Publications

Low Vision Behavior Simulation

2026

Assessment of Newly Designed Fonts for Visual Accessibility

PLoS ONE 2026

Gordon E Legge

Yingzi Xiong

Qingying Gao

Rachel Gage

Taylor Knickel

Charles Bigelow

KEYWORDS: Font Design, Visual Accessibility, Low Vision, OCR

This study compares behavioral and automated methods for evaluating font accessibility for both normal and low-vision readers. Twenty-two normally sighted young adults were tested using a computerized MNREAD test with five fonts under normal conditions and simulated moderate low vision (20/90 acuity). Nine of eleven OCR models showed changes in reading acuity similar to humans, suggesting automated methods offer a practical alternative to intensive human testing for revealing subtle accessibility differences across typefaces.

2025

Walk, See, and Trace: A New Method for Object Labeling in Real-Life Settings

ARVO 2025

Qingying Gao

Rama Chellappa

Peng Cheng

Kristen Shifflet

Gordon E Legge

Yingzi Xiong

KEYWORDS: Low Vision, Computer Vision, Assistive Tool

“What is in front of you?” Numerous studies attempt to understand the impact of low vision on object visibility using computer images, while real-life object visibility is underexplored despite greater ecological validity and critical perceptual factors such as binocular parallax and self-motion. We developed a “Walk, See, and Trace” task, in which observers wear light-weight glasses fitted with scene camera as they report and hand-trace the boundaries of objects in real environment. Here we report the initial evaluation in individuals with normal vision.

Measuring Critical Viewing Distance of Computer Vision Models in Hazard Recognition

ARVO 2025

Seungeon Han

Clara Kim

Qingying Gao

Kristen Shifflet

Rama Chellappa

Peng Cheng

Gordon E Legge

Yingzi Xiong

KEYWORDS: Low Vision, Computer Vision, Model Assessment

As computer vision (CV) applications gain popularity in assisting blind and low-vision individuals, model evaluation should be tailored to these users’ practical needs. In a hazard warning context, it is critical that an assistive application can detect a hazard before the user reaches a contact range. We propose using critical viewing distance (CVD), a metric motivated by human vision evaluation, to assess CV models for hazard recognition. Here, we demonstrate how CVD changes across models, lighting conditions, and hazard types to prove the necessity of such a metric.

VI-OCR: “Visually Impaired” Optical Character Recognition Pipeline for Text Accessibility Assessment

Scientific Reports 2025

Qingying Gao

Roberto Manduchi

Pradeep Y Ramulu

Gordon E Legge

Yingzi Xiong

KEYWORDS: Low Vision, OCR, Text Accessibility, Contrast Sensitivity

We present VI-OCR, a framework combining contrast sensitivity function-based image filtering with OCR models to assess text legibility for individuals with low vision. We benchmarked multiple OCR systems and vision-language models across three reading tasks: letter acuity using ETDRS charts, word acuity using MNREAD charts, and scene text recognition using complex real-life images. Models like Qwen2.5-VL and GPT demonstrated human-like performance in replicating how text recognizability changes with specified levels of visual impairment.

2024

Creating A “Visually Impaired” Character Recognition Model for Text Accessibility Assessment

ARVO 2024

Qingying Gao

Roberto Manduchi

Pradeep Y Ramulu

Gordon E Legge

Yingzi Xiong

KEYWORDS: Low Vision, Computer Vision

Low vision individuals use their residual vision in their daily life to read text such as price tags, street signs, medicine labels. However, there is no objective tool for evaluating text accessibility for low vision. … We aim to design a new framework combining OCR and human contrast sensitivity functions (CSF) to simulate the text recognition capability of low vision.

Generative Model

2026

A Very Big Video Reasoning Suite

2026

Maijunxian Wang

Ruisi Wang

Juyi Lin

Ran Ji

Thaddäus Wiedemer

Qingying Gao

Dezhi Luo

Yaoyao Qian

Lianyu Huang

Zelong Hong

Jiahui Ge

Qianli Ma

Hang He

Yifan Zhou

Lingzi Guo

Lantao Mei

Jiachen Li

Hanwen Xing

et al.

