Get started
Work

From Fluent to Useful: Generative AI That Models Purpose, Audience, and Presenter for Scientific Communication

Year: 2026
Type: conference-paper
Abstract: Modern generative AI produces fluent text, polished slides, and clean diagrams -yet still fails when an artifact must serve a specific purpose for a specific reader, used by a specific presenter.The missing piece is not fluency but a model of why content is being produced, for whom (presenter and audience alike), and how it should adapt as goals shift.My completed and published work develops five systems across the scientific communication pipeline: ADAPTIVE IE for intentdriven extraction; Persona-Aware Slide Generation for audience reframing rather than blanket simplification; GPA for reconciling divergent group preferences; SciDoc2Diagrammer-MAF, whose multi-aspect critics distinguish purposeful abstraction from genuine omission or hallucination; and SMART-Editor, which models cascading edits across multimodal layouts.Together they show that aligning with intent, audience, and structure is necessary -but cannot answer whether the resulting artifacts actually communicate.I therefore propose three directions in priority order: (RQ1) a goal-driven framework that measures the educational utility of document-to-video generation through IRT-calibrated diagnostic questions, validated against measured learning outcomes and accompanied by inter-annotator agreement studies on human effectiveness judgments; (RQ2) presenter-side personalization that treats the presenter -not just the audience -as a firstclass user; and (RQ3) a unified SuperPersonalization benchmark for transferable user preferences.RQ3 is scoped to be deferrable to post-dissertation work if RQ1 expands.The thesis shifts the target from generative AI that produces content that looks correct to systems whose outputs demonstrably communicate.
DOI: 10.18653/v1/2026.acl-srw.87
Language: English
Open Access status: hybrid
OpenAlex ID: W7166814100
FWCI:
Cites:
Cited by:
Related to: 10