acceptodds
Withdrawn from ICLR 2027

SciSlide: Scientific Slide Generation as Verifiable Argument Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Preparing slides for a scientific talk is a document-to-multimodal generation task: text, figures, and equations must be placed on a visual canvas to guide an audience through the paper's causal argument. We formulate this task as argument reconstruction rather than extractive summarization, with explicit verification at each step. SciSlide builds a multi-slide argument graph in which every node carries a verbatim source span and every figure attaches to its supporting claim, enabling string-based verification of coverage and grounding. We score decks with sixteen computable metrics that operationalize the assertion–evidence paradigm for technical presentations, and learn only the layout component (coordinate prediction and geometric composition) from 7,500 conference decks paired with their papers. Against four published systems and a prompted frontier LLM, SciSlide leads on assertion headlines, argument-driven slide order, and figure–claim alignment; human raters prefer its decks in 65% of pairwise comparisons. For layout, autoregressive coordinate generation outperforms classification by 6.2 percentage points in strict exact match.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
paper-to-slides generation, scientific presentation generation, multimodal document generation, autoregressive layout generation, argument structure, grounded generation, reference-free evaluation, assertion-evidence paradigm, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Jd1EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SciSlide, a system that turns scientific papers into presentation slides. It first builds an argument graph from the paper, then plans slides around individual claims and their supporting evidence, and finally generates the layout with a fine-tuned language model. It also introduces a dataset of 7,500 paper–slide pairs and 16 automatic evaluation metrics. The results are encouraging, including the human evaluation. My main concern is that the paper makes stronger claims about verification than its checks support. I also have questions about the evaluation setup and how much each component contributes to the final improvement.

Critical strengths
  1. The problem is useful, and the overall approach makes sense. Separating argument extraction, slide planning, and layout generation gives the system a clear structure.
  2. The evaluation goes beyond visual appearance. I appreciate the attempt to measure content coverage, slide order, and figure placement. The paper also discusses cases where a metric gives a good score to a poor presentation. This is useful.
  3. Several experimental checks strengthen the results. The paper includes component ablations, a Gemini baseline explicitly prompted with assertion–evidence rules, and an additional evaluation using a graph extracted by GPT-4o. The human study also provides evidence beyond the automatic metrics.
Critical weaknesses
  1. The verification checks do not establish that the scientific claims are correct. This is my main concern. In Section 3.3, a node passes the grounding check if its source span appears in the paper. That establishes where the text came from, but it does not establish that the assigned role, inferred relation, or rewritten slide headline is correct. Similarly, placing a figure on the slide associated with its assigned node does not establish that the figure supports the headline. For example, a headline could turn a result observed under one experimental setting into a general claim, while still pointing to a valid source passage. A figure could also be attached to the wrong claim during graph construction and then pass the later placement check. I would like the authors to narrow this statement and directly check a sample of graph relations, rewritten headlines, and figure–claim links against the source papers. The agreement with hand-computed metric values in Appendix B checks the implementation of the rules; it does not resolve this question.
  2. The argument metrics still make it difficult to separate presentations from closer agreement with the system’s representation. SciSlide plans its content using an argument graph, while NC@budget and CER evaluate outputs against that graph. The GPT-4o graph experiment is a useful check, and I appreciate its inclusion. However, it uses the same extraction procedure and still measures node realization through verbatim eight-word matches. A correct paraphrase can therefore miss credit, while copied text can receive credit without explaining the argument clearly. I also do not see a common slide-count or total content budget across systems. Limiting each slide to five content items does not control how many slides a system produces or how long each item is. This matters when comparing coverage.
  3. The human results are promising, but the reported preference rate needs more context. Section 5.3 reports an average score of 3.8 for SciSlide and 3.5 for PPTAgent. The headline result of 65% preference combines comparisons against all five baselines. I would like to see the preference rate against each baseline, especially PPTAgent. The pairwise preferences are derived from five-point ratings rather than collected as direct pairwise choices. Please explain how the two raters’scores are combined and how ties are handled. There is also a dependency issue to clarify: the 150 pairs come from only 30 source papers, and each SciSlide deck is reused in five comparisons. Please explain whether the confidence interval accounts for this grouping.
  4. The ablations do not yet show which components improve the final presentation quality. Appendix E shows that removing components worsens their corresponding metrics. These results are useful, but they mainly show that the components affect the properties they were designed to target. For example, removing assertion labels lowers the assertion-headline score, while removing graph conditioning lowers graph-based coverage. I would like to see a small human comparison of the full system against versions without graph conditioning and without correction.The learned layout component also needs a clearer connection to presentation quality. Matching the coordinates of one reference layout is a reasonable prediction task, but several different layouts may work equally well. It remains unclear how much the exact-match improvement translates into better readability or visual organization.
  5. Please reconcile the BLR results. Table 1 reports a BLR of 0.011 for SciSlide, while the full-system value in the A3 ablation is 0.09. If these refer to different configurations, that needs to be stated.
  6. The explanation for the geometry-loss result is too strong. The reference layout can minimize both the coordinate prediction loss and the geometry penalty. Having additional valid layouts does not itself establish that the objectives conflict.
Questions for authors

