Arman Akbari

I'm a first year PhD student in Computer Engineering at Northeastern University, co-advised by Prof. Yanzhi Wang and Prof. Xuan Zhang. My research interests focus on Efficient AI especifically for LLMs, Multi-modal LLMs, and Diffusion Models. I got my B.Sc. in Computer Science from University of Tehran.

Email  /  Google Scholar  /  LinkedIn  /  Github

Research

My research centers on Efficient AI. I develop methods to accelerate pre-training, fine-tuning, and inference of generative models. My work includes model compression, optimization for large language and vision models, and designing fast, robust architectures that reduce computational cost without sacrificing performance.

While Multi-modal Large Language Models (MLLMs) excel at natural image tasks, their ability to extract equations from technical diagrams remains untested. We present CircuitSense, a benchmark of 8,006+ problems evaluating circuit understanding across three tasks—Perception, Analysis, and Design—with emphasis on deriving symbolic equations from visual inputs. dditionally, we propose a hierarchical synthetic pipeline that generates schematics and block diagrams with guaranteed ground-truth equations.

We propose VOTE, an efficient and general framework for the optimization and acceleration of VLA models. VOTE is a novel tokenizer-free fine-tuning approach for parallel accurate action prediction, which reduces computational overhead and accelerates inference speed. Our method achieves state-of-the-art performance with 35 times faster inference and 145 Hz throughput.

In this work, we systematically investigate how variations in the pedestrian pose—including leg status, elbow status, and body orientation—as well as individual joint occlusions, affect detection performance. We evaluate five pedestrian-specific detectors (F2DNet, MGAN, ALFNet, CSP, and Cascade R-CNN) alongside three general-purpose models (YOLOv12 variants) on the EuroCity Persons Dense Pose (ECP-DP) dataset.

In this paper, a geometry-based algorithm is presented which can find grasp poses based on the geometry of the unknown object and propose the ones which may lead to successful grasping. For the grasp contacts computation part, the presented algorithm produces a finite number of key points based on the 2D shape of the object from a specific point of view. Afterward, it will narrow down the candidate points and output a finite number of successful grasp poses based on three grasp quality metrics

Professional Services

• Conference Reviewer
• NuerIPS'25, ICCV'25

Template credit: source code