Rogério Guimarães

Rogério Guimarães

PhD Student in
Computation and Neural
Systems at Caltech.

B.S. in Computer Science
and in Linguistics and
Philosophy at MIT

Publications

Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning Permalink

Rogério Guimarães & Pietro Perona (ECCV 2026)

We propose Progressive Seed Pruning (PSP), an inference-time scaling method for diffusion and flow-matching models that scores intermediate denoised estimates and progressively narrows a large pool of candidate seeds, spending compute on promising trajectories instead of resampling a fixed-size pool. Across diffusion and flow-matching backbones, PSP consistently improves reward-guided selection and achieves higher GenEval scores and better human evaluation on prompt-alignment than best-of-N, importance-sampling, and tree-search baselines at matched compute.

Diffusion-Based Action Recognition Generalizes to Untrained Domains Permalink

Rogério Guimarães, Frank Xiao, Pietro Perona & Markus Marks (WACV 2026)

We propose using features generated by a Vision Diffusion Model (VDM), aggregated via a transformer, to achieve human-like action recognition across domain shifts. We find that generalization is enhanced by the use of a model conditioned on earlier timesteps of the diffusion process to highlight semantic information over pixel level details in the extracted features. Our model sets a new state-of-the-art across three generalization benchmarks, bringing machine action recognition closer to human-like robustness.

Humans and Mice Navigate Mazes Alike. Can AI Beat Them? Permalink

Rogério Guimarães, Alina Zhang, Frank Xiao, Evan Z. Wang, Jieyu Zheng & Pietro Perona (CCN 2025)

We tested human participants on a virtual binary maze game adapted from a prior mouse study and found not only similar performance, but also striking parallels in learning dynamics between humans and mice. We then used this embodied navigation task to compare AI agents with both species, and showed that two canonical agents, a Deep Q-Learning (DQN) model and a Large Language Model (LLM), were outperformed by the biological learners.

Mice in the Manhattan Maze: Rapid Learning, Flexible Routing and Generalization, With and Without Cortex Permalink

Jieyu Zheng, Rogério Guimarães, Jennifer Hu, Pietro Perona & Markus Meister (CCN 2024)

We designed the “Manhattan Maze,” a reconfigurable navigation task that allows systematic map design over a vast space of possible layouts, to study cognitive flexibility in mice. Naive wildtype mice rapidly learned complex maps within two days and accelerated at learning new ones. Acortical mice, lacking the hippocampus and most of the neocortex, learned more slowly at first but still approached optimal performance, suggesting the cortex is not strictly required for navigating the Manhattan Maze.

Text-image Alignment for Diffusion-based Perception Permalink

Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães & Pietro Perona (CVPR 2024)

We use automatically generated captions to improve the text-image alignment of a diffusion backbone in downstream visual tasks such as semantic segmentation, depth estimation and object detection. Our method also achieves improves the SOTA in both single-domain and cross-domain tasks.