GENG,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://gsunshine.github.io/
Reasoning models can scale test-time computation with problem difficulty: harder problems induce longer reasoning traces. Generative modeling in continuous domains lacks an analogous mechanism; the computational budget per sample is fixed by the framework rather than adapted to the instance. This thesis develops generative models that allocate computation according to instance complexity and user intent.
My completed work establishes two endpoints of this computational spectrum. MeanFlow derives one-step generation from an exact identity relating average and instantaneous velocity. Trained from scratch, MeanFlow models are standalone, stable, and scalable, and substantially narrow the quality gap between one-step and many-step generation on ImageNet. Equilibrium reasoners are trained with at most sixteen recurrent iterations, yet at inference can adaptively run from a single iteration to more than one thousand, without task-specific priors or external verifiers.
The proposed work unifies these endpoints. MeanFlow training itself solves a fixed-point problem, while an equilibrium backbone allows a one-step generative model to perform latent, adaptive computation in response to instance complexity and user intent. At the core of Equilibrium Mean Flows, the MeanFlow identity gives rise to three coupled fixed-point systems that share a single Jacobian. The resulting model is designed to halt early on easy instances and continue iterating on difficult ones. This framework aims to enable one-step generative models to construct variable-length internal refinement processes, making the allocation of computation a learned capability of the model itself.
Thesis Committee:
J. Zico Kolter (Chair)
Tai-Sing Lee
Deepak Pathak
Kaiming He (Massachusetts Institute of Technology)
Contact
Matt Stewart