Maks Ovsjanikov

Far from merely clustering tokens, full Transformers act as pattern-forming dynamical systems that natively amplify traveling and rotating waves—an inductive bias that can be engineered at initialization to accelerate training.

Label-free consistency training can elevate spatial reasoning in LRMs to levels comparable with traditional supervised methods, revolutionizing how we approach model training.

Text-to-3D models can get stuck in "sink traps" where they ignore your prompts, but unlocking their unconditional generation power lets you edit shapes they otherwise couldn't.

Extract watertight meshes from 3D Gaussian Splatting with unprecedented fidelity, even capturing notoriously difficult thin structures like bicycle spokes.

Even state-of-the-art VLMs exhibit systematic failures in reasoning about the physical feasibility of actions in 3D environments, despite high semantic confidence.

Generative drifting's empirical success is no longer a mystery: it's secretly score matching, but with frequency-dependent convergence bottlenecks that explain the preference for Laplacian kernels.

Ditching ReLU can provably prevent oversmoothing in deep GNNs by inducing bifurcations that create stable, heterogeneous node representations.