Amin Tabrizian

PhD candidate in Computer Science
The George Washington University

amin_tabrizian [at] gwu [dot] edu

CV / Scholar / GitHub / LinkedIn / Calendar

I'm a PhD candidate in Computer Science at George Washington University, where I work under Prof. Peng Wei in the Intelligent Aerospace Systems Lab. Most of my research is about LLM agents and reinforcement learning in aviation and autonomous driving, and about how to tell whether they are safe enough to use there. My flight planning work is funded by NASA and done together with MIT Lincoln Laboratory and Honeywell.

This fall I'm teaching Large Language and Vision Models (MAE 6291 / CSCI 6907), a graduate course at GW. Since June I've also been an AI engineer intern at Cignus Consulting, where I build an LLM agent for aviation intelligence and work on its AI safety.

Before GW, I studied mechanical engineering at the University of Tehran.

News

Aug 2026New preprint, A Paragraph is Worth a Thousand Captions, on training vision-language retrievers with long text.
Jul 2026Two papers at ICML 2026 workshops: FRAMe at LM4Plan and EO-Agents at AI for Science.
Jun 2026Joined Cignus Consulting as an AI Engineer Intern.
Jun 2026Our paper on LLM-augmented RL for air traffic management is out in Transportation Research Part E.
Jun 2026Paper on LLM safety assessment at non-towered airports at AIAA Aviation 2026.
May 2026Delay-aware on-ramp merging paper accepted to the IEEE Internet of Things Journal.
Dec 2025Invited talk at Honeywell Advanced Technology on LLM flight planning with RAG memory and a multimodal coach agent.
Oct 2025Talk at the INFORMS Annual Meeting in Atlanta.
Jul 2025Neil Y. Chen Memorial Best Student Paper Award (Air Transportation Systems) at AIAA Aviation 2025 for the Chain-of-Thought Flight Planner. Four of our papers were at the conference.

Earlier

2025Organized the NASA System-Wide Safety Program Workshop.
2025Talk at the AAAI Spring Symposium Series in San Francisco.
2024LLM flight planning paper at IEEE/AIAA DASC in San Diego.
2024Survey on reinforcement learning in aviation published in Engineering Applications of Artificial Intelligence.
2024Three papers at IEEE conferences (iThings, ICCCN, MetaCom).
2024Talk at the INFORMS Annual Meeting in Seattle.
2024Organized the Safe and Reliable AI Workshop.

Selected papers

Full list on Google Scholar.

Diagram of the delay-aware encoder feeding a soft actor-critic merging agent, with a roadside unit sending delayed observations

IoT-J

Delay-Aware Reinforcement Learning for Highway On-Ramp Merging under Stochastic Communication Latency

A. Tabrizian, Z. Huang, A. Aziz, P. Wei

IEEE Internet of Things Journal, 2026

A roadside unit tells the merging car where the other cars are, but that picture can be up to two seconds old. We treat merging as a random-delay MDP and give a SAC agent a GRU encoder that recovers the current state from stale observations and its own recent actions, plus a physics-based safety check. Over 99% merge success in dense SUMO traffic built from NGSIM data.

paper arXiv code

System diagram of the planner LLM, multimodal coach agent, and retrieval memory

ICML-W

End-to-End LLM Flight Planning with RAG-based Memory and Multi-modal Coach Agent

A. Tabrizian, A. Aziz, A. Ullah, M. Ghazanfari, P. Razzaghi, P. Wei

ICML 2026 Workshop on Planning in the Era of Language Models (LM4Plan)

An LLM writes eVTOL flight plans as lists of waypoints. A second, multimodal agent looks at a rendered map of each plan and critiques it, and a retrieval memory brings back earlier plans and their reviews. Across four planner models, the full system gives the most valid plans, up to 93.8%.

arXiv workshop code

Grid of radar charts comparing flight time, near mid-air collisions, loss of well clear, and speed changes across trained multi-agent RL policies

Framework diagram: an LLM proposes reward weights for MARL training in the BlueSky simulator, then recommends a trained policy from user preferences and feedback

TR-E

Adaptive Air Traffic Management via Large Language Model Augmented Reinforcement Learning

S. Chen, A. Tabrizian, P. Wei, Y. Wang, X. Qu

Transportation Research Part E: Logistics and Transportation Review, vol. 214, 2026

For urban air mobility traffic, an LLM proposes reward-weight settings offline, and the system later picks among pre-trained multi-agent RL policies to match what the user cares about. This takes most of the manual reward tuning out of the loop.

paper

Pipeline that turns one image caption into positive captions, negative captions, and a description paragraph for contrastive fine-tuning

arXiv

A Paragraph is Worth a Thousand Captions: Rethinking Text Supervision for Vision-Language Retrieval

M. Ghazanfari, A. Tabrizian, A. Aziz, B. Wang, P. Wei

arXiv preprint, 2026

CLIP-style retrievers learn from short captions and struggle with long descriptions. Fine-tuning only BLIP's text encoder on synthetic paragraphs for 500K images matches Long-CLIP-L on ShareGPT4V and beats it by more than 14 points on DOCCI, with no change to the architecture.

