PaperlayerView exact revision on arXiv

Source-linked AI summary

Bridging the Gulf of Envisioning: Cognitive Design Challenges in LLM Interfaces

Hariharan Subramonyam, Roy Pea, Christopher Lawrence Pondoc, Maneesh Agrawala, Colleen Seifert

arXiv:2309.14459v2cs.HC

TL;DR

LLM interfaces offer flexible natural-language interaction, but users lack a clear account of how goals become intentions and prompts that yield appropriate outputs. The paper theorizes envisioning as this missing process, identifies capability, instruction, and intentionality gaps, and examines interface patterns intended to narrow them. It concludes that LLM interfaces should better support intention formation, task planning, and output evaluation, while noting important variation in users’ generative abilities.

  • Problem

    Research has underexplored how end-users form intentions when interacting with LLMs, despite challenges connecting goals, prompts, and outputs.

  • Method

    The paper combines cognitive science and HCI perspectives to theorize envisioning, define three gaps, and analyze existing LLM interface designs.

  • Results

    The paper characterizes the gulf of envisioning as capability, instruction, and intentionality gaps and identifies interface features that address these gaps.

  • Takeaways & Limitations

    Human-LLM interfaces should support users in developing intentions, formulating prompts, and evaluating outputs rather than relying only on unconstrained natural-language input.

  • Takeaways & Limitations

    The model is limited by substantial differences among people in identifying intentions from goals and performing open-ended generative tasks.

Abstract

from arXiv · show

Large language models (LLMs) exhibit dynamic capabilities and appear to comprehend complex and ambiguous natural language prompts. However, calibrating LLM interactions is challenging for interface designers and end-users alike. A central issue is our limited grasp of how human cognitive processes begin with a goal and form intentions for executing actions, a blindspot even in established interaction models such as Norman's gulfs of execution and evaluation. To address this gap, we theorize how end-users 'envision' translating their goals into clear intentions and craft prompts to obtain the desired LLM response. We define a process of Envisioning by highlighting three misalignments: (1) knowing whether LLMs can accomplish the task, (2) how to instruct the LLM to do the task, and (3) how to evaluate the success of the LLM's output in meeting the goal. Finally, we make recommendations to narrow the envisioning gulf in human-LLM interactions.

1 INTRODUCTION

LLMs offer broad, open-ended capabilities, but users must carefully formulate intentions and prompts to align outputs with their goals. The paper names the resulting distance between intentions and prompts the gulf of envisioning, comprising capability, instruction, and intentionality gaps.

  • LLMs can generate novel, coherent, contextually relevant content across tasks including essays, artwork, and software interfaces.
  • Vague intentions or prompts can produce generic, irrelevant, or off-topic responses, making effective prompting challenging.Iteration may improve outputs but can become inefficient for longer generations.
  • LLM interfaces vary along functional flexibility, intent specificity, and output determinacy, including open-ended generation with multiple possible solutions.
  • The gulf of envisioning is the distance between users’ initial intentions and prompts that anticipate how LLM capabilities and training data can produce high-quality output.
  • The three envisioning challenges concern whether the LLM can accomplish a task, how to instruct it, and what output to expect.

2 INTENTIONS AND INTERACTIONS IN CONVENTIONAL SOFTWARE SYSTEMS

Conventional interaction models distinguish goals, intentions, action specification, execution, and evaluation, but intention formation has received comparatively little attention. The paper links intention formation to cognitive task processes and system mental models that jointly connect goals with actions and outcomes.

  • Norman’s seven stages include establishing a goal, forming an intention, specifying and executing actions, perceiving and interpreting system state, and evaluating outcomes.
  • Intentions translate abstract goals into concrete action sequences and draw on declarative knowledge, procedural knowledge, decisions, and cognitive skills.
  • Users rely on system mental models to predict and explain how interface actions will affect interaction outcomes.
  • Intention formation and action specification interact: users develop task intentions while simulating how the system may execute selected actions.
  • LLM interactions are challenging because the link between user intentions and system actions is less clear and users lack adequate mental models of LLMs.

3 ENVISIONING INTENTIONS IN LLM INTERACTIONS

The paper theorizes envisioning as the cognitive process of developing goals into prompt specifications for generative LLM tasks. It identifies capability, instruction, and intentionality gaps arising from uncertain capabilities, difficult instruction, and evaluation without conventional planning and execution supports.

