acceptodds
Under review as a conference paper at ICLR 2027

Slides2MindMap: Reconstructing Cognitively Grounded Knowledge Hierarchies from Lecture Slides

Abstract

Generating mind maps from lecture slides can help learners efficiently assimilate fragmented knowledge, promising substantial benefits for intelligent tutoring systems. However, dedicated automatic evaluation methods remain underexplored, while the generation is also challenging, requiring a global-local knowledge focus balance and handling large-scale, heterogeneous content. We formulate the Slides2MindMap task, which aims to reconstruct cognitively efficient knowledge hierarchies from a course's slide deck collection. For systematic evaluation, we introduce S2M-Bench, a benchmark comprising 12,774 slide pages with expert-annotated mind maps spanning 24 university courses. S2M-Bench includes a cognitive-science-grounded evaluation framework that integrates comparison with expert-annotated references, structure conformity analysis, and VLM-as-a-Judge. To address this task, we propose AutoMindMap, an agentic framework inspired by the Structure Building Framework Theory. AutoMindMap comprises Skeleton Laying for global scaffold anchoring, Iterative Knowledge Integration augmented by context-aware and branch-isolated summarization, and Dual-Stage Refinement with a local-global decoupling mechanism. The framework reconciles local knowledge faithfulness with global coherence, and adapts to slide-specific features. Experiments on S2M-Bench demonstrate that AutoMindMap outperforms baselines and achieves superior robustness across different models and scenarios, underscoring its educational application value.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI for Education & Learning Technologies, Slides-based Datasets & Benchmarks, LLM-based Agentic Systems, Knowledge Mining

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AbRaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This Paper formulate the Slides2MindMap task, proposes S2M-Bench, a benchmark for evaluating automatic mind-map generation from lecture slides and AutoMindMap, an agentic framework for constructing mind-map from slides inspired by the Structure Building Framework. The authors evaluate the generated mind-map using ground-truth based metrics, structure conformity and VLM-Judge.

Critical strengths
  1. The proposed benchmark is large and multi-dimensional, which could serve to evaluate automatic methods in this subdomain.
  2. The pipeline of AutoMindMap is also well designed, with multiple stages for skeleton construction, iterative knowledge, and refinement.
Critical weaknesses
  1. The evaluation mainly focuses on the generated map itself or its similarity to the human reference but does not directly evaluate whether the generated map is faithfully grounded in the original slides. There might therefore be a hallucination problem. Based on the prompt of VLM-Judge, only the generated map, excluding the original slides, is given to the VLM for evaluation. Therefore, the Content Accuracy score seems to mainly evaluate map itself, rather than the faithfulness between the generated map and the original slides.

  2. This task is subjective, and many possible taxonomies may exist. As the authors show in Table 18, the EC-F1 of human re-annotation is only 54.73. In addition, only limited details about the ground-truth annotation process are provided in the paper. Therefore, I am skeptical about the reliability of the GT-based metrics for this open-ended task.

Questions for authors
  1. More details about the annotation process need to be clarified. For example, how many annotators were involved for each course? Were the annotations constructed independently? How were the annotation instructions and rubrics provided, and how were disagreements resolved?

  2. Is there any evaluation that directly measures whether the generated concepts and relations are supported by the original slides?

Reviewer_CctVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents the Slides2MindMap task, aiming to reconstruct the cognitive organizational concept hierarchy at the course level from a set of heterogeneous course lecture slides. The article introduces the S2M-Bench benchmark dataset, which contains 12,774 pages of slides from 24 university courses and expertly annotated mind maps, and combines evaluation frameworks such as reference-based metrics, structural checks, and VLM as judges. The authors also propose AutoMindMap, a proxy workflow based on large language models (LLMs): first, plan the global framework, gradually integrate the slide content through specific summary and editing tools for branches, and then perform local and global optimizations separately. In the benchmark test, AutoMindMap outperforms the evaluated automated baseline methods in most reported metrics, while the manually annotated results still perform better.

Critical strengths

1.This paper presents a very meaningful scenario. It aims to reconstruct the thinking process from the slides into a mindmap, in order to assist learners in comprehending the entire course.