arXiv

Knowledge graph of NASA Earth observation datasets, platforms, and instruments

Pipeline from knowledge graph to graph neural network to three LLM agents that filter, generate, and judge hypotheses

ICML-W

EO-Agents: A Three-Agent LLM Pipeline for Earth Observation Hypothesis Generation

M. Ghazanfari, A. Tabrizian, A. Mehrabian, P. Wei

ICML 2026 Workshop on AI for Science

A graph neural network over NASA's Earth observation knowledge graph suggests which datasets to study together, and three LLM agents filter the pairs, write hypotheses, and grade them. From 1,475 datasets we get 160 hypotheses. The ranking holds up across different judge models even though the raw scores do not.

arXiv workshop

VFR sectional chart of the San Francisco Bay Area around Half Moon Bay

Framework diagram: radio transcripts, weather, ADS-B tracks, and charts go into a language model that outputs a safety classification and reasoning

AIAA

Towards Automated Air Traffic Safety Assessment Around Non-Towered Airports Using Large Language Models

T. Darrell, M. Ghazanfari, J. K. Kam, A. Bayen, A. Tabrizian, P. Wei

AIAA Aviation Forum, 2026

At airports without a tower, pilots keep each other safe by announcing themselves on a shared radio frequency. We give radio transcripts, weather reports, ADS-B tracks, and VFR charts to vision-language models for post-flight safety review. On a nominal vs. danger task, even open models reach a macro F1 above 0.85.

paper arXiv

Map of the Dallas Fort Worth area with wind hazard polygons, an origin, and a destination

The same map with the LLM's planned route drawn around the hazards

AIAA

Chain-of-Thought Flight Planner: End-to-End LLM Routing Under Wind Hazards

A. Tabrizian, M. Ghazanfari, P. Wei

AIAA Aviation Forum, Las Vegas, 2025

Best Student Paper Award, Air Transportation Systems

We ask an LLM to route an air taxi around wind hazard zones while following what the operator says they want, and compare five prompting strategies. The best one gives a valid plan 98% of the time.

paper pdf code

Satellite map of the Dallas Fort Worth area with wind hazard polygons between an origin and a destination

Flight Plan Recommender chat interface where the operator asks for a smooth, short flight and the assistant explains its weights

DASC

Using Large Language Models to Automate Flight Planning under Wind Hazards

A. Tabrizian, P. Gupta, A. Taye, J. Jones, E. Thompson, S. Chen, T. Bonin, D. Eberle, P. Wei

IEEE/AIAA Digital Avionics Systems Conference (DASC), San Diego, 2024

Where the flight planning work started: can an LLM read a wind hazard forecast and produce a flight plan that stays clear of it?

paper

Wheel diagram of reinforcement learning application areas in aviation

EAAI

A Survey on Reinforcement Learning in Aviation Applications

P. Razzaghi, A. Tabrizian, W. Guo, S. Chen, A. Taye, E. Thompson, A. Bregeon, A. Bahari, P. Wei

Engineering Applications of Artificial Intelligence, vol. 136, 2024

How reinforcement learning has been used across aviation, from flight planning and air traffic management to attitude control, maintenance, and revenue management.

paper arXiv

Other papers

Transformer or CNN? Benchmarking Real-Time Detection Transformer and YOLOv8 for Small UAS Autonomous Landing. C. Vu, M. Ghazanfari, K. Dong, A. Taye, A. Tabrizian, P. Wei. AIAA Aviation, 2025.
Small UAS Landing Site Detection with ArUco Markers and Deep Learning Based Computer Vision. J. Xu, C. Vu, A. Taye, A. Tabrizian, P. Wei. AIAA Aviation, 2025.
Synthetic Data Generation for Computer Vision Based Autonomous Landing for Small UAS Package Delivery. D. Ding, C. Vu, A. Taye, A. Tabrizian, P. Wei, Z. Zhao. AIAA Aviation, 2025.
RADIO: Reinforcement Learning-Aided Deployment of Wi-Fi Routers in 5G Networks for Indoor Drone Orchestrating. A. Famili, A. Tabrizian, T. Atalay, A. Stavrou. IEEE iThings, Copenhagen, 2024.
RAPID: Reinforcement Learning-Aided Femtocell Placement for Indoor Drone Localization. A. Famili, A. Tabrizian, T. Atalay, A. Stavrou, P. Wei. IEEE ICCCN, Big Island, 2024.
GREEN: Precise Geolocation in Metaverse using Reinforcement Learning-Enabled Sensor Placement. A. Famili, T. Atalay, A. Tabrizian, A. Stavrou. IEEE MetaCom, Hong Kong, 2024.

Teaching

Large Language and Vision Models (MAE 6291 / CSCI 6907) Fall 2026

Instructor, The George Washington University

Linear System Dynamics (MAE 3134) Spring 2026

Teaching Assistant, The George Washington University

Measurement Systems and Instrumentation Fall 2021