Pith. sign in

REVIEW 4 major objections 5 minor 60 references

The paper claims that a three-stage agentic pipeline, AutoMindMap, reconstructs course-level knowledge hierarchies from raw lecture slides and outperforms all automated baselines on its own benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 00:32 UTC pith:WO6CBMCL

load-bearing objection The benchmark is the real deliverable; the cognitive-efficiency claim outruns the evidence, but the method is credible and worth refereeing. the 4 major comments →

arxiv 2608.00610 v1 pith:WO6CBMCL submitted 2026-08-01 cs.AI

Slides2MindMap: Reconstructing Cognitively Efficient Knowledge Hierarchies from Lecture Slides

Yuzhi Wang , Rongjun Ye , Shengyuan Chen , Huachi Zhou , Jiaqi Bai , Chuang Zhou , Zhicong Hong , Xiao Huang This is my paper
classification cs.AI
keywords mind map generationlecture slidesknowledge hierarchy reconstructioncognitive loadLLM agentStructure Building Frameworkeducational AIVLM-as-a-Judge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Lecture slides fragment and interleave knowledge, so mind maps built to teach from them need both faithful local content and a coherent global hierarchy. The paper tries to establish that this hierarchy construction should be modeled as a discourse-comprehension process, not a text-extraction task, and that its three-stage agentic framework AutoMindMap does it better than any existing automated method. It also establishes a benchmark, S2M-Bench, with 12,774 slides and expert-annotated mind maps for 24 university courses, plus a three-part evaluation combining ground-truth alignment, structural conformity, and VLM judging. On the benchmark, AutoMindMap is the only automated method that produces structurally valid maps on every run while scoring higher than every baseline on every metric, though it still trails human experts. The significance, if true, is a deployable first-draft mind-map tool and a standard way to measure progress in this new task.

Core claim

The paper defines Slides2MindMap: given a course's slide decks, generate a mind map—a connected directed acyclic graph with a single root—that faithfully reflects the intrinsic knowledge hierarchy and presents it in a cognitively efficient way. To evaluate it, the authors construct S2M-Bench, 24 university courses spanning 267 decks and 12,774 pages with expert-annotated maps averaging 121 concepts per course, and a three-facet evaluation: ground-truth alignment (hierarchy-weighted node F1, edge-connectivity F1, and matched-edge connectivity F1), structural conformity (Pass@1 validity plus a Structural Load Factor penalizing shallow/deep leaves and high fan-out, grounded in the 7±2 working-m

What carries the argument

The central mechanism is the Structure Building Framework (SBF), a cognitive-psychology account of discourse comprehension in which a reader lays a foundation, maps coherent new information onto the emerging structure, and shifts to a new substructure when coherence breaks. AutoMindMap converts these three processes into: (1) Skeleton Laying—an LLM planner locks a level-2 backbone, a deck-to-branch bipartite assignment, expansion guidelines, and a prerequisite order, anchoring global structure before any detail is added; (2) Iterative Knowledge Integration—a dynamic reading-memory state bridges fragmented discourse units, a masked partial view keeps the model focused on the relevant branch,

Load-bearing premise

The load-bearing premise is that expert-annotated reference maps, the 7±2-based structural-load cutoffs, and a vision-language-model judge are valid proxies for what makes a hierarchy cognitively efficient; Appendix B concedes that no real classroom or learner-feedback experiments were run, so the pedagogical-value claim is only as strong as those proxies.

