🤖 AI

Slides2MindMap: Reconstructing Cognitively Efficient Knowledge Hierarchies from Lecture Slides

This paper introduces the Slides2MindMap task and the S2M-Bench benchmark to address the challenge of automatically reconstructing cognitively efficient knowledge hierarchies from lecture slides, proposing the AutoMindMap framework which leverages a structure-building approach to achieve superior global coherence and local faithfulness compared to existing baselines.

Original authors: Yuzhi Wang, Rongjun Ye, Shengyuan Chen, Huachi Zhou, Jiaqi Bai, Chuang Zhou, Zhicong Hong, Xiao Huang

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuzhi Wang, Rongjun Ye, Shengyuan Chen, Huachi Zhou, Jiaqi Bai, Chuang Zhou, Zhicong Hong, Xiao Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to learn a new subject, like how a car engine works or the history of ancient Rome. You have a stack of lecture slides in front of you. These slides are like a chaotic treasure hunt: they jump from big ideas to tiny details, they repeat themselves, and sometimes they show a picture without explaining what it means. Your brain has to work overtime to sort this mess into a clear story. This is where mind maps come in. Think of a mind map as a super-organized family tree for ideas. Instead of a long, boring list, it puts the main topic in the center and branches out into smaller, connected ideas, just like how your brain naturally connects thoughts. For a long time, making these maps from messy slides has been a job for humans, taking hours of careful thinking. But now, scientists are asking: Can a computer do this for us? Can an AI look at a messy pile of slides and instantly build a perfect, easy-to-understand map that helps us learn faster? This is the big question researchers are tackling in the field of Educational AI, specifically focusing on how machines can understand and organize human knowledge.

Enter the team from The Hong Kong Polytechnic University and Shanghai Jiaotong University, who have built a new tool called AutoMindMap. They realized that simply asking a computer to "summarize these slides" doesn't work well. It's like asking a chef to make a soup by just throwing all the ingredients into a pot without a recipe; you might get something edible, but it won't taste right. The computer needs to understand the story of the lecture, not just the words. To test their idea, they created a giant playground called S2M-Bench. This isn't just a few slides; it's a massive collection of 12,774 slide pages from 24 different university courses, ranging from math to computer science. For every single course, human experts drew the "perfect" mind map by hand, giving the researchers a gold standard to measure against.

The researchers found that old methods were like trying to build a house by stacking bricks randomly. Some methods just grabbed the titles of the slides, missing the deep connections. Others tried to merge small chunks of text, but the final result was a tangled mess with no clear structure. Their new tool, AutoMindMap, works differently. It acts like a smart architect who follows a three-step plan based on how humans actually learn. First, it does Skeleton Laying: before building anything, it looks at the whole course to figure out the main pillars and the order of topics, creating a sturdy frame. Next, it performs Iterative Knowledge Integration: it reads the slides section by section, like a student taking notes, constantly updating its memory to connect new ideas to old ones without getting confused. Finally, it uses Dual-Stage Refinement: it acts as both a micro-manager and a big-picture critic. It fixes small errors in the details (like a typo in a node name) and then steps back to look at the whole map to ensure the branches are balanced and not too crowded.

The results? The new tool is a game-changer. When tested against the human-made "gold standard" maps, AutoMindMap built structures that were far more logical and easier to understand than any other automated method they tried. It successfully handled the messy, jumping nature of slides, turning them into clear, branching trees of knowledge. While it didn't quite reach the perfection of a human expert (who can still spot the subtlest nuances), it came remarkably close, proving that an AI can indeed learn to organize knowledge in a way that feels natural to the human mind. The researchers suggest this could be a huge help for students everywhere, turning hours of confusing lecture notes into a single, clear picture of what they need to learn. They have even made their code and data available for others to try, hoping to spark a new wave of smarter, more helpful educational tools.

Technical Summary: Slides2MindMap

Problem Definition

The paper addresses the Slides2MindMap task: reconstructing cognitively efficient knowledge hierarchies from large-scale, fragmented lecture slide collections. Unlike well-structured textbooks, lecture slides often lack explicit hierarchical organization, featuring flexible, pedagogy-driven narratives with interleaving, repetition, and visual noise. Existing methods face two primary challenges:

Paradigm Gap: Traditional knowledge representation frameworks (e.g., Knowledge Graphs) focus on dense triple extraction and entity anchoring for retrieval, lacking the concise, progressive, and branched structure required for human visual understanding and cognitive assimilation.
Slide-Specific Heterogeneity: Slide decks are large-scale (often exceeding 100K tokens per course) and structurally disordered. Direct extraction or merging methods struggle to maintain global logical coherence while preserving local knowledge faithfulness, often leading to cognitive overload or structural collapse.

The task is defined as generating a connected directed acyclic graph (DAG) $M = (V, E, r)$ where $r$ is the course root, $V$ represents concept nodes, and $E$ represents hierarchical edges, such that the map faithfully generalizes the intrinsic knowledge hierarchy of the course while presenting a coherent global structure.

Methodology: AutoMindMap

To address these challenges, the authors propose AutoMindMap, an LLM-based agentic framework inspired by the Structure Building Framework (SBF) theory from cognitive psychology. The framework treats mind map construction as a dynamic discourse comprehension process involving three distinct stages:

1. Skeleton Laying (Foundation)

Mirroring the "foundation laying" phase of SBF, this stage establishes a global scaffold before detailed construction.