NA

Reviewer_1WHHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SciSlide, a system for generating scientific presentation slides through argument reconstruction. It extracts an argument graph with source spans and links to figures and tables, plans slides around assertion headlines and supporting evidence, and generates layouts using a fine-tuned Qwen2.5 model. Deterministic metrics are used for evaluation and page-local correction. The paper also introduces a corpus of 7,500 paper–deck pairs and 16 evaluation metrics. Evaluation includes comparisons with five alternatives on 300 test papers, human ratings of 180 decks from 30 papers, and layout and component ablations. The results show improvements on several proposed metrics and higher overall human ratings, while the layout experiments report an advantage for autoregressive coordinate generation over a classification head.

Critical strengths

S1. The three-layer design makes the links between source evidence, slide content, and layout explicit. The human evaluation supports its practical value, with a mean suitability rating of 3.8 versus 3.5 for PPTAgent.

S2. The 7,500-pair corpus and analysis of metric failure modes contribute to the study and evaluation of slide generation.

Critical weaknesses

W1. The grounding checks establish source traceability but do not verify that the evidence supports the claim. Specifically, as the only example in this work (The reviewer thought there should have been more) in Figure 6, the second SciSlide slide has a headline about 2D human pose estimation, but displays the table for 3D pose and mesh performance on Human3.6M and 3DPW (it is kind of blurry, so please correct the reviewer if it is not). The third slide has a headline about 3D pose estimation and mesh recovery, but displays the 2D COCO results table. The two tables appear to have been exchanged. I cannot tell whether this happened during generation or when Figure 6 was assembled, but it needs to be resolved given the work has an explicit mechanism to prevent this.

W2. The methodological advance over prior discourse-based slide generation is not sufficiently established, as ArcDeck (Ozden et al., 2026) already uses rhetorical structure to guide slide planning. SciSlide adds explicit source links and deterministic checks, but its benefit over this closely related prior approach is not evaluated.

Questions for authors

Q1: Are the apparent 2D/3D table swaps in Figure 6 generation errors or figure-assembly errors, and what evidence supports the claimed correctness of claim–evidence mappings?

Q2: Compared with ArcDeck, what is the main technical advance in SciSlide’s argument representation and verification, and what evidence establishes its benefit under comparable generation settings?

Q3: Were raters given the source papers to assess factual accuracy and claim–evidence support? Please also report preference rates against individual baselines and explain how the confidence interval accounts for repeated comparisons involving the same paper.

Reviewer_T5D3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In the paper, the authors turn research papers into talk slides by first rebuilding the paper's argument. A state of the art model (Gemini is the current case ) pulls out the key pieces of the paper (problem, method, evidence, claim), each tied to an exact quote from the PDF. It then links them into a graph and attaches each figure to the claim it supports. Slides follow the order of this graph, and each slide has a full-sentence headline (e.g., "Our method beats X on Y") with the supporting figure underneath, instead of a topic title over bullets. The only trained part is the layout model, a fine-tuned Qwen 14B that decides where things go on the slide, trained on 7.5k real conference decks. Slides that fail the authors' checks get rewritten by Claude. They evaluate with 16 rule-based metrics on 300 papers against four existing systems and plain Gemini, plus a human study on 180 decks, where SciSlide is preferred 65% of the time.

Critical strengths

One of the stronger aspects of the paper is the evaluation framework. All sixteen metrics are implemented as deterministic programs, avoiding the variability and known biases of LLM-based judging. The authors also check the implementations against hand-computed values on 200 units per metric, with reasonably strong agreement. I think this makes the evaluation reproducible and potentially useful beyond the proposed system. My concern is less about the reliability of the implementations and more about whether the metrics capture the properties they claim to measure.

