Skip to content

SlideBot: A Multi-Agent Framework for Generating Informative, Reliable, Multi-Modal Presentations

🕒 Published (v1): 2025-11-12 23:12 UTC · Source: Arxiv · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SlideBot is a three-stage, multi-agent pipeline that generates university-level slide decks from a topic string by coordinating a Retriever, Code Generator, and Enhancer under a central Moderator. It grounds content via RAG (arXiv API or BM-25 over textbooks), encodes instructional design principles from Cognitive Load Theory (CLT) and the Cognitive Theory of Multimedia Learning (CTML), and outputs LaTeX Beamer slides with embedded instructor comments. Human evaluations in CS and biomedical domains show consistent improvements over Microsoft Copilot across informativeness, reliability, and practicality.

Problem

Existing LLM-based slide generators rely on parametric knowledge, producing hallucinated or outdated content and ignoring pedagogical structure. Single-prompt approaches fail at simultaneous multimodal generation (text, layout, visuals), yielding incoherent or broken outputs. There is no prior system that integrates retrieval, structured instructional planning, code generation, and iterative instructor feedback into a unified pipeline for educational slide creation.

Method

SlideBot decomposes generation into three sequential stages orchestrated by a central Moderator agent:

  1. Content Retrieval: A Retriever agent queries external corpora—arXiv API (keyword matching) for academic literature or BM-25 over textbooks—and returns summaries with citation metadata. The Moderator filters these summaries before passing them downstream.

  2. Slide Draft Generation: The Moderator constructs a structured slide plan using a predefined guide embedding CLT/CTML principles (signaling via bolded terms and bullet hierarchies, coherence via per-slide conciseness, spatial contiguity via text-visual pairing). The plan is handed to a Code Generator agent, which translates it into LaTeX Beamer code. A validation loop sends compilation errors back to the Code Generator with the error message and revision suggestions until the code compiles cleanly.

  3. Presentation Enhancement: An Enhancer agent inserts figures via prewritten parametric macros and adds instructor-view-only comment slides (elaboration prompts, misconception warnings). The Moderator then returns the finished deck to the user, who can issue revision requests that re-enter the loop.

All inter-agent communication routes through the Moderator, which enforces quality gates at each handoff.

Key Contributions

  • Modular multi-agent harness with a central Moderator coordinating four specialized agents (Retriever, Moderator-planner, Code Generator, Enhancer) across three pipeline stages.
  • Integration of CLT and CTML instructional design principles into the slide plan template, operationalized as structural constraints rather than post-hoc prompts.
  • Iterative code compilation/repair loop that minimizes LaTeX hallucinations and ensures executable output.
  • Dual-survey evaluation protocol pairing student (surface quality) and domain-expert (accuracy, instructor utility) perspectives across two domains (CS and biomedical).
  • Instructor-facing interactivity: revision requests re-enter the Moderator loop; instructor comments are embedded as speaker-view annotations.

Results

All scores are on a 1–5 scale, SlideBot vs. Copilot (lower variability is better):

  • Explanation Style (student): 3.96 vs. 2.24 (+1.71)
  • Conceptual Accuracy (expert): 3.57 vs. 2.71 (+0.86)
  • Topic Coverage (expert): 4.10 vs. 2.67 (+1.43)
  • Credibility (student): 4.36 vs. 1.93 (+2.42)
  • Variability in suitability (student): 0.33 vs. 1.47 (−1.14; SlideBot more consistent)
  • Structure & Flow (student): 4.04 vs. 3.42 (+0.62)
  • Overall Suitability (student): 3.67 vs. 2.09 (+1.58)
  • Instructor Utility (expert): 3.76 vs. 2.14 (+1.62)

Baselines: Microsoft Copilot (integrated PowerPoint AI) and GPT-4o-mini Direct Prompting. CS study used GPT-4o-mini; biomedical study also tested model-size scaling (results truncated in the provided text).

Limitations

  • Very small evaluation samples: 15 CS students, 13 biomedical students, 7 CS experts, 4 biomedical experts—insufficient for statistical significance claims.
  • Subjective Likert-scale surveys; no automated or reproducible benchmark.
  • Only two baselines tested (Copilot, Direct Prompt); no comparison to other multi-agent or RAG-based slide systems.
  • LaTeX/Beamer output limits adoption in environments that require PowerPoint or Google Slides.
  • arXiv keyword-matching retrieval is shallow; semantic retrieval or vector search not employed.
  • Iterative instructor feedback loop is described but not systematically evaluated (no ablation of feedback rounds).
  • Biomedical results and model-scaling results are truncated in the available text—full quantitative picture is incomplete.

Relevance to Harnesses / Meta-Harnesses

SlideBot is a textbook instance of a domain-specific agentic harness: a Moderator agent acts as an orchestrator that sequences specialized sub-agents, enforces handoff contracts (summary → plan → code → compiled artifact), and runs a repair sub-loop when compilation fails—exactly the pattern seen in code-generation and research meta-harnesses. The modular corpus interface (swap arXiv for BM-25 textbooks) and the extensible agent library (drop in new figure macros or retrieval sources) mirror the plugin/adapter patterns used in general-purpose harness frameworks. The code-validation feedback loop—where the Moderator re-invokes the Code Generator with the error trace—is a concrete instantiation of the self-repair loops studied in harness research. For someone tracking harness design, this paper offers a clean case study of how orchestration logic (Moderator), quality gates (compilation check), and iterative refinement combine in a production-adjacent system.