What would settle it

Give students the same course content as either an AutoMindMap map, a Direct-method map, or a human-expert map, then measure comprehension and self-reported cognitive load: if learner rankings deviate from the benchmark's ranking, the cognitive-efficiency claim fails even if the benchmark ordering holds. A cheaper check is to recompute method rankings with the SLF depth cutoff changed from 6 to 8 or the fan-out cutoff from 7 to 10 and see whether the ordering flips.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Students and instructors get a practical first-draft mind-map generator for an entire course's slide collection: the method produces structurally valid maps on every run and leaves a smaller gap to expert quality than any tested baseline.
  • Retrieval-oriented document-hierarchy methods such as BookRAG and PageIndex are shown to be the wrong tool for this task—they yield valid trees but the worst knowledge-faithfulness and pedagogy scores, supporting the paper's claim that slide hierarchies need reasoning-based reconstruction, not formatting-cue extraction.
  • The benchmark provides a reusable, multi-faceted yardstick—ground-truth alignment, structural conformity, and VLM judging—so future mind-map-generation methods can be ranked on a common scale rather than on ad hoc accuracy metrics.
  • The framework's advantage persists across different backbone LLMs and across course categories, and it is largest on concept-heavy courses where direct generation collapses most visibly, suggesting the design addresses exactly the failure mode it targets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The masked partial-view plus memory-state design implies that long-context hierarchy construction is inherently a stateful sequential process; this design likely transfers to other long-document structuring tasks such as textbook outlining or research-survey organization, though the paper only evaluates lecture slides.
  • The Structural Load Factor's cutoffs (leaf depth 2–6, fan-out ≤7) come directly from Miller's 7±2 heuristic; if interactive mind-map tools relieve working-memory load through zooming and folding, the cutoffs may be miscalibrated, and a learner-behavior study could recalibrate them without changing the framework.
  • The macro-refinement stage makes a testable auxiliary claim: rendered images carry structural information that text relation lists lose. Feeding the same critic graph-level imbalance statistics instead of the image would isolate whether the visual channel is truly load-bearing.
  • Because the paper's Appendix B reports that no real educational scenarios were run, an immediate extension is a classroom experiment comparing learning outcomes from AutoMindMap maps versus baseline and expert maps; such a study would test the benchmark's central proxy directly.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Slides2MindMap, a task of reconstructing a single mind map from an entire lecture-slide collection, along with S2M-Bench (12,774 slide pages, 24 courses, expert-annotated maps) and AutoMindMap, a training-free, LLM-agent pipeline with three stages: Skeleton Laying, Iterative Knowledge Integration, and Dual-Stage Refinement. The evaluation combines GT-based metrics (HN-F1, EC-F1, MEC-F1), a structural-load indicator (SLF), and a VLM-as-a-Judge with three Likert dimensions (CA, SQ, PE). The headline experiments report that AutoMindMap achieves Pass@1 = 100%, the lowest SLF among automated methods, and the best scores on all other metrics, with ablations and cross-model/cross-category robustness analyses in the appendix. The paper is explicit in Appendix B that no real educational-scenario experiments or learner feedback were collected.

Significance. If accepted at face value, the contribution is substantial: S2M-Bench is the first benchmark of its kind for slide-to-mindmap reconstruction, with non-trivial scale and expert annotation effort; AutoMindMap is a coherent, well-motivated system whose design is grounded in the Structure Building Framework; and the appendix includes valuable checks—human re-annotation, a Qwen3.5-Plus re-judging, human–model kappa, stability runs, and full prompts/tool specifications. The ranking claims are supported by converging evidence from multiple metrics rather than a single number. However, the cognitive-efficiency interpretation rests on unvalidated proxies, so the benchmark's pedagogical claim is stronger than the evidence. The absence of learner-outcome validation and the use of a same-family VLM for refinement and evaluation are the main risks; neither appears to be a fatal internal error, but both need to be confronted explicitly.

