Skip to content

Four-stage windowed pipeline architecture for deck-level question generation

Design · All designs
Evidence · 1 claim (1 for) · 1 study (1 design), q1 · 0 of 1 report an effect size · 1 claim rests on one study

Description

The article's organizing architecture is a multi-pass generative pipeline: a Window Planner that "Analyzes overlapping windows of slides, creating a contact sheet image to infer titles, summaries, modalities, roles, and a provisional question budget"; a Deck Synthesis phase merging window analyses into one coherent deck plan; a Slide Annotator generating questions per slide using deck context; and a Reconciliation phase evaluating the full-deck annotation for redundancies and unbalanced coverage. The windowed design is intended to mitigate LLM context limits while preserving the presentation's narrative flow.

Design Implications

Context

Requirements

  • All outputs must conform to strongly typed JSON schemas (Pydantic models SlidePlan, SlideAnnotationModel, ReconciliationModel) to ensure data integrity

Constraints

  • The multi-pass architecture incurs substantial LLM inference latency and API costs on large decks

Target Learners

  • undergraduate and graduate students in technical lecture courses

Learning Goals

  • deck-level pedagogical scaffolding of comprehension questions across a full presentation

Claims

Examples

-

Key Sources

  • Salsman, J. (2026). Slide Deck Q&A Quality Assurance App: A Multi-Stage Pipeline for Pedagogical Question Generation. https://arxiv.org/abs/2605.26428. doi:10.48550/arxiv.2605.26428
Page metadata
Field Value
type design
id four-stage-windowed-question-pipeline
title Four-stage windowed pipeline architecture for deck-level question generation
description The article's organizing architecture is a multi-pass generative pipeline: a Window Planner that "Analyzes overlapping windows of slides, creating a contact sheet image to infer titles, summaries, modalities, roles, a...
status review
generated by process:settle-candidates on 2026-10-11
trust unverified — no one has confirmed this page's content
sources 1 — salsman-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.