AI PhD Researcher · Software Engineering
Linqiang (Jackson) Guo
AI-driven mobile automation · LLM agents · Multimodal VLM post-training
Software Engineering PhD student at Concordia University and Graduate Student Researcher at Mila, building LLM-based agents that reason about state, recover from failure, and understand user interfaces. Former software engineer at Walmart Global Tech.
Latest News
-
Oct 2026
Glad to be one of the contributors to AgentPersonaBench: Benchmarking Persona-Driven User Simulation, a MatrAIx-led benchmark with 46 authors across 31 institutions, now on arXiv. It tests whether persona conditioning actually steers agent behavior in survey, chat, web and desktop-app environments.
-
Aug 2026
Our paper Learning to Adapt: Few-Shot Generalization of Mobile Agents to Unseen Applications received the Best Paper Award at the ECCV 2026 Workshop on Multimodal Digital Agents (MDA), with an invited Spotlight Talk.
-
Aug 2026
Our paper Q-Reflector: Learning GUI Action Reflection with Generated Conditions was accepted to the ECCV 2026 Workshop on Multimodal Digital Agents (MDA) (Poster + Lightning Talk).
-
Aug 2026
Our paper CoAdapt-GUI: Joint Workflow Context and Policy Adaptation for Unseen GUI Applications was accepted to the ECCV 2026 Workshop on Multimodal Digital Agents (MDA) (Poster + Lightning Talk).
-
Aug 2026
Our paper Beyond the GUI Paradigm: Do Mobile Agents Need the Phone Screen? was accepted to the ECCV 2026 Workshop on Multimodal Digital Agents (MDA) (Poster).
-
Aug 2026
Our paper Learning to Adapt: Few-Shot Generalization of Mobile Agents to Unseen Applications was accepted to the ECCV 2026 Workshop on Multimodal Digital Agents (MDA) (Poster + Lightning Talk).
-
Jul 2026
Joined Mila — Quebec Artificial Intelligence Institute as a Graduate Student Researcher, working on multimodal / GUI agents and RL-based post-training.
-
Dec 2025
Our paper Agent-SAMA: State-Aware Mobile Assistant was accepted to AAAI 2026.
-
2025
MAPLE was accepted to the COLM 2025 Workshop.
-
2025
GUIWatcher was accepted to ICSE 2025.
-
Jan 2025
Started as a Teaching Assistant at Concordia University.
-
Jan 2024
Began my PhD and joined the SPEAR Lab in Montreal.
Research at the intersection of AI and software systems.
My work focuses on intelligent mobile software systems — designing agents that combine large language models, vision-language pipelines, and finite-state reasoning to automate and test complex mobile applications, and to stay state-aware, robust, and recoverable when things go wrong. I'm also a Graduate Student Researcher at Mila — Quebec Artificial Intelligence Institute, working on RL-based post-training (GRPO, DPO) for multimodal GUI agents.
Before my PhD I built and maintained microservices for customer personalization and recommendation at Walmart Global Tech, integrating GenAI into search and tuning Milvus on Kubernetes. I hold an MS in ECE (AI track) from the University of Miami and a BS in Electrical Engineering from Missouri.
- Montreal, Canada
- PhD @ Concordia University
- LLM agents & mobile automation
- linqg327@gmail.com
Education
-
2024 — Present
PhD, Software Engineering
Concordia University · Montreal, CA
-
2020 — 2022
MS, Electrical & Computer Engineering (AI track)
University of Miami · Miami, US
-
2016 — 2019
BS, Electrical Engineering
University of Missouri Columbia · US
Selected Papers
-
ECCV 2026 WS
Learning to Adapt: Few-Shot Generalization of Mobile Agents to Unseen Applications
Best Paper Award · ECCV 2026 Workshop on Multimodal Digital Agents (MDA) · Spotlight Talk + Poster
Studies how a mobile GUI agent can adapt to an app it has never seen from just two user demonstrations, introducing the AndroidControl-AppAdapt benchmark and two episodic training methods, FO-MAML and R-ICL-Train. Meta-learned FO-MAML gives the strongest gains on unseen apps (68.79% goal-level success, +3.21 over the fine-tuned baseline), and a key finding is that matching training to the test-time adaptation mechanism matters more than the mechanism itself.
-
ECCV 2026 WS
Q-Reflector: Learning GUI Action Reflection with Generated Conditions
ECCV 2026 Workshop on Multimodal Digital Agents (MDA) · Poster + Lightning Talk · arXiv · Demo · Code
A standalone Qwen3-VL-8B reflector that first generates an action’s pre- and post-conditions, then checks the before/after screenshots to judge whether the action succeeded and actually moved the task forward. It reaches 82.16% accuracy on 1,082 verified AndroidWorld transitions and lifts MobileWorld task success by up to 17 points when plugged into GUI agents.
-
ECCV 2026 WS
CoAdapt-GUI: Joint Workflow Context and Policy Adaptation for Unseen GUI Applications
ECCV 2026 Workshop on Multimodal Digital Agents (MDA) · Poster + Lightning Talk · arXiv
A test-time adaptation framework that lets a mobile GUI agent adapt to an app it has never seen, jointly refining a transferable workflow context and a LoRA policy from its own rollouts and rewards. It reaches 45.0% on AndroidWorld-Generalization (vs. 37.5% for policy-only adaptation) and raises AndroidWorld Plus from 38.6% to 52.9%.