I also appreciate that the authors critically examine their own metrics rather than assuming that higher scores necessarily imply better presentations. The examples involving verbatim copying and nearly empty slides demonstrate how some metrics can be optimised without improving quality. They further examine correlations with human ratings while controlling for system identity, which makes the analysis more meaningful. Interestingly, the argument-related metrics show relatively weak associations with human judgments (NC@budget ρ=0.24 and DTW-Flow ρ=-0.22), compared with assertion-headline ratio (ρ=0.38). This doesn't necessarily invalidate the argument representation, since overall presentation quality and argument fidelity are not equivalent, but it leaves the central claim less directly supported. To me, the evaluation methodology is a more convincing contribution than the evidence for verifiable argument reconstruction.

Critical weaknesses

My main concern is that much of the evaluation measures how closely SciSlide follows the structure and presentation criteria it was designed around. This is not necessarily a problem for checking whether the system works as intended, but it becomes difficult to interpret these results as evidence of better argument reconstruction. The argument metrics are evaluated against graphs produced using SciSlide's extraction procedure, while figure–claim alignment is already encouraged by how assets are attached to nodes and headlines are generated. The correction loop further optimises these metrics before the final comparison. Although the authors re-extract the graph using GPT-4o, the underlying representation and extraction rules remain unchanged. I think the reported improvements are meaningful within the proposed framework, but it is unclear how well they would hold against an independently annotated argument structure. The graph ablation shows that graph conditioning matters for the chosen metrics, but it does not fully resolve this concern.

The human evaluation is encouraging, but I don't think it sufficiently validates the main contribution. The study covers 30 papers with two raters per deck, and the difference between SciSlide and PPTAgent (3.8 versus 3.5) is not accompanied by a significance test for that particular comparison. More importantly, the raters assess overall presentation suitability rather than whether the scientific argument has been reconstructed correctly. Table 6 shows that the argument metrics have weaker within-system correlations with human ratings than the assertion-headline metric, although this may partly reflect what the raters were asked to evaluate. The graph itself is also not validated against human annotations, so we cannot tell whether its causal edges and claim–evidence relationships are correct. I am additionally unsure how much of the improvement can be attributed to the proposed representation, given that SciSlide uses Gemini 3 Pro for extraction and planning and Claude Sonnet 4.6 for correction, while the baselines use their existing pipelines. A controlled comparison using the same underlying models, along with an independent graph evaluation, would make the contribution much clearer.

Questions for authors
  1. Appendix A explains how graph nodes are matched to baseline content using 8-grams, but I am still unclear how the slide-level role sequences are constructed for DTW-Flow when the baselines do not produce rhetorical roles. Could the authors clarify this procedure? Since the reference graph is generated using SciSlide's extraction method, I am concerned that differences in lexical matching or role assignment could influence the comparison.

  2. Ret and Hall are both listed as comparable metrics in Table 2, but neither appears in the main results in Table 1. Could the authors report these values for all systems? They seem particularly relevant given the emphasis on source grounding and preserving scientific content.

  3. Section 5.4 reports that correction improves AHR by around 0.10 and reduces OOB from 0.032 to 0.002. Since Table 1 reports results after correction, could the authors provide the pre-correction scores as well? It would also be useful to apply an equivalent correction procedure to the baselines, where feasible, to understand how much of SciSlide's advantage comes from the argument-based generation process rather than the subsequent metric-guided refinement.

  4. The GPT-4o re-extraction experiment uses the same extraction procedure, including the predefined roles, span constraints and edge rules. Could the authors evaluate the extracted graphs against human annotations for a smaller subset of papers, perhaps 20–30? In particular, I would be interested in the correctness of node roles, causal relationships and figure attachments, and whether the improvements in NC@budget, CER and DTW-Flow remain when evaluated against these independently constructed graphs.

  5. In Section 5.2, explicitly prompting Gemini 3 Pro with assertion–evidence instructions increases AHR to 0.68, but NC@budget remains at 0.43. What happens when Gemini 3 Pro is also provided with the extracted argument graph? This seems like a useful comparison to determine whether the improvement primarily comes from having a structured argument representation or from the remaining components of SciSlide. It would also help address the differences in model configurations between SciSlide and the baselines.

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

Related papers

Discussion (0)

Sign in to comment.