PaperlayerView exact revision on arXiv

Source-linked AI summary

NeuroDivSim: An Interactive Tool for Model-Based Reflection on Cognitive Diversity in Interface Design

Eske Beckefeld, Henrik H. J. Detjen

arXiv:2610.11590v1cs.HC

TL;DR

Simulated users raise concerns about validation, stereotyping, and representational bias, while designers need to translate knowledge about cognitive diversity into concrete interface decisions between empirical evaluations. NeuroDivSim uses an LLM to propose inspectable task, interface, and environment models for human review, then deterministically simulates them with explicit cognitive reference configurations. Participants rated both understanding the system’s basic purpose and completing the intended tasks M=4.80 (SD=0.42) on five-point scales.

  • Problem

    Simulated users raise concerns about validation, stereotyping, and representational bias, while designers need to translate knowledge about cognitive diversity into concrete interface decisions between empirical evaluations.

    Source: §1 · p. 2

  • Method

    NeuroDivSim uses an LLM to propose inspectable task, interface, and environment models for human review, then deterministically simulates them with explicit cognitive reference configurations.

    Source: §1 · p. 2, §3 · p. 3, §3.1 · p. 5

  • Results

    Participants rated both understanding the system’s basic purpose and completing the intended tasks M=4.80 (SD=0.42) on five-point scales.

    Source: §4 · p. 8

  • Takeaways & Limitations

    Designers can hold a modeled usage situation constant while varying cognitive assumptions and tracing their consequences to interaction steps and recommendations.

    Source: §1 · p. 2, §5 · p. 9

    The pilot does not establish effects on design decisions or interface quality, nor correspondence between simulation and observed behavior.

    Source: §5.3 · pp. 9–10

Abstract

from arXiv · show

Recent approaches to simulated and synthetic users offer new ways to support design, but raise questions about how computational representations of users should contribute to design practice. We present NeuroDivSim, an interactive tool that explores simulation as an inspectable mechanism for reflecting on cognitive diversity during design and prototyping. Rather than using an LLM to act as a simulated user, NeuroDivSim uses generative AI to construct inspectable task, interface, and environment models from a usage scenario. After human review, these models are combined with explicit cognitive reference configurations and processed through deterministic simulation. This enables designers to hold a modeled usage situation constant while varying cognitive assumptions and tracing their consequences to interaction steps and rule-based design recommendations. We further report an exploratory pilot evaluation (N=10) that provided formative insights into how participants engaged with the workflow and informed subsequent refinements to the presentation of models, simulation results, and recommendations.

1 Introduction

Simulated users can add design perspectives, but their use raises concerns about validation, stereotyping, and representational bias, especially for heterogeneous populations. NeuroDivSim instead makes modeled assumptions inspectable and tests them through deterministic simulation, with a pilot informing refinements to its presentation.

1 Introduction

  • Using simulated users as representations of human participants raises validation, stereotyping, and representational-bias concerns, especially for heterogeneous populations.Neurodiversity highlights the challenge: designers must translate research, guidelines, and user involvement into concrete decisions between opportunities for empirical evaluation.Source: §1 · p. 2
  • NeuroDivSim uses an LLM to create inspectable task, interface, and environment models, then applies cognitive reference configurations through deterministic simulation.Designers can hold the modeled situation constant while varying assumptions and tracing consequences to interaction steps and rule-based recommendations.Source: §1 · p. 2
  • The tool supports reflection on modeled assumptions rather than predicting how neurodivergent people will behave.Its architecture makes consequences of selected assumptions explicit while a design can still be revised.Source: §1 · p. 2
  • An exploratory pilot with ten prospective users provided formative evidence about workflow engagement and informed refinements to the presentation of models, results, and recommendations.The paper frames simulation as an inspectable reflection mechanism, not a surrogate for empirical participants.Source: §1 · p. 2

2 Related Work

The related work motivates modeling cognitive diversity without treating diagnoses as behavioral specifications. NeuroDivSim combines task-step inspection and counterfactual comparison while distinguishing its approach from synthetic users that can raise representation and validity concerns.

