PerceptUI: LLM Agents as Human-Aligned Synthetic Users for UI/UX Evaluation

PerceptUI is a new AI framework that uses persona-conditioned large language models to evaluate user interfaces by simulating how specific users would respond to UX questions. The system achieves human-level accuracy through contrastive learning and prompt evolution, potentially accelerating product development by reducing reliance on costly human testing and A/B tests.

PerceptUI addresses a genuine friction point in product development: UI/UX evaluation remains expensive and time-consuming despite being critical for success. Traditional approaches require recruiting participants or running extended A/B tests, creating bottlenecks during early-stage iteration. This framework shifts evaluation earlier in the development cycle by training multimodal language models to predict user responses with persona-specific conditioning.

The technical innovation lies in its two-stage training approach. Contrastive reflection fine-tuning extracts lessons from human decisions by analyzing teacher-generated rationales, while reflective prompt-evolution learns from model failure traces to improve accuracy. This addresses a key limitation of previous LLM-based evaluators: they tend to produce generic critiques reflecting model biases rather than authentic user perspectives. The ability to generate natural-language rationales alongside predictions makes the system more interpretable and actionable for designers.

For product teams and design-focused companies, this represents a meaningful efficiency gain. Earlier, cheaper feedback enables faster iteration cycles and reduces dependency on expensive user research during prototyping phases. However, this tool complements rather than replaces human testing entirely—complex emotional responses, accessibility concerns, and edge cases still require real user validation.

The framework's demonstrated generalization to unseen questions and personas, combined with population-level response distribution capability, suggests practical applicability across industries. As LLMs continue improving, persona-conditioned evaluation could become standard in design workflows, particularly for teams with resource constraints or rapid iteration requirements.

• →PerceptUI enables persona-specific UI/UX evaluation by training multimodal LLMs to simulate user responses with human-level accuracy.
• →Two-stage training combining contrastive reflection and prompt evolution reduces model bias and improves prediction reliability.
• →The framework accelerates product development by providing cheaper, earlier feedback than traditional human testing and A/B tests.
• →System generalizes to unseen questions and personas while generating interpretable natural-language rationales for designer feedback.
• →Practical deployment could become standard for design teams, particularly those with limited user research budgets or tight iteration timelines.