DFT-GEN preserves 93% of exam info in dyslexia-friendly texts

Generic text simplification often strips away the exact details learners with dyslexia still need—terms, constraints, or source evidence. In response, Jiaqian Yu and colleagues from the human-computer interaction community created DFT-GEN, a stakeholder-informed transformation framework specifically built for content-heavy educational materials. Rather than relying on a generic LLM refinement loop, DFT-GEN layers a deterministic Dyslexia Accessibility Controller (DAC) on top of protected-span preservation. This layer turns stakeholder and expert preferences into reproducible rules for visual-unit length, chunk spacing, source/task separation, highlighting budget, and reviewable risk flags—so the final output is both readable and verifiable.

DFT-GEN was evaluated with two purpose-built diagnostics: DCFI for fidelity safety and B-DVAS-VL for rendered visual accessibility. On 2,280 bilingual exam-style items, DFT-GEN won 93% of English and 64% of Chinese pairwise visual-accessibility judgments over same-backbone controls. In a controlled pilot with dyslexic adult readers, participants maintained answerability on exam questions while reporting reduced reading effort. The authors also released code with the preprint, making the framework immediately usable for researchers and edtech developers who need to adapt dense learning materials without losing the substance that makes them useful.