-
ECCV 2026 WS
Beyond the GUI Paradigm: Do Mobile Agents Need the Phone Screen?
ECCV 2026 Workshop on Multimodal Digital Agents (MDA) · Poster · arXiv
Asks whether mobile agents really need the screen by letting off-the-shelf coding agents operate Android through the command line instead of the GUI. Claude Code reaches 71.8% on AndroidWorld and 51.9% on MobileWorld, beating every reproducible GUI baseline, and every CLI agent wins on the new CLI-Advantage suite with far fewer steps.
-
arXiv 2026
AgentPersonaBench: Benchmarking Persona-Driven User Simulation
Preprint · 2,460-task benchmark across survey, chat, web & app surfaces · arXiv · Code
A 2,460-task benchmark that tests whether a persona assigned to an agent actually steers what it does, judged only from observable actions across survey, chat, web and desktop-app environments. Leading models reach up to 84.7% full-pass adherence, yet adherence drops across interaction surfaces and under multi-attribute personas.
-
AAAI 2026
Agent-SAMA: State-Aware Mobile Assistant
A state-aware multi-agent framework that models app execution as a finite state machine — screens as states, actions as transitions — which four specialized agents build in real time to plan, verify each step and recover from errors. It reaches 84.0% success on Mobile-Eval-E, 80.0% on SPA-Bench and 63.7% on AndroidWorld, improving task success by up to 12% and recovery by 13.8% over prior methods.
-
ICSE 2025
GUIWatcher: Automatically Detecting GUI Lags by Analyzing Mobile Application Screencasts
ICSE 2025 · arXiv
A framework that detects GUI lags in screencasts recorded during mobile app testing, using computer vision to find janky, long-loading and frozen frames and to prioritize the most severe ones. It reaches 0.91 precision and 0.96 recall on real-world app tests and runs continuously in a production testing pipeline.
-
COLM 2025 WS
MAPLE: A Mobile Assistant with Persistent Finite State Machines for Recovery Reasoning
The workshop version of the Agent-SAMA idea: a mobile assistant that keeps a persistent finite-state-machine memory of the app’s navigation flow, with screens as states and actions as transitions, so it can check its progress and reason its way back after errors. The approach was later extended into Agent-SAMA (AAAI 2026).
-
CMIG
Registration on DCE-MRI Images via Multi-Domain Image-to-Image Translation
Computerized Medical Imaging and Graphics (IF 7.422)
-
Under Review
Screencast-Based Analysis of User-Perceived GUI Responsiveness
Under review · arXiv
Presents MobileGUIPerf, a lightweight black-box technique that measures GUI responsiveness straight from testing screencasts, computing response time and finish time for every user interaction. On 2,458 annotated interactions from 64 Android apps it detects interactions with 0.96 precision and 0.93 recall, and it analyzes thousands of screencasts daily in an industrial pipeline.
-
Under Review
PopSweeper: Automatically Detecting and Resolving App-Blocking Pop-Ups to Assist Automated Mobile GUI Testing
A lightweight, screenshot-based tool that detects app-blocking pop-ups during automated GUI testing and locates the control to close them — such pop-ups appeared in 42.5% of the recent top apps studied. Paired with DroidBot, it cut time lost to pop-ups by 59.27% and pop-up-caused redirects by 80%.
Where I've Worked
-
MilaJul 2026 — Present
Graduate Student Researcher
- Research on multimodal / GUI agents.
- RL-based post-training for agentic models (GRPO, DPO).
-
SPEAR LabJan 2024 — Present
Research Assistant
- Research intelligent mobile software systems with a focus on AI-driven automation and UI understanding.
- Develop methods to improve agent reasoning, state awareness, and robustness in complex mobile environments.
- Co-authored multiple papers at AAAI, ICSE, TOSEM, and COLM.
-
ConcordiaJan 2025 — Jan 2026
Teaching Assistant
- Assist teaching Software Engineering: software design, UML modeling, testing, project management.
- Lead weekly tutorials and support students via office hours and online discussions.
- Guide students through core SE concepts and GitHub-based project workflows.
-
WalmartAug 2022 — Dec 2023
Software Developer
- Built and maintained microservices for customer personalization, data ingestion, and recommendation using Java, Spring, and cloud tools.
- Optimized performance of the Milvus vector database on Kubernetes.
- Integrated GenAI workflows into search, improving query understanding and relevance ranking.
- Deployed scalable API endpoints and streamlined CI/CD with data science and platform teams.
-
CPECCJan 2020 — Feb 2020
Software Engineer Intern
- Developed RESTful web services with Spring Boot, Spring MVC, Maven, and Hibernate.
- Tested features with JUnit and Mockito; used Git/GitHub for version control and review.
Tools & Tech
-
01
Languages
What I think in
-
02
AI / ML
What I train & evaluate
-
03
Backend & Platforms
What I’ve shipped at scale
Honors & Awards
-
2026
Best Paper Award — ECCV 2026 Workshop on Multimodal Digital Agents (MDA), for Learning to Adapt.
-
2023
Concordia University International Tuition Award of Excellence
-
2023
Concordia University Split Graduate Fellowship
Beyond Work
Life with
Pocky
Coffee, long walks, and my dog Pocky — a look at the moments away from the keyboard.
Step insideLet's Connect
Interested in collaborating on AI agents, mobile automation, or software-engineering research? I'd love to hear from you.
Email me