Process: An LLM planner extracts salient elements (e.g., file metadata, titles) from the slide decks to generate:
- Backbone Nodes ( $V_s$ ): A set of second-level nodes directly connecting to the root, representing the top-level abstract concepts.
- Mapping ( $B$ ): A bipartite graph linking backbone nodes to relevant slide decks.
- Expansion Guidelines ( $Z$ ): Skill-style principles for downstream branching, outlining anticipated structures and risks.
- Prerequisite Sequence ( $S$ ): An ordering for processing decks.
Goal: To anchor the global structure logic, preventing structural shifts and loss of architectural sight during subsequent expansion.

2. Iterative Knowledge Integration (Mapping & Shifting)

This stage emulates the "mapping" (intra-branch) and "shifting" (cross-branch) processes, incrementally assimilating concepts.

Discourse Segmentation: Slides are partitioned into contiguous discourse units based on semantic continuity rather than format cues.
Context-Aware Summarization: For each discourse unit, an LLM extracts local concept relations ( $N^k_j$ ) while updating a dynamic Memory State ( $U_t$ ). This state acts as a semantic bridge, retaining the essence of prior content to reconstruct knowledge logic continuity across fragmented slides.
Controlled Integration: The local subgraph is integrated into the global map ( $M_t$ ) using a partial view ( $M'_{t-1}$ ) to constrain context to relevant branches, avoiding noise.
Cross-Branch Associating & Repair: The system retrieves candidate nodes from other branches to identify potential cross-links and performs structural local repairs to prevent loops or broken hierarchies.
Tool-Based Execution: All updates are executed via reasoning-enhanced atomic tool calls (e.g., add_node, delete_edge) to ensure verifiable and controllable graph editing, mitigating hallucination.

3. Dual-Stage Refinement

To resolve residual structural imperfections without causing mutual destruction between local and global edits, a decoupled refinement process is employed:

Micro-Level Refinement: Independent branch-level refinement using an Editor-Critic loop. A critic agent identifies local anomalies (e.g., overgrown branches, guideline violations), and an editor agent executes structural edits via tools.
Macro-Level Refinement: A global critic evaluates the rendered mind map image ( $I$ ) alongside the textual list to assess global coherence, branch balance, and cognitive load. This visual feedback allows the system to detect structural issues invisible in text alone (e.g., spatial imbalance).

Key Contributions

S2M-Bench: The first dedicated benchmark for evaluating automated knowledge hierarchy reconstruction from lecture slides. It comprises 12,774 slide pages across 24 university courses with expert-annotated ground-truth mind maps.
Evaluation Framework: A three-faceted evaluation system grounded in cognitive science:
- Ground-Truth Comparison: Metrics including Hierarchy-aware Node F1-Score (HN-F1), Edge Connectivity F1-Score (EC-F1), and Matched-Edge Connectivity F1-Score (MEC-F1).
- Quantitative Structure Conformity: Metrics like Pass@1 (structural validity) and Structural Load Factor (SLF) to measure cognitive load.
- VLM-as-a-Judge: A vision-language model evaluator scoring Content Accuracy, Structure Quality, and Pedagogical Efficiency based on rendered images and text.
AutoMindMap Framework: An agentic system that successfully balances local knowledge faithfulness with global logical coherence by mimicking human discourse comprehension (SBF theory), outperforming existing baselines.

Experimental Results

Experiments on S2M-Bench demonstrate that AutoMindMap significantly outperforms baselines, including direct LLM generation, chunk-merge strategies, and document hierarchy indexing methods (BookRAG, PageIndex).

Performance: AutoMindMap achieves superior scores across all metrics. With GPT-5.4-mini as the VLM, it reaches an average VLM-Judge score of 3.72 (vs. 3.01 for the best baseline) and a 100% Pass@1 rate, indicating perfect structural validity.
Robustness: The framework shows strong generalization across different LLM backbones (DeepSeek-V4-Flash, GPT-5.4-mini) and diverse course categories (Math, Concept Understanding, Synthetic Application).
Ablation Studies: Removing any core component (Skeleton Laying, Iterative Integration, Dual-Stage Refinement, or Visual Feedback) leads to performance degradation, confirming the necessity of the global-local balance and the decoupled refinement mechanism.
Comparison to Humans: While AutoMindMap narrows the gap significantly, human-annotated maps still achieve the highest scores, highlighting the inherent difficulty and open-ended nature of the task.

Significance and Claims

The paper claims that Slides2MindMap represents a critical step toward intelligent education by automating the creation of cognitively efficient learning scaffolds.

Pedagogical Value: By reconstructing fragmented slide content into progressive, hierarchical structures, the system facilitates the assimilation of knowledge, aligning with Advance Organizer and Cognitive Load theories.
Technical Advancement: The work demonstrates that treating knowledge extraction as a dynamic, cognitive process (via SBF) rather than a static extraction task yields superior results in handling the heterogeneity of real-world educational materials.
Benchmarking: S2M-Bench provides a necessary standard for evaluating not just accuracy, but the cognitive efficiency and structural soundness of generated knowledge representations, moving beyond simple retrieval metrics.

The authors note limitations, including the current focus on concept structure (excluding visual design generation), dependence on document parsing tools (MinerU), and the high computational cost of the iterative agentic process. They position the work as a foundation for future research in Educational AI and knowledge structure visualization.