major comments (4)
  1. [§Evaluation Metrics (SLF, VLM-as-a-Judge) and Appendix B] The paper's central claim is that AutoMindMap produces 'cognitively efficient' hierarchies, but no learner-based validation is provided. SLF uses hand-set cutoffs (depth >7, fan-out >7, Eq. in 'Structural Load Factor'), and the CA/SQ/PE dimensions are scored by a VLM using rubric heuristics. Appendix B explicitly states that 'we don’t conduct experiments in real educational scenarios to collect human feedback.' As written, the evaluation establishes that AutoMindMap is best under these author-defined proxies; it does not establish that the maps reduce cognitive load or improve comprehension. I recommend either adding a small calibration study (e.g., human recall/complexity judgments on a subset) or, if that is out of scope, narrowing the title/abstract and task definition claims to 'proxy-based structural and semantic quality' and clearly labeling the cognitive-efficiency claim as a hypo
  2. [§VLM-as-a-Judge, Table 2, Appendix G.1–G.3] The main VLM-Judge scores come from GPT-5.4, while AutoMindMap's best variant uses GPT-5.4-mini for visual refinement. This creates a same-family preference confound for the magnitude of the CA/SQ/PE differences, even though Appendix G.2's Qwen3.5-Plus re-evaluation preserves the relative ranking. The paper should present the Qwen-based scores as a co-primary analysis or, at minimum, quantify the judge-model interaction. Additionally, Table 17 reports Pearson correlations up to 0.896 among CA, SQ, and PE; the claim that these are three distinct cognitive-efficiency dimensions is weakened. I suggest reporting a composite score and discussing redundancy, or providing discriminative validity evidence beyond kappa with humans.
  3. [Table 2 and Appendix F.4] The headline comparison reports a single run with no variance estimates over the 24-course test set. The stability analysis in Table 11 covers only 3 courses with temperature 0.5, which is not the setting of the main tables. Given that some margins are modest (e.g., EC-F1: 22.96 vs. 20.67; SLF: 1.82 vs. 5.26 is large, but head-to-head differences among variants in Table 12 are small), the paper should report means and standard deviations (or at least per-course score ranges) for the main comparison, or justify why the 24-course aggregate is stable enough for the ranking claim.
  4. [Appendix G.4, Table 18, and §Ground-Truth-based Comparison] The GT metrics are central to the claim that AutoMindMap is more faithful than baselines, but their ceiling is low and sensitive to annotator variation: human re-annotations against the original GT reach only HN-F1 61.64, EC-F1 54.73, MEC-F1 83.67 on the sampled courses. The matching procedure (Appendix E.1) also relies on LLM judgment with thresholds (τ=0.6, top-k=3) whose sensitivity is not reported. I am not claiming circularity—the matching is external to AutoMindMap—but the paper should provide a sensitivity analysis of the GT metrics to the matching thresholds and report inter-annotator variance for the re-annotation sample, otherwise the absolute GT numbers are hard to interpret.
minor comments (5)
  1. [Appendix G.1] The human–model and human–human kappa values in Table 15 are computed from a small sample (48 generated results for 8 courses). Report the 95% confidence intervals, and state how many human annotators scored each map.
  2. [Figure 4] The ablation figure's y-axis metric is not clearly labeled, and it would help to reference Table 12 directly in the main text for the 'slightly underperforms on MEC-F1 and SLF' sentence.
  3. [§Experiment Results] The phrase 'across every metric' in the main-results paragraph is too strong when the comparison is to Human; Human is excluded by the sentence 'except for Human', but the sentence could still be read as claiming strict Pareto dominance over all baselines on every row of Table 2. Clarify that the claim excludes the Human row and that SLF is lower-is-better.
  4. [Appendix F.3] Cost comparisons in Table 10 are single-run estimates with no variance; since cost is an explicit trade-off, report the range across courses.
  5. [References] PageIndex is cited via a blog URL (Zhang, Tang, and Team 2025). A peer-reviewed or archival reference would be preferable, or the paper should state clearly that the baseline reimplements the blog description.

Circularity Check

0 steps flagged

No significant circularity; AutoMindMap's evaluation is grounded in external checks, not in its own definitions.

full rationale

The claimed derivation chain is empirical rather than analytic. The task definition (Def. 2) does not fix the method's outputs; AutoMindMap's skeleton/integration/refinement stages are not defined in terms of the evaluation metrics. GT-based metrics (Appendix E.2) are fixed formulas over matched node sets, and the matching algorithm (Appendix E.1) is a uniform preprocessing step applied to all methods. VLM-Judge is not a rewrite of AutoMindMap: it is validated independently by human-model quadratic weighted kappa (Table 15: 0.807 average) and by a cross-model re-evaluation with Qwen3.5-Plus (Table 16), and the paper acknowledges same-family bias in Appendix G.1. SLF thresholds (depth>7, fan-out>7) are taken from external working-memory and rubric sources (Miller 1956; Kedaj et al. 2014), not fitted to AutoMindMap. Appendix B's admission that no real educational-scenario experiments were run is a limitation on the external validity of 'cognitive efficiency,' but it does not make the benchmark ranking circular. No equation is defined in terms of the model's outputs, and no parameter is fitted to the test set. The only concerns are author-designed benchmark and same-family judge; these are validity risks, not circular reductions.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claims rest on benchmark and proxy assumptions: what a valid or cognitively efficient map is, what the reference maps represent, and whether VLM judgments stand in for learners. There are no invented physical entities. The pipeline also has hand-set constants for retrieval, matching, and refinement that affect reported scores.