KEYWORDS: Video Reasoning, Benchmark, VLM

This paper introduces the VBVR Dataset, comprising over one million video clips across 200 reasoning tasks. We present a verifiable evaluation framework that moves beyond model-based judging by incorporating alternative assessment methods, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Our large-scale evaluation reveals early evidence of generalization to unfamiliar reasoning tasks.

Egocentric Bias in Vision-Language Models

2026

Maijunxian Wang

Yijiang Li

Bingyang Wang

Tianwei Zhao

Ran Ji

Qingying Gao

Emmy Liu

Hokin Deng

Dezhi Luo

KEYWORDS: VLM, Perspective-taking, Egocentric Bias

We introduce FlipSet, a diagnostic benchmark testing how well vision-language models can infer how the world appears from another’s viewpoint. Testing 103 models revealed systematic egocentric bias where most performed below chance level. We discovered a compositional deficit: models handle theory-of-mind and mental rotation separately but fail catastrophically when integration is required, suggesting fundamental limitations in spatial reasoning.

Vision-Language Models Mistake Head Orientation for Gaze Direction

ACL 2026 2026

Zory Zhang

Pinyuan Feng

Bingyang Wang

Tianwei Zhao

Suyang Yu

Qingying Gao

Hokin Deng

Ziqiao Ma

Yijiang Li

Dezhi Luo

KEYWORDS: VLM, Gaze Direction, Social Cognition

We investigated how well vision-language models can identify where people are looking. We created 1,360 real-world photos showing individuals gazing at tabletop objects while controlling head positioning. Results revealed a substantial performance gap between VLMs and humans, with models relying on head orientation rather than eye appearance to infer gaze direction. Fine-tuning experiments on transformer-based vision models suggest this bias stems from training data rather than architectural limitations.

2025

ICML 2025 2025

Yijiang Li

Qingying Gao

Tianwei Zhao

Bingyang Wang

Haoran Sun

Haiyun Lyu

Robert D Hawkins

Nuno Vasconcelos

Tal Golan

Dezhi Luo

Hokin Deng

KEYWORDS: VLM, Core Knowledge, Cognitive Development, Benchmark

We introduce CoreCognition, a benchmark testing 12 core knowledge concepts rooted in developmental psychology, and evaluate 230 multimodal language models. We find that these systems struggle with basic-level tasks relative to high-level ones and exhibit reduced, or even absent, scalability as they grow. We propose a method called Concept Hacking demonstrating that improved performance stems from shortcut learning rather than genuine conceptual understanding.

SCSL @ ICLR 2025 2025

Yijiang Li

Bingyang Wang

Tianwei Zhao

Qingying Gao

Hokin Deng

Dezhi Luo

KEYWORDS: VLM, Core Knowledge, Evaluation, Concept Hacking

We introduce Concept Hacking, a paradigm that manipulates concept-relevant information to flip the ground-truth while preserving concept-irrelevant confounds. We assessed 209 models across 45 experiment pairs spanning nine low-level cognitive abilities. Our findings reveal that models either rely on shortcuts or demonstrate illusory understanding, with none approaching human-level performance. Scaling neither imparts core knowledge nor reduces shortcut reliance.

Probing Perceptual Constancy in Large Vision-Language Models

2025

Haoran Sun

Bingyang Wang

Suyang Yu

Yijiang Li

Qingying Gao

Haiyun Lyu

Lianyu Huang

Zelong Hong

Jiahui Ge

Qianli Ma

Hang He

Yifan Zhou

Lingzi Guo

Lantao Mei

Maijunxian Wang

Dezhi Luo

Hokin Deng

KEYWORDS: VLM, Perceptual Constancy, Cognitive Development

We explore how vision-language models maintain consistent object perception when visual conditions change, such as distance, viewing angle, or illumination. We evaluated 155 VLMs across 236 experiments in three areas: color, size, and shape constancy. Testing involved both single-image and video versions of established cognitive psychology tasks, plus experiments under naturalistic conditions. Results reveal considerable variation in model performance, with shape constancy showing notably distinct patterns.

