Skip to content

Addressing Situated Teaching Needs: A Multi-Agent Framework for Automated Slide Adaptation

🕒 Published (v1): 2025-11-24 07:22 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper presents a four-agent hierarchical framework (Lesson Planner → Adaptation Organizer → Executor → Validator) that translates instructors' natural-language pedagogical requests into concrete PowerPoint modifications via MCP tool calls. Evaluated on 16 tasks across 8 real-world course slide decks, it achieves an operational F1 of 0.89 against expert-defined ground truth, outperforming two baselines on intent alignment, content coherence, and factual accuracy.

Problem

Existing AI slide tools (Manus, GLM, AutoPresent) are designed for one-shot generation rather than iterative, instructor-driven adaptation—they ignore pedagogical context, overwrite established visual styles, and lack mechanisms to translate high-level teaching intent into precise, atomic slide operations. No prior system addresses the teacher-centric situated-teaching-needs (audience level, curriculum standards, hot topics) that dominate real-world adaptation workflows.

Method

The framework \(\mathcal{F}\) takes three inputs—Situated Teaching Needs (STN), Original Slides \(S = \{s_1, \ldots, s_N\}\), and Instructor Materials (IM)—and produces an adapted presentation \(S' = \mathcal{A}(S, R, M)\) through two phases:

  1. Strategic Planning — The Lesson Planner (LP) classifies requests as Recompose (structural overhaul) or Refine (targeted edits). For Recompose, LP builds a full instructional design (objectives, student analysis, lesson flow), performs a gap analysis against existing slides, and emits a high-level adaptive guideline. For Refine, it directly parses the instruction into an actionable directive. The Adaptation Organizer (AO) then augments with live web retrieval (Tavily Search API) and produces a slide-by-slide JSON manifest of ADD/MODIFY/KEEP operations with granular execution steps.

  2. Execution and QA — The Executor (EX) calls a custom MCP tool suite (extract_slide_text, get_slide_images_info_for_llm, manage_text, manage_image) built atop python-pptx for in-situ modification (no slide regeneration). The Validator (VA) performs per-slide content and typographic checks against the manifest in an iterative feedback loop with EX until all discrepancies are resolved.

Key Contributions

  • Formative study (N=5 educators, semi-structured interviews) identifying four friction themes in slide adaptation and three derived design goals.
  • Hierarchical Plan-Act multi-agent architecture that decouples curriculum contextualization (LP/AO) from fine-grained execution (EX/VA).
  • Custom MCP tool layer wrapping python-pptx for atomic, in-situ slide modification, avoiding style-destroying regeneration.
  • Operation-level evaluation methodology using precision/recall/F1 against expert-built ground-truth operation sets.
  • LLM-as-judge pipeline (Gemini-2.5-Pro, \(T=0\)) scoring four dimensions: Intent Alignment, Content Coherence, Factual Accuracy, Visual Clarity.

Results

  • Operational F1: Ours 0.89 (P=0.90, R=0.88) vs. GLM 0.76 (P=0.73, R=0.83) vs. PPT-MCP 0.20 (P=0.31, R=0.18).
  • Intent Alignment (LLM-judged, 1–5): Ours 3.75 vs. GLM 3.56 vs. PPT-MCP 2.53.
  • Content Coherence: Ours 2.78 vs. GLM 2.19 vs. PPT-MCP 1.88.
  • Factual Accuracy: Ours 2.38 vs. PPT-MCP 2.13 vs. GLM 1.50.
  • Visual Clarity: Ours 3.16 ≈ PPT-MCP 3.31 ≈ GLM 3.22 (no significant gain).
  • Processing time: PPT-MCP 17 s < Ours 159 s < GLM 330 s; Ours splits as ~117 s planning + ~41 s execution+validation.
  • Evaluation corpus: 8 OER slide decks (10–25 slides each), 16 tasks (8 recompose, 8 refine).

Limitations

  • Factual accuracy remains low in absolute terms (2.38/5) despite leading baselines; hallucination from retrieval is a residual problem.
  • Visual clarity shows no improvement over baselines due to the in-situ modification approach compromising legibility when inserting content.
  • Small evaluation scale (16 tasks, 8 decks, N=5 formative interviews) limits generalizability.
  • LLM-as-judge evaluation (Gemini-2.5-Pro) for output quality introduces potential model-specific bias.
  • Operational ground truth was built for only 5 of 16 tasks by three raters, limiting scope of that evaluation.
  • No user study with real instructors actually deploying the system in live course preparation.

Relevance to Harnesses / Meta-Harnesses

This paper is a concrete instantiation of a domain-specific multi-agent harness: a fixed pipeline of specialized sub-agents (LP → AO → EX → VA) with structured inter-agent communication (JSON manifests), external tool integration (MCP tool layer, web retrieval API), and a built-in quality-assurance feedback loop. The separation of planning agents from execution agents—and the use of a validator to drive iterative correction—directly mirrors the Plan-Act-Verify pattern common in general-purpose agent harnesses, but applied narrowly to document editing. The custom MCP tool suite (atomic manage_text/manage_image operations) exemplifies how harness designers define restricted action spaces to keep execution agents grounded. For harness researchers, the operational F1 evaluation methodology—comparing executed operation sequences against expert ground truth using precision/recall—offers a transferable framework for measuring agent faithfulness to high-level intent.