free parameters (5)
  • cross-branch retrieval K = 3
    Number of candidate nodes retrieved for cross-branch association (Appendix D); hand-set with no reported sensitivity analysis.
  • similarity threshold gamma_sim = 0.6
    Threshold for cross-branch node retrieval (Appendix D); hand-set, directly affects cross-link density.
  • refinement rounds = 2
    Maximum number of editor-critic refinement rounds (Appendix D); not swept in the paper.
  • node matching thresholds tau and top-k = tau=0.6, k=3
    Used in the GT-based matching algorithm (Appendix E.1); controls all GT-based metric scores.
  • SLF abnormality thresholds = leaf depth 1 or >=7; non-leaf fan-out >7
    Structural Load Factor definition in Section S2M-Bench; grounded in Miller's 7 plus/minus 2 but implemented with arbitrary cutoffs.
axioms (6)
  • domain assumption A mind map is a rooted connected DAG with one central root; Pass@1 requires exactly one root, no isolated nodes, and no cycles.
    Definition 1 and Pass@1 in Section S2M-Bench. This structural definition excludes other valid knowledge representations and is the basis for structure-conformity evaluation.
  • domain assumption Cognitive efficiency can be operationalized by working-memory capacity (Miller 7 plus/minus 2) and progressive differentiation via simple depth and fan-out cutoffs.
    Used for SLF and VLM-Judge SQ/PE criteria. The paper runs no human learning-behavior experiment.
  • domain assumption Expert-annotated ground-truth maps are a faithful reference for the intrinsic knowledge hierarchy of a course.
    Dataset construction uses lecturer/PhD annotators and refinement rounds. Appendix G.4 shows human re-annotation varies substantially, so this is an assumption rather than a proven absolute.
  • domain assumption VLM-as-a-Judge scores approximate human judgments of mind map quality.
    Appendix G.1 reports quadratic weighted kappa 0.807 versus human-human 0.845. The assumption is required for all VLM-judge results.
  • domain assumption MinerU parsing and VLM image summaries adequately represent slide content.
    All non-human methods consume MinerU Markdown. Appendix B acknowledges dependence on parsing quality.
  • domain assumption LLM agents can reliably execute atomic graph-editing tools so that structural constraints hold.
    AutoMindMap relies on tool calls and compliance checks to achieve Pass@1=100 percent. This is an empirical property of the LLM, not a formal guarantee.

pith-pipeline@v1.3.0-alltime-deepseek · 31860 in / 14912 out tokens · 174831 ms · 2026-08-05T00:32:15.506561+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Slides2MindMap: Reconstructing Cognitively Efficient Knowledge Hierarchies from Lecture Slides." pith.science (2026). https://pith.science/paper/WO6CBMCL

@misc{pith2026260800610,
  author       = {Pith},
  title        = {Pith review of: Slides2MindMap: Reconstructing Cognitively Efficient Knowledge Hierarchies from Lecture Slides},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WO6CBMCL}},
  note         = {Machine review of arXiv:2608.00610}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generating mind maps from lecture slides can help learners efficiently assimilate fragmented knowledge, promising substantial benefits for intelligent education. However, dedicated automatic generation and evaluation frameworks remain underexplored and challenging, requiring a global-local knowledge focus balance and handling large-scale, heterogeneous slides. 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 ground-truth-based comparison, structure conformity analysis, and VLM-as-a-Judge. To address this task, we propose AutoMindMap, an agentic framework inspired by the Structure Building Framework. AutoMindMap comprises Skeleton Laying for global scaffold anchoring, Iterative Knowledge Integration augmented by context-aware 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 pedagogical application value.

Figures

Figures reproduced from arXiv: 2608.00610 by Chuang Zhou, Huachi Zhou, Jiaqi Bai, Rongjun Ye, Shengyuan Chen, Xiao Huang, Yuzhi Wang, Zhicong Hong.

