Language technologies increasingly mediate access to scientific information, yet their impact depends on whether they improve real reading experiences for diverse audiences. We study large language model (LLM)-generated scientific summaries as an accessibility intervention for neurodivergent adults, focusing on readers with ADHD and/or autism. Using a mixed-methods design, we first conducted a formative qualitative study with 10 neurodivergent adults to identify design needs. We then ran a quantitative experiment with 200 neurodivergent adults comparing five summary styles across lower- and higher-complexity scientific texts, measuring comprehension, trust, engagement, and cognitive load. Results show that writing-style effects depend on text complexity: structured LLM-generated summaries improved comprehension for higher-complexity articles but reduced comprehension for lower-complexity articles relative to a baseline human-generated extractive summary. Trust increased for simpler LLM summary styles, while engagement and cognitive load showed fewer robust writing-style effects. Our findings suggest that accessible NLP systems should adapt explanatory support to reader needs and text difficulty, rather than offering a single “best” summary style.

Online charts are ubiquitous, yet remain systematically inaccessible to screen reader users. This paper presents the process for developing the MATCHA Chart Summarizer (MCS), an interactive, browser-based system, leveraging multimodal LLM’s to generate accessible spoken chart summaries and support natural-language querying. MCS aims to enhance functional comprehension and narrow the reasoning gap for screen reader users when interacting with data visualizations. Through a three-stage mixed-methods, user-based evaluation (N = 126 total), we conduct a needs assessment, design our prototype, and demonstrate that LLM-generated summaries can improve task accuracy and increase user confidence, informativeness, and usability. We also quantify how visualization flaws, such as missing labels or poor contrast, induce semantic drift in generated chart summaries. We contribute an accessible assistive NLP system, an impact-oriented evaluation framework, and evidence-based design recommendations for automated chart summarization.