2.S2M-Bench is a substantial curation effort. It contains 267 decks and 12,774 pages across 24 courses, includes text-, image-, and hybrid-heavy materials, and provides expert-annotated course-level maps.

3.The evaluation is broader than exact match to one reference map. It combines hierarchy-aware concept and relation metrics, structural validity measures, and rubric-based VLM judgments. The authors also report human-model agreement, cross-model judge results, stability tests, failure cases, and cost analysis.

Critical weaknesses

1.This paper conducts an evaluation of the quality of mind maps. However, an even more intriguing point is: Just how much help do these mind maps provide for students' learning? For example, can combining an existing mind map with LLM lead to better review strategies or the discovery of knowledge relationships than vanilla LLM?

2.The benefit of using a VLM for evaluation is not sufficiently established. The mind maps are represented as structured concept graphs, which could be serialized into a semi-structured format such as YAML and evaluated by an LLM against the same rubric. The current setup does not show whether visual input provides information that cannot be recovered from the structured representation, or whether the VLM scores are mainly driven by the rendered layout. The authors should compare image-based VLM judging with text-only LLM judging on the same examples and rubric. This would clarify whether visual evaluation is necessary for assessing the knowledge structure, and what additional value it provides for judging visual organization.

Questions for authors

See Weakness.

Reviewer_AZ8wOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the problem of converting course presentation slides to highly structured mind maps, the Slides2MindMap task formulated in this paper. The task is challenging because course slides are usually heterogeneous and unstructured; the necessary information is spread across the slides, making it hard for the model to produce coherent mind maps. To this end, the paper proposes S2M-Bench, a benchmark comprising 12,774 slide pages with 24 expert-annotated mind maps, and further proposes AutoMindMap, an agentic framework leveraging LLM and pre-defined tools (increasing/decreasing nodes, etc.) to build the mind maps. The paper demonstrates results with a series of metrics including comparison with expert maps in different dimensions and LLM evaluations.

Critical strengths
  • Converting course presentation slides to a highly structured mind map is a practical and useful application for both teachers and students, and can facilitate many education scenarios.
  • The proposed evaluation metrics are thorough, covering many aspects including comparison with expert maps and LLM evaluation of the constructed mind map quality.
Critical weaknesses
  • The paper appears to lack technique innovation. The main method builds on prompting LLM to build and augment the mind map, and the applied techniques are mainly prompt engineering. Besides the proposed benchmark, the paper is mainly an engineering work.
  • The paper does not compare with some existing work. For example, BoomMapper seems to be a very similar work to convert powerpoint presentations into mind maps. The paper claims this work is the first to Slides2MindMap task, which could be overstated.
  • The paper should consider include more open-source models and strong private models. Currently it only uses GPT 5.4-mini and GPT 4.1-mini, and I am curious if latest model like GPT 6 can directly resolve this problem.
  • (Minor) The proposed benchmark involves expert mind maps. While expert mind maps are valuable, these are hard to be extended, and 24 mind maps seem to be quite limited. if trained by a model or even used as a validation dataset, the model could quickly overfit to that.
Questions for authors

I am mainly worried about the technical innovation of this work and the comparison with existing work, especially BoomMapper[1]. This work's contribution seems highly overlap with what BoomMapper trying to bring. Please consider focusing on these two aspects in rebuttal to my review.

[1] BoomMapper: An Innovative Web-Based System for Transforming PowerPoint Presentations into Mind Maps

2 comments

I thank the authors for their replies, which address many of my concerns. Specifically, I appreciate the more detailed discussion of related work, and the responses highlight the unique advantages of the proposed AutoMindMap. In addition, I appreciate the comparison with the latest model; the limitations of SOTA models on this benchmark add value to this work. I remain somewhat reserved about the novelty of the proposed modules. Although the authors highlight five modules in the rebuttal, some appear incremental to me. For example, step-wise association and repair involves similarity-based retrieval and prompting an LLM to check structural validity. However, the contributions in scale and evaluation are clear to me. I will increase my rating.

Dear Reviewer AZ8w,

Thank you for your thoughtful follow-up and for recognizing the contributions of our work. We are glad that our response helped address many of your concerns. We appreciate your constructive feedback and will incorporate key additional experimental results into the revised manuscript.

Best regards,

Authors of Paper 57046

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.