Figure 1
Figure 1. Figure 1: Illustration of Slides2MindMap. Left: mind maps provide more intuitive structure for visual learning, but need to strike a global-local balance. Right: the flexible organiza￾tion of slides makes explicit hierarchy extraction difficult. from course materials remains labor-intensive and expertise￾demanding, motivating the pursuit of automated genera￾tion (Zhang et al. 2024), particularly forslide-oriented mi… view at source ↗
Figure 2
Figure 2. Figure 2: Two typical cases of the heterogeneity. Left: in￾termediate taxonomies of descendant concepts might be im￾plicit and need to be inferred. Right: a lot of noise and knowl￾edge interleaving are for pedagogical purposes, and hierarchy reconstruction needs further reasoning. concepts into progressive and meaningful hierarchies around a central theme, distilling dense relational networks into an intuitive, bran… view at source ↗
Figure 3
Figure 3. Figure 3: Overall framework of AutoMindMap. The main pipeline receives the parsed Markdown files of a course, then lays the initial skeleton with key associative information. The mind map is incrementally expanded with memory and tool-augmented integration. The map is further refined by an editor-critic framework that decouples macro-level and micro-level repair. Second, the integration of N k j into the global mind… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study result. The corresponding detailed data is in Table 12. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between AutoMindMap and Direct on different courses categories. Each metric is normalized by its maximum value across all results, resulting in a [0,1] scale. AutoMindMap performs better on each category, especially in Concept Understanding, showing the advantage in courses with more complex knowledge hierarchy. Appendix F pro￾vides more detailed analysis. detailed experiments are in Appendix F … view at source ↗
Figure 6
Figure 6. Figure 6: Further Statistic of S2M-Bench. E Evaluation Metrics E.1 Procedure for Concept Nodes Match A prerequisite for GT-based metrics is to establish a matched node-pair set C ⊆ V pred × V gt that aligns concepts with semantically equivalent meanings despite surface-form vari￾ations. To accommodate the inherent diversity in concept naming (e.g., abbreviations, synonyms, and paraphrases), we design a multi-stage m… view at source ↗
Figure 7
Figure 7. Figure 7: Examples of different categories in terms of content modality. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of correlation among main evaluation metrics, and scatter diagrams. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distributions of leaf-node layers (top) and direct child-node fan-out (bottom) in mind maps generated by different [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of the mind maps generated by different baselines. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: (Continued) Visualization of the mind maps generated by different baselines. [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visualization of the mind maps generated in the ablation study. [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 48 canonical work pages · 1 internal anchor

  1. [1]

    , author=

    The Role of Visual Learning in Improving Students' High-Order Thinking Skills. , author=. Journal of education and practice , volume=. 2016 , publisher=

    work page 2016

  2. [2]

    Computers , volume=

    A systematic review of mind maps, STEM education, algorithmic and procedural learning , author=. Computers , volume=. 2025 , publisher=

    work page 2025

  3. [3]

    2006 , publisher=

    The mind map book , author=. 2006 , publisher=

    work page 2006

  4. [4]

    , author=

    The use of advance organizers in the learning and retention of meaningful verbal material. , author=. Journal of educational psychology , volume=. 1960 , publisher=

    work page 1960

  5. [5]

    Psychology of learning and motivation , volume=

    Cognitive load theory , author=. Psychology of learning and motivation , volume=. 2011 , publisher=

    2011

  6. [6]

    Prospects , volume=

    Concept mapping to facilitate teaching and learning , author=. Prospects , volume=. 1995 , publisher=

    work page 1995

  7. [7]

    arXiv preprint arXiv:2407.05458 , year=

    A survey of models for cognitive diagnosis: New developments and future directions , author=. arXiv preprint arXiv:2407.05458 , year=

    Pith/arXiv arXiv

  8. [8]

    International Workshop on Learning Technology for Education Challenges , pages=

    BoomMapper: An Innovative Web-Based System for Transforming PowerPoint Presentations into Mind Maps , author=. International Workshop on Learning Technology for Education Challenges , pages=. 2025 , organization=

    work page 2025

  9. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Coreference graph guidance for mind-map generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  10. [10]

    International Conference on Soft Computing and Data Mining , pages=

    A Presentation Mining Framework: From Text Mining to to Mind Mapping , author=. International Conference on Soft Computing and Data Mining , pages=. 2024 , organization=

    work page 2024

  11. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Slidevqa: A dataset for document visual question answering on multiple images , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  12. [12]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Efficient mind-map generation via sequence-to-graph and reinforced graph refinement , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2021

  13. [13]

    World Wide Web , volume=

    Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities , author=. World Wide Web , volume=. 2024 , publisher=

    work page 2024

  14. [14]

    ACM Computing Surveys , volume=

    A systematic survey of text summarization: From statistical methods to large language models , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    work page 2025

  15. [15]

    arXiv preprint arXiv:2501.13958 , year=

    A survey of graph retrieval-augmented generation for customized large language models , author=. arXiv preprint arXiv:2501.13958 , year=

    arXiv

  16. [16]

    Evaluation and program planning , volume=

    An introduction to concept mapping for planning and evaluation , author=. Evaluation and program planning , volume=. 1989 , publisher=

    work page 1989

  17. [17]