  • LLM interfaces replace preset actions with unconstrained natural-language input, expanding the operational scope through which users express intentions.
  • Generative tasks involve variable, repetitive cycles of cognitive processes rather than a definitive process for achieving a goal.
  • Envisioning develops prompt specifications by planning topics, structure, constraints, tone, style, and desirable qualities before asking the LLM to generate output.
  • More developed intentions may improve output satisfaction, but formulating them requires effortful mental exploration of possible plans and values.
  • Iteration can incur user-experience and conversational costs while increasing fixation on early outputs and limiting exploration of alternatives.
  • Instruction gap: The instruction gap reflects difficulty expressing intentions precisely because small wording changes can produce substantially different outputs.
  • Intentionality gap: The intentionality gap occurs when users evaluate outputs without the planning and execution processes that would provide task context and a cohesive mental model.
  • The three gaps together comprise the gulf of envisioning, linking user intentions to successful LLM execution and evaluation.

4 EXAMINING THE ENVISIONING GAP IN THREE LLM INTERFACES

The paper analyzes three existing LLM interfaces through planning, execution, and evaluation, examining how their features address the three envisioning gaps. The supplied passage frames this analysis but does not report interface-specific findings.

  • The analysis examines three existing LLM interfaces and maps their design features to planning, execution, evaluation, and the three envisioning gaps.

4.1 Writing using ChatGPT

ChatGPT provides broad writing support and basic affordances for exploring tasks, but users still must formulate actionable intentions and learn prompt effects largely through trial and error. Its features offer guidance and organization without fully supporting planning and evaluation.

  • ChatGPT supports drafting, outlining, elaboration, grammar and style checks, paraphrasing, and synthesis across writing tasks.Its writing support is diverse and dynamically evolving.
  • Planning: Example prompts help users understand the model’s action space, but their high-level granularity does not support concrete planning.
  • Planning: Different chats organize planning and execution lines, yet they do not help users formulate intentions or break goals into actionable steps.
  • Instruction: Users mainly discover prompt interpretation through trial and error using regeneration, prompt editing, and separate chats.
  • Instruction: Without regeneration history, users struggle to compare outputs or learn how slight language changes affect model interpretations.
  • Evaluation: Custom instructions align model responses with user context but require users to foresee intentions upfront, while output assessment can remain difficult without a comprehensive mental model.

4.2 Creative Coding using Spellburst

Spellburst supports exploratory creative coding by combining examples, prompt assistance, semantic operators, and code-level feedback. These affordances help users understand possibilities, connect prompts to code changes, and pursue alternative creative iterations.

  • Spellburst bridges natural-language creative intents and code by generating executable computational artwork and supporting merges, variations, and branching.
  • Capability gap: Example sketches show which creative coding tasks suit Spellburst, while autocomplete and semantic operators expose possible directions for extending ideas.
  • Instruction gap: Interleaved code comments identify where a user’s prompt produced a specific code change, making prompt-output relationships more visible.
  • Instruction gap: Spellburst frames the LLM as an expert creative coder and uses a crowd-sourced taxonomy to drive prompt autocomplete and steer executable code generation.

4.3 Software Development using Cursor

Cursor integrates LLM assistance into a code editor through examples, contextual interaction modes, codebase questions, and references. These affordances support programming workflows but leave specific planning and input-output comparison challenging.

  • Cursor pairs programming with an LLM by supporting project conversations, code changes, and bug fixing within a VSCode-based editor.
  • Planning: Users can interact through selected text, keyboard shortcuts, the editor, and a dedicated chat sidebar.
  • Planning: Example files and tooltips show available tasks, actions, and context-setting options, but high-level examples do not support specific planning.
  • Capability gap: Codebase questions, inline comments, and rerunning without context help users inspect language interpretation, although missing regeneration history complicates input-output comparisons.
  • Intentionality gap: References to snippets, files, and documentation reduce context-setting effort and steer style or implementation choices without inherently providing a mental model of LLM operation.

5 RECOMMENDATIONS FOR DESIGNING INTERFACES FOR ENVISIONING

The paper proposes interface patterns that help users envision LLM interactions by externalizing prompt-output histories, suggesting prompts, exposing alternatives, explaining outputs, tailoring prompting, and preserving manual control.

