Research publications by Wayner Barrios, from multimodal reasoning and video understanding to visual grounding, medical AI, and efficient inference. Find papers, summaries, and links to the open source implementations.
These publications examine how models connect visual observations with language and reasoning. My work on video asks how a language query can identify a relevant moment in a long sequence. The visual grounding work studies which information an instruction should draw from an image. More recent evaluation work asks whether an explanation contains the steps needed to justify its final answer.
For a closer look at the methods and their scope, read the MoDA visual grounding guide and the CRYSTAL reasoning guide. The vLLM-MLX inference guide follows the systems side of this research, explaining batching, caching, and how to interpret the reported Apple Silicon benchmarks. Each guide links the original paper and implementation; the publication list below also includes earlier work and collaborations in medical AI.
Publications
Latest · Multimodal reasoning
A 6,372-instance benchmark that checks every intermediate step. Across 20 models, none maintained more than 60% step accuracy in the correct order; Causal Process Reward improves step-level consistency by 32% without manual annotations.
Introduces native MLX inference with continuous batching and multimodal prefix caching, reaching up to 525 tokens per second and 28x faster repeated image queries on M4 Max.
Uses instruction-guided channel modulation to improve fine-grained visual grounding across three MLLM families and 12 benchmarks with less than 1% additional FLOPs.
Combines AI biomarkers from chest radiographs, clinical notes, and ECG signals to study pneumonia mortality using routinely collected hospital data.
Generates talking faces directly from first-person text in a one-stage pipeline, without requiring audio, landmarks, or pose during inference.
Introduces multi-layer learnable attention masks that prioritize informative tokens across transformer layers while reducing redundant computation.
Improves language grounding in long videos by identifying describable windows and pruning irrelevant segments before matching a query.
Introduces Bladder4Net to classify whole-slide histopathology images into low-risk and high-risk bladder cancer groups for prognosis support.
Builds a weakly supervised CT pipeline for COVID-19 classification, lesion segmentation, and quantification from scan-level labels.
Analyzes the training and testing gaps between proposal, classification, and localization modules in event detection pipelines, then introduces practical fixes.
Uses object and scene semantics to generate class-specific action proposals and prune unrelated activities in real time on ActivityNet.
Building Wiqonn and researching multimodal perception, video understanding, and efficient inference.
Get in touch