    Higher education , volume=

    Concept mapping, mind mapping and argument mapping: what are the differences and do they matter? , author=. Higher education , volume=. 2011 , publisher=

    work page 2011

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Make your llm fully utilize the context , author=. Advances in Neural Information Processing Systems , volume=

    work page

  19. [19]

    Behaviormetrika , volume=

    Exploring the psychometric properties of the mind-map scoring rubric , author=. Behaviormetrika , volume=. 2019 , publisher=

    work page 2019

  20. [20]

    PageIndex Blog , year =

    Mingtian Zhang and Yu Tang and PageIndex Team , title =. PageIndex Blog , year =

    work page

  21. [21]

    arXiv preprint arXiv:2512.03413 , year=

    BookRAG: A Hierarchical Structure-aware Index-based Approach for Retrieval-Augmented Generation on Complex Documents , author=. arXiv preprint arXiv:2512.03413 , year=

    arXiv

  22. [22]

    2013 , publisher=

    Language comprehension as structure building , author=. 2013 , publisher=

    work page 2013

  23. [23]

    Vanderbilt University Center for Teaching , volume=

    Bloom’s taxonomy , author=. Vanderbilt University Center for Teaching , volume=. 2010 , publisher=

    work page 2010

  24. [24]

    , author=

    The magical number seven, plus or minus two: Some limits on our capacity for processing information. , author=. Psychological review , volume=. 1956 , publisher=

    work page 1956

  25. [25]

    Acta Informatica Pragensia , volume=

    Effective mind maps in e-learning , author=. Acta Informatica Pragensia , volume=

    work page

  26. [26]

    JALT 2013 conference proceedings , volume=

    Likert-scale questionnaires , author=. JALT 2013 conference proceedings , volume=

    work page 2013

  27. [27]

    International symposium on graph drawing , pages=

    Graphviz—open source graph drawing tools , author=. International symposium on graph drawing , pages=. 2001 , organization=

    work page 2001

  28. [28]

    arXiv preprint arXiv:2409.18839 , year=

    Mineru: An open-source solution for precise document content extraction , author=. arXiv preprint arXiv:2409.18839 , year=

    Pith/arXiv arXiv

  29. [29]

    Approaches to specialized genres , pages=

    Genre, pedagogy, and PowerPoint design: A multimodal move analysis of linguistics lecture slides , author=. Approaches to specialized genres , pages=. 2020 , publisher=

    work page 2020

  30. [30]

    , author=

    Updating a situation model: a memory-based text processing view. , author=. Journal of Experimental Psychology: Learning, Memory, and Cognition , volume=. 1998 , publisher=

    work page 1998

  31. [31]

    Communications of the ACM , volume=

    MapReduce: simplified data processing on large clusters , author=. Communications of the ACM , volume=. 2008 , publisher=

    2008

  32. [32]

    arXiv preprint arXiv:2404.16130 , year=

    From local to global: A graph rag approach to query-focused summarization , author=. arXiv preprint arXiv:2404.16130 , year=

    Pith/arXiv arXiv

  33. [33]

    Education and information technologies , volume=

    A systematic literature review on educational recommender systems for teaching and learning: research trends, limitations and opportunities , author=. Education and information technologies , volume=. 2023 , publisher=

    work page 2023

  34. [34]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Tutorialbank: A manually-collected corpus for prerequisite chains, survey extraction and resource recommendation , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  35. [35]

    Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025) , pages=

    Beyond Linear Digital Reading: An LLM-Powered Concept Mapping Approach for Reducing Cognitive Load , author=. Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025) , pages=

    work page 2025

  36. [36]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Tree-KG: An expandable knowledge graph construction framework for knowledge-intensive domains , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  37. [37]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    What should i learn first: Introducing lecturebank for nlp education and prerequisite chain learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  38. [38]

    International Journal of Information Technology , volume=

    Logical concept mapping and social media analytics relating to cyber criminal activities for ontology creation , author=. International Journal of Information Technology , volume=. 2023 , publisher=

    work page 2023

  39. [39]

    arXiv preprint arXiv:2404.19146 , year=

    Automated construction of theme-specific knowledge graphs , author=. arXiv preprint arXiv:2404.19146 , year=

    Pith/arXiv arXiv

  40. [40]

    Alexandria Engineering Journal , volume=

    EduVQA: A multimodal Visual Question Answering framework for smart education , author=. Alexandria Engineering Journal , volume=. 2025 , publisher=

    work page 2025

  41. [41]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Awaking the slides: a tuning-free and knowledge-regulated AI tutoring system via language model coordination , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

    work page

  42. [42]