  • Design Pattern 1 – Visually Track Prompts and Outputs: Visually tracking prompts and outputs through connected nodes lowers cognitive load and supports deliberate prompt authoring across alternative pathways.
  • Design Pattern 2 – Suggest Ideas for Prompting: Prompt suggestions and examples assist unfamiliar users, support ideation, and emphasize clarity and precision in language.
  • Tenets: Clear, precise written language helps bridge the gap between human intention and LLM output, while LLMs can serve as cognitive partners in task formulation.
  • Design Pattern 3 – Provide Multiple Outputs: Providing multiple outputs lets users compare options and connect prompt changes to changes in final outputs.
  • Design Pattern 4 – Make the Output Explainable: Explainable outputs help users understand prompt interpretation and make generated content easier to edit manually.
  • Design Pattern 5 – Use domain-specific prompting strategies: Domain-specific prompting strategies steer outputs toward usable end goals while reducing ambiguity across prompts.
  • Design Pattern 6 - Allow manual control of output: Manual control through direct editing helps users incorporate their values and intentions into exploratory LLM interactions.

6 DISCUSSION

The discussion frames envisioning as a cognitive process for translating bespoke goals into LLM intentions, prompts, and evaluations. It identifies gaps in users’ system understanding and proposes interface guidance and empirical research to support this process.

  • The authors identify capability, instruction, and intentionality gaps as distinct misalignments in specifying and evaluating LLM actions.These concern what the LLM can do, how to communicate the task, and what output to expect and accept.
  • LLM interactions differ from conventional systems because users can express bespoke goals while lacking familiar affordances for planning and evaluation.The shift is from deterministic functions toward dynamic intelligent agents.
  • Novice users may lack mental models that explain LLM processes or predict outputs from inputs.This limits their ability to formulate action specifications and evaluate whether results are good enough.
  • Future work should develop user-friendly models of LLM operations that preserve details needed to predict interaction outcomes.Suggested supports include interactive visualizations and simplified explanatory frameworks.
  • Users’ open-ended exploration can be difficult when abstract or vague inputs require iterative refinement toward quality and relevance.The discussion asks how interfaces should balance Intent Specificity, Functional Flexibility, and Output Determinacy.
  • The model is limited by substantial differences in people’s ability to identify intentions from goals and solve open-ended problems.The discussion links these differences to divergent thinking and cognitive capacities such as memory and imagination.
  • The proposed HCI direction calls for empirical comparisons of intention-explicit prompts and evaluations of AI systems’ value to users.The authors also suggest asking an LLM to repeat its understanding of task instructions.

7 RELATED WORK

The related work reviews documented LLM challenges, prompting techniques, and human-AI interface design. It positions the paper’s contribution as a cognitive account of why flexible LLMs make precise intention communication difficult.

  • 7.1 Documented Challenges of LLMs: LLM explainability is difficult because users cannot always determine how prompts affect outputs or whether apparent reasoning reflects actual model reasoning.This difficulty contributes to users’ struggles to build adequate mental models.
  • 7.1 Documented Challenges of LLMs: LLM outputs may be nonsensical, incoherent, factually incorrect, or difficult to verify, creating usability challenges across contexts.The passage notes that accuracy concerns are especially relevant in domains such as medical and military applications.
  • 7.1 Documented Challenges of LLMs: LLMs can encode racial, gender, and political biases, while jailbreaks may elicit toxic outputs or leak private information.Contributing factors include training corpora, data labeling, and model architectures.
  • 7.2 Prompting Strategies: Prompt engineering uses techniques for setting rules, structuring outputs, and guiding models toward users’ intended directions.Few-shot prompting supplies examples of expected inputs and outputs, whereas chain-of-thought supplies reasoning steps.
  • 7.3 Designing Human-AI Interfaces: Human-AI systems shift interaction from assistant-like use toward collaboration because models exhibit human-like cognitive abilities.Design guidelines therefore address action consequences, diverse model options, and accountability for errors.
  • 7.3 Designing Human-AI Interfaces: The literature has emphasized prompt strategies and prompt authoring, but the paper identifies a need to understand why precise intentions remain difficult to convey.The authors address this need with a cognitive framework for prompt-based interfaces.

8 CONCLUSION

The conclusion introduces the gulf of envisioning as a cognitive challenge in successfully formulating intentions for desired LLM responses. It divides this gulf into three gaps and connects the framework to design recommendations.

  • The paper applies a cognitive framework to characterize the dynamics and complexities of prompt-based interfaces such as ChatGPT.
  • The gulf of envisioning captures the challenge of successfully formulating intentions that elicit a desired response from an LLM.
  • The capability gap concerns what intentions an LLM can perform, the instruction gap concerns how to express them, and the intentionality gap concerns evaluating outputs.
  • The paper argues that, for LLM interfaces, intentions function as actions and should be supported through design recommendations for envisioning.
Loading 2309.14459v2…