TLDR: This research paper details insights from a design thinking workshop held at the deRSE25 conference, focusing on developing innovative user interface (UI) concepts for Large Language Models (LLMs). Participants identified common LLM use cases, evaluated existing interfaces’ strengths and weaknesses, and visualized new interaction paradigms emphasizing flexible context management, dynamic conversation branching, and enhanced user control. The findings are used to advance the authors’ own whiteboard-based UI approach, guided by a human-centered design process, advocating for greater attention to UI innovation in future LLM development.
Large Language Models (LLMs) have become indispensable tools across many fields, from writing code to generating creative content. However, despite their advanced capabilities, the way we interact with them often feels rigid and limited. Most LLM interfaces still follow a simple, linear chat format, which can hinder complex tasks and lead to frustrations like losing context or dealing with irrelevant information.
Recognizing this gap, researchers Maximilian Frank and Simon Lund organized a design thinking workshop at the deRSE25 conference. The goal was to collaboratively explore and develop innovative user interface (UI) concepts for LLMs. This initiative, detailed in their paper, aimed to move beyond the conventional chat paradigm and create more flexible, user-friendly interaction models.
Understanding User Experiences with LLMs
The workshop brought together scientists from various disciplines, many of whom regularly use LLMs. Participants shared how they use these AI tools, highlighting both their benefits and drawbacks. On the positive side, users praised LLMs for their efficiency, time-saving potential, and ability to handle monotonous tasks like writing documentation. They appreciated the models’ helpful demeanor, ease of use, and capacity to generate multiple suggestions or improve text quality.
However, users also voiced significant concerns. A major issue was the reliability and accuracy of LLM outputs, with frequent mentions of “hallucinations” (plausible but incorrect information) and “faulty code.” Other dislikes included repetitive responses, the model’s tendency to forget context in long discussions, and its limitations with less common software. Beyond functionality, ethical considerations like data privacy, copyright issues, and the environmental impact of training large models were also raised. Interestingly, while these criticisms were abundant, direct complaints about the existing UI were scarce, suggesting users have simply grown accustomed to the current linear chat format rather than finding it optimal.
Innovative UI Concepts from the Workshop
The core of the workshop involved participants sketching and presenting new UI solutions. These ideas offered fresh perspectives on how users could better manage conversations and context:
Branching Conversations: Several concepts focused on allowing users to branch off conversations. Imagine dragging a useful AI response into a new panel to start a separate, focused discussion, or hovering over a message to create a temporary sub-conversation for clarification.
Context Management: Ideas included an “auto-forget” function to remove irrelevant information from the LLM’s memory and checkboxes next to messages to selectively include or exclude them from the active context. This gives users precise control over what the AI “sees.”
Weighting and Highlighting: Participants envisioned ways to assign importance to specific parts of a conversation. This could involve pinning entire messages or highlighting text segments, influencing how much impact they have on subsequent AI responses.
Output Formatting: One concept proposed a dropdown menu to easily select desired output formats (e.g., JSON, YAML) for AI responses, moving beyond the need for complex prompt engineering.
Also Read:
- Beyond Chatbots: How LLMs Are Learning to Build Interactive User Interfaces
- Bridging Human Intuition and AI Efficiency in Design Optimization
The Authors’ Vision: A Dynamic Whiteboard UI
Building on these insights, the authors are developing their own innovative UI concept: a dynamic whiteboard interface. This system moves away from the linear chat by representing all user prompts and LLM outputs as discrete, movable elements on a virtual whiteboard. Users gain granular control through a visible “context window” where they can explicitly select, reorder, and manage the information available to the LLM. This approach aims to provide greater flexibility, transparency, and control, allowing users to save specific arrangements as “blueprints” for recurring tasks.
The development of this whiteboard UI is guided by the human-centered design (HCD) process, an iterative methodology that prioritizes user needs and feedback. The workshop’s findings, particularly on visualizing branching, implementing weighting functions, and hover-based interactions, are now informing the refinement of this concept. The researchers hypothesize that as LLM capabilities mature, UI usability and user experience will become crucial differentiators, making UI innovation a key focus for future AI development.
For more in-depth information, you can read the full research paper available at arxiv.org/pdf/2508.18784.
