acceptodds
Withdrawn from ICLR 2027

GraphTalk: Slide-Anchored Topic Graphs with Supporting Document Evidence for Presentation Narration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating spoken narration (a talk track) for an existing slide deck requires following the slides' topical structure while grounding each slide in evidence from long supporting documents. We present GraphTalk, a training-free pipeline built around an explicit, interpretable structured planning layer that replaces the flat, high-volume source context used by prior approaches. GraphTalk induces atomic topics from the slides, enriches them with retrieved document evidence, ranks them over a similarity graph with a PageRank-seeded variant of DivRank, and allocates them to slides by capacity-constrained matching in a shared embedding space. On two benchmarks (M3AV academic lectures and SEC EDGAR financial filings), two generation models, and a dual-ladder of deterministic and LLM-as-judge metrics, its central result is the strongest document-grounded slide coverage and relevance, holding across deterministic metrics and three LLM judges, at a token budget comparable to the strongest baselines. A small human study supports the coverage advantage. Automatic judges are unreliable for supporting-document usage, where our human study is more trustworthy and also favours GraphTalk. We validate judge reliability with multi-judge agreement, an image-grounded judge, and perturbation tests.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Talk Track Generation, DivRank, Structured Planning, Multi-Source Grounding, Bipartite Matching, Long-context Generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_z53cOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GraphTalk is a training-free pipeline for generating per-slide narration from an existing slide deck and supporting documents. It extracts slide-anchored topics, enriches them with document evidence, and uses DivRank and capacity-constrained matching to allocate topics before drafting and polishing narrations. Experiments on M3AV and SEC EDGAR with two generators report improved coverage and relevance under LLM judges. A small human study supports coverage and document usage but gives lower scores on several other quality dimensions.

Critical strengths
  1. The task setting is relevant and the pipeline is modular. Existing slide decks do need grounded narration, and exposing topics and evidence could be useful for inspection. However, this limited strength does not compensate for the weak experimental comparison and evaluation problems below.
Critical weaknesses
  1. Weak comparisons and limited evidence for the planner. Experiments use Pixtral-12B and Gemma-3-27B with only Direct, Summ, and simple RAG baselines, leaving stronger multimodal and structured-generation comparisons untested. Tables 10 and 14 show that most gains come from slide anchoring, while DivRank and ILP contribute little. Missing prompts and retrieval details also prevent checking comparison fairness.
  2. Unconvincing evaluation. SEC EDGAR relies on LLM judges, while human evaluation covers only six presentations and two team annotators. Table 5 rates GraphTalk substantially below Direct/RAG on relevance, faithfulness, and several other dimensions despite higher coverage. This conflict remains unexplained, and incomplete failure/cost reporting further weakens the conclusions.
Questions for authors
  1. Clarification: Are prompts, visual inputs, context limits, and polishing matched across methods, and what evidence isolates the planner's benefit beyond slide anchoring?
  2. Clarification: How do you explain the lower human-rated relevance and faithfulness in Table 5, and how were evaluation subsets and failed judge calls handled?
Reviewer_MCC7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents GraphTalk, a training-free pipeline for generating narration for an existing slide deck using supporting documents. It extracts topics from the slides, enriches them with document evidence, and applies graph ranking and global allocation to plan the narration. Experiments cover academic and financial presentations with two generators, automatic evaluation, component ablations, and a small human study. The results primarily support improved coverage of slide content using supporting documents.

Critical strengths
  1. The paper addresses a practical problem of generating narration that follows an existing slide deck while adding relevant information from supporting documents. Organizing evidence around slide topics is a well-motivated and interpretable approach to this task.
  2. The evaluation covers two domains and two generators, with multiple judges, human evaluation, and component ablations. The token-matched RAG and LLM-planning controls strengthen the evidence for the coverage gains.
Critical weaknesses
  1. The contribution of global graph planning beyond slide anchoring is insufficiently established. The evidence mainly supports organizing information around slides, leaving the paper’s central methodological contribution weakly supported without a clear benefit over simpler per-slide selection and deduplication.
  2. In the human evaluation, GraphTalk shows advantages only in coverage and supporting-document usage (SDUsage). These gains do not clearly establish better overall narration quality, since the unresolved trade-off with relevance and faithfulness leaves the method’s practical benefit uncertain.
Questions for authors
  1. With topics, evidence, and generation budget held fixed, what improvement does the full planner provide over simple per-slide selection and deduplication?
  2. How do human and automatic scores compare on the same six presentations, particularly for relevance and faithfulness? Could the authors provide per-presentation results?
  3. Does the coverage advantage remain when narration length or speaking time is matched, without sacrificing relevance or faithfulness?
Reviewer_TQatOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GraphTalk, a training-free framework for document-grounded slide narration. GraphTalk generates presentation narration through two phases: Global Topical Planning and Local Slide Realization. Experiments demonstrate strong performance in slide coverage and relevance. A human study further supports its advantages in coverage and supporting-document usage.

Critical strengths
  • The paper addresses a practically meaningful problem of generating coherent, document-grounded narration for existing presentation slides.

  • The proposed structured planning framework provides an interpretable mechanism for organizing supporting evidence and allocating topics across slides.

Critical weaknesses
  • The human evaluation is limited to six presentations and two annotators. More importantly, GraphTalk performs substantially worse than the baselines on all human-rated dimensions except coverage and supporting-document usage. These results appear inconsistent with the automatic evaluations and raise concerns about the overall narration quality.

  • Inter-judge agreement is low for source faithfulness, supporting-document usage, and diversity, with limited analysis of the reasons behind these discrepancies and their implications for evaluation reliability.

  • The paper lacks qualitative examples of generated narration. Providing side-by-side comparisons of GraphTalk, baseline outputs, and ground-truth narration, together with the corresponding slides and supporting evidence, would help readers better assess narration quality, factual grounding, and the differences between methods.

Questions for authors
  • Could the authors explain the low performance on several human evaluation metrics and consider conducting a larger human study?

  • Could the authors further explain the low inter-judge agreement for these metrics?

  • Could the authors provide qualitative examples comparing GraphTalk with baseline methods and ground-truth narrations, along with the corresponding slides and supporting-document evidence, to better illustrate the differences in narration quality in the final version?

  • Several recent works on slide generation [1, 2, 3] are not discussed in the related work.

[1] Chengzhi Liu et al. Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations. In International Conference on Learning Representations (ICLR), pages 133655–133696, 2026.

[2] Tarik Can Ozden et al. Narrative-Driven Paper-to-Slide Generation via ArcDeck. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520–537, 2026.

[3] Hao Zheng et al. DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation. In Findings of the Association for Computational Linguistics (ACL), pages 31545–31558, 2026. URL: https://aclanthology.org/2026.findings-acl.1578/

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

Related papers

Discussion (0)

Sign in to comment.