Please note: This PhD defence will take place in DC 2314 and online.
Mohammad Hasan Payandeh, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Jian Zhao
Effective human–LLM communication depends on the input provided to LLMs, which shapes how models interpret users' intentions and generate responses. Prompt engineering research has explored ways to improve this input through clearer instructions, iterative refinement, and alternative formulations. While this input may include both user requests and supporting context, comparatively less attention has been given to how users can directly explore and control the context supplied to LLMs. Context refers to task-relevant information that grounds model responses, such as prior conversations, code, data, or images. In many LLM-based systems, this information is selected or assembled automatically, with limited opportunities for users to inspect, validate, or manipulate it. Consequently, the context provided to an LLM may not fully reflect the information users consider relevant to their tasks.
This thesis investigates user-driven context as an important, but underexplored, dimension of human–LLM communication through four research focuses. SenseSync highlights the importance of studying context by showing how variations in context enable information exploration. NoteEx examines the effects of user-driven context on LLM responses, showing that user-selected context can improve responses. ThinkInk investigates the necessity of user-driven context across tasks, demonstrating its essential role in LLM-assisted spatial inking, compared with its beneficial role in information exploration and response improvement in the preceding projects. Finally, DuckDuckTalk explores levels of granularity in context manipulation, showing how context can be configured through bundles that define different agents, rather than through individual entities as in NoteEx and ThinkInk. Together, these projects contribute a comparative understanding of user-driven context, interaction designs that support its manipulation, and empirical evidence of its effects across different human–LLM interactions. They demonstrate that the value and necessity of context control vary across tasks and interaction settings. More broadly, this thesis argues that improving human–LLM communication requires not only refining how users express their requests, but also enabling them to explore, inspect, and manipulate the contextual information that shapes LLM responses.
To attend this PhD defence in person, please go to DC 2314. You can also attend virtually on Zoom.