Amirhossein Abaskohi

Amirhossein Abaskohi

Computer Science PhD Student at UBC

About me

Greetings! I’m Amir, short for Amirhossein Abaskohi (امیرحسین عباسکوهی). Feel free to call me Amir 😄.

I am currently embarking on my journey as a second-year PhD student at the University of British Columbia, under the supervision of Prof. Giuseppe Carenini and Prof. Peter West. I am part of the UBC Natural Language Processing Group.

My current research sits at the intersection of two lines of work. The first is the efficiency and accuracy of deep research agents, especially for enterprise systems: how we build, evaluate, and improve agents that can carry out long-horizon, evidence-grounded investigations reliably. This includes better benchmarking of deep research agents (DRBench) and efficient retrieval-augmented generation systems (MCompassRAG and SproutRAG). Within this line, I am currently exploring how such agents prematurely commit to conclusions before enough evidence is available, and how to predict and repair these premature commitments; this is the focus of my recent work DeepRewind, which treats the agent’s evolving epistemic state as a typed graph of sources, evidence, claims, hypotheses, and commitments, and uses a world model and consistency monitor to block risky conclusions and roll them back when later evidence invalidates them. The second line of work is homogeneity in language models, i.e., how a lack of diversity in model outputs limits performance in creative, open-ended agent settings such as deep research, where the exploration plan itself needs to be diverse to be effective; this is what I explore in UnpredictaBench, which characterizes the unpredictability and randomness of different agents.

Beyond these two lines, I am also broadly interested in inference-time scaling and adaptation of world models, computer-use agents, and multi-agent debate systems — directions I see as closely tied to making agentic systems both more efficient and more reliably diverse in how they reason and plan. I also remain interested in natural language inference and reasoning, long multimodal document reasoning, controlled generation, and information retrieval — threads from my earlier work that continue to inform how I approach agentic design today.

I earned my Master of Science in Computer Science from UBC with distinction (A+), supervised by Prof. Giuseppe Carenini and Prof. Issam H. Laradji. My research centered on reasoning over long, multimodal documents, spanning both how such documents are represented and how they are queried. My thesis, Multimodal Understanding of Long Documents: From Topic Modeling to Question Answering, develops this in two directions: CEMTM, a topic modeling technique for long documents that combines contextual embeddings with cross-modal alignment, and FM²DS, a synthetic-data-generation pipeline for multimodal, multi-hop question answering in low-resource settings.

Prior to that, I earned my BSc in Computer Engineering from the University of Tehran. During my undergraduate studies, I had the privilege of collaborating with Prof. Azadeh Shakeri, Prof. Yadollah Yaghoobzadeh, and Prof. Behnam Bahrak, enriching my academic experience.

Throughout this journey, I have completed internships at Microsoft Research, Huawei Research, and ServiceNow Research, and collaborated with researchers at Google, Meta AI, and Salesforce AI. These experiences have enriched my research perspective and broadened my understanding of cutting-edge developments in natural language processing.

Beyond academia, I am enthusiastic about knowledge dissemination and community engagement. I share insights and discoveries through Medium and YouTube, fostering dialogue and knowledge exchange. I have also honed my software engineering skills through practical work, contributing to back-end development at System Group in Summer 2022 and MateMachine for a year starting November 2022.

When I’m not working or studying, I love getting lost in the world of novels and comics 📚. They spark my imagination and give me a fun break from my busy life 🌟. I don’t spend a lot of time outdoors, but I really enjoy it when I do. Living in Vancouver, with its amazing natural beauty, gives me plenty of chances to explore. Whether it’s a quiet hike in the forest, a walk along the beach, or watching a colorful sunset, being in nature always makes me feel calm and refreshed. Even though life can be hectic, these little moments in nature remind me to slow down and appreciate the world around me. 🌿✨

I am happy to answer any questions and chat about research/collboration! Reach me at aabaskoh@cs.ubc.ca.

News

Misc

  • My path into natural language processing began with a broader fascination: the origins of human language and how tongues have formed, spread, and transformed across history. That curiosity still shapes how I approach NLP, which sits at the intersection of linguistics, computer science, and AI. Two books capture this interest especially well for me, Guy Deutscher’s The Unfolding of Language, on how grammatical structure emerges and evolves over time, and Nicholas Ostler’s Empires of the Word, a sweeping history of how languages have risen and spread across civilizations.

  • In my free time, I enjoy cooking. I am far from an expert, but I love experimenting in the kitchen and playing with flavors until something turns out just right.

  • Whenever I get the chance to step outside, I take it. Vancouver’s landscapes make that easy, and a bit of time in the forest or by the water is usually all I need to reset and slow down. 🌿

  • I am a devoted horror fan, most of all in fiction. I have made my way through nearly all of Stephen King’s books, and I am always happy to talk shop about them 📚. I am also deep into the Marvel universe, following it across both the films and the comics.

  • I am a big fan of gaming, especially story-driven titles that pull you into a rich narrative. Survival horror is my particular soft spot, where the tension and puzzles keep me hooked. If you are a gamer too, feel free to check out my Steam profile!