2.1 Neurodiversity and Ability-Based Design

  • Neurodivergent people vary in characteristics, strategies, contexts, and access needs, motivating interaction-relevant dimensions rather than generalized deficits or diagnoses as behavioral specifications.Participatory, strengths-based work and Ability-Based Design foreground capabilities in context and system responsibility to accommodate heterogeneous abilities.Source: §2.1 · pp. 2–3

2.2 Model-Based Inspection and Reflection

  • Model-based inspection can support counterfactual reasoning by holding parts of an interaction constant while varying assumptions; NeuroDivSim structures this around task steps and compares parameterized cognitive and contextual assumptions.Its purpose is identifying where assumptions produce a modeled concern, not calibrated prediction.Source: §2.2 · p. 3

2.3 Synthetic Users and the Boundaries of Simulation

  • Synthetic-user approaches can add perspectives but raise concerns about validity, stereotyping, representational bias, transparency, and substitution for empirical research.NeuroDivSim instead uses an LLM to construct an editable usage model and explicit configurations and rules to expose how assumptions propagate, without claiming individual behavior.Source: §2.3 · p. 3

3 Design of NeuroDivSim

NeuroDivSim separates usage-scenario modeling from cognitive assumptions and deterministic simulation, so designers can vary explicit assumptions while holding a modeled situation constant. Its stepwise outputs and rule-based recommendations expose how modeled conditions lead to design concerns.

3 Design of NeuroDivSim

  • The system separates scenario interpretation from simulation: designers inspect and revise structured task, interface, and environment models before combining them with cognitive assumptions.The LLM maps scenario information and optional screenshots to model attributes and task steps; after model confirmation, processing is deterministic.Source: §3 · pp. 3–4, §3.1 · p. 5

3.1 Formalizing the Usage Situation

  • Task, interface, and environment models keep distinct sources of interaction demand separate, with task steps providing structure for later simulation events and recommendations.Attributes use a shared 0–100 scale and can be reviewed and revised before confirmation.Source: §3.1 · p. 5

3.2 Representing Cognitive Assumptions

  • ADHD-informed and dyslexia-informed reference configurations apply predefined cognitive parameters to the same modeled situation, rather than representing individuals or population averages.Their numerical values are heuristic operationalizations, not empirically estimated characteristics of diagnostic populations.Source: §3.2 · p. 6

3.3 Deterministic Cognitive Simulation

  • The simulation tracks dynamic cognitive states and derives metrics and threshold-based events across task steps, allowing aggregate differences to be traced to modeled conditions.In the illustrated run, attention decreases while fatigue increases; event markers identify steps with conditions such as elevated error risk or reduced attention.Source: §3.3 · pp. 6–7
  • In the hotel-booking run, simulated completion time was 399 seconds for the ADHD-informed configuration, 389 seconds for dyslexia-informed, and 373 seconds for generic; aggregate error-risk values were 88.75, 86.72, and 81.07, respectively.These are model outputs, not measured user data or predictions of how people with ADHD or dyslexia would perform.Source: §3.3 · p. 7

3.4 From Simulation States to Design Recommendations

  • Rule-based recommendations link a potentially demanding interaction step to a design response and retain the task step, modeled factors, explanation, and suggestion that triggered it.The LLM does not generate recommendations; designers can inspect why a step was flagged and modify the interface, scenario model, or assumptions before rerunning the simulation.Source: §3.4 · pp. 7–8

4 Formative Pilot Evaluation

In an exploratory pilot with ten prospective users, participants reported high ratings for understanding the system's purpose and completing tasks, while result presentation and findability received the lowest rating. The formative findings informed interface refinements but do not establish simulation validity or effects on design outcomes.

