Accuracy and Satisfaction in Multi-Turn LLM Dialogues for NFR Assessment

As Large Language Models (LLMs) become standard tools for software development, current evaluation benchmarks focus almost exclusively on functional correctness. This leaves a significant gap in understanding how well these models handle Non-Functional Requirements (NFRs)—vague, context-dependent, and system-wide concerns like security and regulatory compliance. This study investigates the accuracy and quality of multi-turn dialogues between 49 programmers and GitHub Copilot, focusing on HIPAA-derived NFRs within the iTrust codebase.

Researchers established an expert-annotated ground truth for 148 NFRs across three dimensions: requirement satisfaction level, reasoning, and code localization. Participants engaged in multi-turn dialogues to assess these requirements, rated the system's responses, and completed satisfaction surveys based on the PARADISE framework.

The study reveals a concerning disconnect between user perception and objective performance. Participants expressed high levels of agreement with the LLM's assessments (91–94% across all dimensions), yet the system's actual accuracy was poor. Specifically, the F1 score for requirement satisfaction was 0.381, and code localization accuracy was notably low at 0.203.

When modeling user satisfaction, the researchers found that dialogue characteristics significantly influence the developer experience. Verbose responses and a high volume of information-providing turns were negatively correlated with user satisfaction. Conversely, proactive interactions—where the agent takes initiative in the dialogue—were positively associated with higher satisfaction scores.

These findings suggest that developers may be prone to over-relying on LLM-based assistants for complex tasks like NFR assessment, even when the system provides inaccurate information. Because users perceive the output as high-quality, they may fail to critically evaluate the reasoning or code locations provided by the AI. This highlights an urgent need for evaluation benchmarks that prioritize multi-turn reasoning and accuracy in NFR assessment rather than just single-turn functional output.