Where to Show Demos in Your Prompt: A Positional Bias of In-Context Learning
Kwesi A. Cobbina, Tianyi Zhou
EMNLP 2025 2025
Published
In-Context LearningLLMPrompt EngineeringNLP
Kwesi Adu Cobbina
PhD Candidate in Computer Science · University of Maryland, College Park
Advised by Prof. Tianyi Zhou and Prof. Tom Goldstein
I am a PhD candidate in Computer Science at the University of Maryland, advised by Prof. Tianyi Zhou and Prof. Tom Goldstein. My research sits at the intersection of natural language processing, multimodal learning, and optimization. I study how large language models encode and leverage in-context information — from positional biases in demonstration placement to efficient reasoning trace compression — with the goal of making LLMs more robust, efficient, and equitable.
College Park, MD · PhD expected January 2027
Research Interests
In-Context LearningLarge Language ModelsMultimodal LearningNatural Language ProcessingLLM Fairness & RobustnessOptimization
Advised by Prof. Tianyi Zhou and Prof. Tom Goldstein at University of Maryland, College Park
News
I'll be interning at Netflix HQ: as a Machine Learning Intern working on LLMs and recommendation systems.
Started working with Tom Gooldstein as a Graduate Research Assistant at the University of Maryland.
Paper accepted at EMNLP 2025: Where to Show Demos in Your Prompt: A Positional Bias of In-Context Learning.
Paper accepted at NeurIPS 2025: ColorBench: Towards Evaluation on Color Cognition Capabilities of LLMs.
Awarded M.Sc. in Computer Science from the University of Maryland, College Park.
Paper accepted at NAACL 2025: My LLM Might Mimic AAE—But When Should It?
Paper accepted at IEEE VIS 2024: Evaluating the Semantic Profiling Abilities of LLMs for Natural Language Utterances.
Received the Dean's Fellowship from the University of Maryland (2021–2023).
Joined TianyiLab as a Graduate Research Assistant under Prof. Tianyi Zhou.
Joined the University of Maryland as a Graduate Student.
Selected Publications
* equal contribution
ColorBench: Towards Evaluation on Color Cognition Capabilities of LLMs
Yijun Liang, Ming Li, Chenrui Fan, Ziyue Li, Dang Nguyen, Kwesi A. Cobbina, Shweta Bhardwaj, Jiuhai Chen, Fuxiao Liu, Tianyi Zhou
NeurIPS 2025 2025
Published
Multimodal LearningLLM EvaluationColor CognitionBenchmark
My LLM Might Mimic AAE—But When Should It?
Sandra C. Sandoval*, Christabel Acquaye*, Kwesi A. Cobbina*, Mohammad N. Teli, Hal Daumé III
NAACL 2025 2025— * Equal contribution
Published
Language FairnessAAELLMNLPEthics
Beyond the Prompt: Approaches to Knowledge Boundary Detection in Large Language Models — A Systematic Literature Review
Brandon C. Colelough, Davis Bartels, Dina Demner-Fushman, Ishan Tamrakar, Kwesi Cobbina, Mike Ledford, Srividya Ponnada, Xinchen Yang, Yuexi Chen
IEEE TNNLS 2026— Journal article under review
Under Review
Knowledge BoundariesLLMSurveyNLP
Selected Projects
Meta-Vector: In-Context Vector Arithmetic
A test-time meta-learning approach that steers model behavior via in-context vector arithmetic, offering a practical alternative to few-shot ICL under strict context budgets. Includes a reproducible codebase and open-source toolkit (in progress).
Meta-LearningIn-Context LearningLLMPyTorch
Positional Bias Analysis Suite for ICL
An analysis suite for quantifying LLM sensitivity to demonstration ordering and placement in prompts. Supports large-scale evaluations across multiple model families and tasks, with distilled prompt-design recommendations.
In-Context LearningEvaluationPythonNLP
ChurchCast
A cross-platform mobile app enabling churches to stream services live, schedule events, and engage communities in a single unified workflow. Features push notifications, content management, and social sharing.
MobileCross-PlatformPush NotificationsContent Management
Teaching
Web Application Development with JavaScript
Teaching Assistant·University of Maryland·Spring 2022 – Spring 2026 · 12 offeringsAll offeringsSpring 2026Fall 2025Summer 2025Spring 2025Fall 2024Summer 2024Spring 2024Summer 2023Spring 2023Fall 2022Summer 2022Spring 2022
Supported course delivery across twelve offerings of this undergraduate web development course covering HTML, CSS, JavaScript, Node.js, and responsive design. Held weekly office hours, graded assignments and exams, and mentored students on full-stack web application projects.
Web DevelopmentJavaScriptNode.jsUndergraduate
Special Topics in Computer Science
Teaching Assistant·University of Maryland·Winter 2022 – Winter 2025 · 4 offeringsAll offeringsWinter 2025Winter 2024Winter 2023Winter 2022
Assisted across four offerings of this 1-credit special topics seminar covering emerging areas in computer science. Supported course logistics, graded assignments, and provided technical guidance to student instructors.
Special TopicsSeminarUndergraduate
Special Topics in Computer Science
Teaching Assistant·University of Maryland·Winter 2022 – Winter 2025 · 4 offeringsAll offeringsWinter 2025Winter 2024Winter 2023Winter 2022
Assisted across four offerings of this 1-credit special topics seminar covering emerging areas in computer science. Supported course logistics, graded assignments, and provided technical guidance to student instructors.
Special TopicsSeminarUndergraduate
Service
Reviewing
- ACL Rolling Review (ARR) — 2026
Mentoring
- Undergraduate Research Mentor, University of Maryland (2023–present)
- AI Instructor & Mentor, ELiTE Emerging Labs (2019–2022)
Memberships
- Association for Computational Linguistics (ACL)
- Association for Computing Machinery (ACM)
- Sigma XI Scientific Research Honor Society
Honors & Awards
Dean's Fellowship — University of Maryland, College Park
2021–2023Computer Science Department Award — GIMPA
2021College Honor Award — GIMPA
2016–2018Eroll King Fellowship — ELiTE Education
2017Contact
I am happy to discuss research collaborations, speaking opportunities, or any questions about my work. The best way to reach me is by email.
kcobbina@umd.eduUniversity of Maryland, College Park · College Park, MD