Probing Mechanical Reasoning in Large Vision Language Models

BiAlign @ ICLR 2025 2025

Haoran Sun

Qingying Gao

Haiyun Lyu

Dezhi Luo

Yijiang Li

Hokin Deng

KEYWORDS: VLM, Mechanical Reasoning, Cognitive Development

We examined mechanical reasoning capabilities across 26 vision language models using 155 cognitive experiments covering system stability, gears and pulley systems, leverage principle, inertia and motion, and fluid mechanics. VLMs consistently underperform compared to humans across all domains, with particular struggles in gear and fluid mechanics reasoning. Performance gains do not correlate with increasing model parameters, suggesting fundamental limitations in current attention-based architectures.

Vision Language Models Know Law of Conservation without Understanding More-or-Less

BiAlign @ ICLR 2025 2025

Dezhi Luo

Haiyun Lyu

Qingying Gao

Haoran Sun

Yijiang Li

Hokin Deng

KEYWORDS: VLM, Conservation, Cognitive Development

We created ConserveBench, comprising 365 cognitive experiments across four dimensions of physical quantities: volume, solid quantity, length, and number. Our findings reveal that VLMs perform well on transformational tasks requiring understanding of reversible operations, yet struggle with non-transformational tasks assessing quantity comprehension. This suggests that understanding the reversibility of operations and understanding the concept of quantity operate independently in these models.

Rethinking the Simulation vs. Rendering Dichotomy: No Free Lunch in Spatial World Modelling

NeurIPS 2025 Workshop 2025

Dezhi Luo

Qingying Gao

Hokin Deng

KEYWORDS: Spatial World Model, Embodied AI, Cognitive Science

We examine spatial world models crucial for computational systems operating in physical environments. Drawing on aphantasia research, we contend that detailed perceptual information is essential for model-based spatial reasoning. We propose that spatial simulation and perceptual experience rely on shared representational structures captured by higher-order indices of perceptual relations. Recent embodied AI developments support this view, showing that rich perceptual details enhance performance in physics-based tasks.

2024

Vision Language Models See What You Want but not What You See

2024

Qingying Gao

Yijiang Li

Haiyun Lyu

Haoran Sun

Dezhi Luo

Hokin Deng

KEYWORDS: VLM, Cognitive development

Knowing others’ intentions and taking others’ perspectives are two core components of human intelligence that are typically considered to be instantiations of theory-of-mind. Infiltrating machines with these abilities is an important step towards building human-level artificial intelligence. … We find VLMs achieving high performance on intentionality understanding but lower performance on perspective-taking. This challenges the common belief in cognitive science literature that perspective-taking at the corresponding modality is necessary for intentionality understanding.

Augmented Reality

2023

Mixed Reality Guided Root Canal Therapy

AE-CAI Workshop 2023

Fangjie Li*

Qingying Gao*

Nengyu Wang

Nicholas Greene

Tianyu Song

Omid Dianat

Ehsan Azimi

* Co-first author

KEYWORDS: Mixed Reality, Root Canal Therapy, Navigation, Augmented Reality

Root Canal Therapy (RCT) is a widely performed procedure in dentistry, with over 25 million individuals undergoing it annually. This procedure is carried out to address inflammation or infection within the root canal system of affected teeth. However, accurately aligning CT scan information with the patient’s tooth has posed challenges, leading to errors in tool positioning and potential negative outcomes. To overcome these challenges, we have developed a mixed reality application using an Optical See-Through Head-Mounted Display (OST-HMD)…