4 Formative Pilot Evaluation

  • The exploratory pilot involved ten participants with software, web, or UX/UI design experience and examined whether prospective users could work with and interpret the workflow.Participants inspected generated models, ran simulations, explored configurations and recommendations, and completed ratings and open-ended questions.Source: §4 · p. 8
  • M=4.80 on five-point scales for both understanding the system’s basic purpose and completing the intended tasks; result presentation and findability received the lowest rating, M=4.00 (SD=0.68).The UEQ-S showed a positive descriptive tendency: pragmatic mean M=1.65, hedonic mean M=2.20, and overall mean M=1.92; these are not standardized benchmark results.Source: §4 · p. 8
  • Participants highlighted structured scenario models, configuration comparisons, and recommendations, while reporting difficulty interpreting metrics and events and relating them to interaction steps.These findings informed refinements to step-centered result presentation, event links, and explanations of attributes, metrics, events, and recommendations.Source: §4 · p. 8
  • As a formative pilot, the findings provide initial insights into workflow engagement and interpretation but do not establish simulation validity or effects on design quality and accessibility outcomes.The study used an earlier prototype version.Source: §4 · pp. 8–9

5 General Discussion

NeuroDivSim is framed as a reflection tool, not a proxy for neurodivergent people: it makes modeled assumptions inspectable while preserving limits on what its outputs establish.

5 General Discussion

  • NeuroDivSim compares explicit cognitive assumptions while holding a modeled interaction constant, tracing their consequences to steps and recommendations.Its purpose is to identify which aspects of an interaction warrant closer examination under current assumptions, alongside accessibility guidance and research with neurodivergent people.Source: §5 · p. 9, §5.1 · p. 9

5.1 Simulation as Design Reflection

  • An LLM constructs a revisable scenario representation, while deterministic downstream states, events, and recommendations make comparisons reproducible and traceable.The trace exposes why an interaction became a candidate for reconsideration, and designers can inspect and correct the system’s interpretation.Source: §5.1 · p. 9

5.2 Inspectability and Boundaries of Cognitive Modeling

  • Generic, ADHD-informed, and dyslexia-informed configurations model shared interaction-relevant dimensions, not individuals or homogeneous populations.Their diagnosis-informed presets are provisional entry points, not behavioral specifications.Source: §5.2 · p. 9
  • Values, weights, thresholds, and relationships are heuristic and uncalibrated, so reproducible precision does not establish empirical validity or model lived experience.The components remain inspectable and replaceable, but transparency alone does not validate them; the system must remain alongside participatory and empirical methods.Source: §5.2 · p. 9
  • NeuroDivSim should not claim empathy, certify accessibility, rank groups, or replace research with neurodivergent people.Diagnosis-labeled presets, precise-looking scores, and recommendations may amplify stereotypes, overreliance, or “accessibility theater”; explanations alone do not prevent automation bias [7].Source: §5.2 · p. 9

5.3 Limitations and Future Work

  • The pilot offered formative engagement insights but did not establish effects on decisions or interface quality, or correspondence with observed behavior.Future evaluation should separately test design utility against practices such as guidelines or cognitive walkthroughs, and model validity through relationships, parameter sensitivity, and calibration with neurodivergent participants.Source: §5.3 · pp. 9–10

6 Conclusion

NeuroDivSim shifts simulation from representing a user toward examining explicit assumptions about an interaction. Its workflow sits between existing user knowledge and later empirical evaluation, with design utility and model validity remaining open questions.

6 Conclusion

  • NeuroDivSim combines LLM-supported formalization of a usage situation with inspectable cognitive reference configurations and deterministic simulation to trace modeled consequences without presenting them as human experience.The paper asks whether simulation can support designers without a generative agent acting as a surrogate for a human user.Source: §6 · p. 10
  • The workflow adds a reflection step between existing knowledge about users and subsequent empirical evaluation, and its formative pilot informed workflow refinement.Design utility and model validity remain for future investigation.Source: §6 · p. 10

A NeuroDivSim Interaction Workflow

Figures 3 and 4 together provide a detailed view of NeuroDivSim’s complete interaction workflow.

A NeuroDivSim Interaction Workflow

B NeuroDivSim Interface Examples

The screenshots show selected NeuroDivSim interface stages using a hotel-booking scenario, as impressions of the interface rather than a complete walkthrough.

Scenario specification

Interpreted interaction flow

Inspectable model requirements

Simulation results overview

Detailed simulation results

  • Fig. 9 presents detailed simulation results for an ADHD-informed reference configuration.Source: Figure 9 · p. 20

Design recommendations

  • Fig. 10 shows example design recommendations for the ADHD-informed reference configuration.Source: Figure 10 · p. 21
Loading 2610.11590v1…