    Journal of Science Education and Technology , volume=

    Constructivist-visual mind map teaching approach and the quality of students’ cognitive structures , author=. Journal of Science Education and Technology , volume=. 2011 , publisher=

    work page 2011

  43. [43]

    Annual Meeting of the Association for Computational Linguistics , year=

    MOOCCube: A Large-scale Data Repository for NLP Applications in MOOCs , author=. Annual Meeting of the Association for Computational Linguistics , year=

    work page

  44. [44]

    Course Concept Expansion in MOOCs with External Knowledge and Interactive Game , DOI=

    Yu, Jifan and Wang, Chenyu and Luo, Gan and Hou, Lei and Li, Juanzi and Liu, Zhiyuan and Tang, Jie , year=. Course Concept Expansion in MOOCs with External Knowledge and Interactive Game , DOI=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , publisher=

    work page

  45. [45]

    2025 , eprint=

    LLM-Assisted Knowledge Graph Completion for Curriculum and Domain Modelling in Personalized Higher Education Recommendations , author=. 2025 , eprint=

    work page 2025

  46. [46]

    WIREs Data Mining and Knowledge Discovery , volume =

    Ferreira-Mello, Rafael and André, Máverick and Pinheiro, Anderson and Costa, Evandro and Romero, Cristobal , title =. WIREs Data Mining and Knowledge Discovery , volume =. doi:https://doi.org/10.1002/widm.1332 , year =

    doi:10.1002/widm.1332

  47. [47]

    Towards Comprehensive Lecture Slides Understanding: Large-Scale Dataset and Effective Method , year=

    Zhang, Enming and Li, Yuzhe and Liu, Yuliang and Zhu, Yingying and Bai, Xiang , booktitle=. Towards Comprehensive Lecture Slides Understanding: Large-Scale Dataset and Effective Method , year=

    work page

  48. [48]

    Zhe Chen and Heyang Liu and Wenyi Yu and Guangzhi Sun and Hongcheng Liu and Ji Wu and Chao Zhang and Yu Wang and Yanfeng Wang , booktitle =. M. 2024 , editor =

    work page 2024

  49. [49]

    Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos , year=

    Lee, Dong Won and Ahuja, Chaitanya and Liang, Paul Pu and Natu, Sanika and Morency, Louis-Philippe , booktitle=. Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos , year=

    work page

  50. [50]

    , author=

    Understanding interobserver agreement: the kappa statistic. , author=. Family medicine , year=

    work page

  51. [51]

    2024 , eprint=

    MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark , author=. 2024 , eprint=

    work page 2024

  52. [52]

    Journal of the American Statistical Association , year=

    Estimates of the Regression Coefficient Based on Kendall's Tau , author=. Journal of the American Statistical Association , year=

    work page

  53. [53]

    and Shneiderman, B

    Johnson, B. and Shneiderman, B. , booktitle=. Tree-maps: a space-filling approach to the visualization of hierarchical information structures , year=

    work page

  54. [54]

    Educational Psychology Review , year=

    Cognitive Architecture and Instructional Design: 20 Years Later , author=. Educational Psychology Review , year=

    work page

  55. [55]

    1988 , url=

    Cognitive flexibility theory : advanced knowledge acquisition in ill-structured domains , author=. 1988 , url=

    work page 1988

  56. [56]

    The Reading Teacher , volume=

    Theories and practices of multimodal education: The instructional dynamics of picture books and primary classrooms , author=. The Reading Teacher , volume=. 2009 , publisher=

    work page 2009

  57. [57]

    arXiv preprint arXiv:2602.12430 , year=

    Agent skills for large language models: Architecture, acquisition, security, and the path forward , author=. arXiv preprint arXiv:2602.12430 , year=

    Pith/arXiv arXiv

  58. [58]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

  59. [59]

    LLM-Based Educational Simulation: Evaluating Temporal Student Persona Stability Across ADHD Profiles

    LLM-Based Educational Simulation: Evaluating Temporal Student Persona Stability Across ADHD Profiles , author=. arXiv preprint arXiv:2605.06307 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2410.17558 , year=

    Clr-bench: Evaluating large language models in college-level reasoning , author=. arXiv preprint arXiv:2410.17558 , year=

    Pith/arXiv arXiv