acceptodds
Under review as a conference paper at ICLR 2027

Translating Research to Slides: Audience-Grounded Presentation Generation via Verifiable Transduction Graphs

Abstract

Generating a scientific talk from a paper is a constrained translation problem: one shared deck must serve listeners with different knowledge and goals while preserving evidence, required conditions, and claim scope. Prior systems address important components of presentation generation—including source-content organization, audience-conditioned planning, and render-grounded refinement—but generally leave implicit how a shared narrative links source meaning to audience purpose, presentation order, and preservation requirements. We formulate this setting as joint-audience semantic transduction: audience needs may change expression and order, but cannot authorize unsupported claims. AudienceDeck aligns source-anchored semantic units and audience specifications into a frozen contract defining evidence bindings, audience goals, required conditions, and presentation budgets. Each adaptation becomes an executable transduction graph that records source meaning, transformation operation, story role, and conditions that must remain visible. Complete candidates are filtered by fidelity, core coverage, structure, slide, and time budgets; claims exposed before their required conditions are rejected before selection. Among feasible graphs, a prefix-sensitive selector estimates inference risk over visible narrative prefixes and terminal goal attainment after the complete narrative. The selected graph is compiled into traceable page contracts and slide objects, then checked through rendered-output verification and bounded repair. On Academic50, AudienceDeck ranks first among seven evaluated systems in macro coverage (0.7811), correctness (0.9744), and correctness-discounted time efficiency (1.9604), while ranking second in uncorrected time efficiency (2.015). On ArcBench, Judge and Quiz means reach 86.6 and 83.8. The resulting selection record connects source bindings, gate outcomes, selector decisions, rendered audits, and benchmark results in one auditable presentation process.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
paper-to-presentation generation, slide generation, audience adaptation, semantic transduction, multi-agent systems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5ibKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work presents AudienceDeck, a semantic transduction system that converts scientific papers into audience-aware presentation slides while preserving scientific fidelity.

Critical strengths

The experiments are comprehensive, covering a wide range of baseline systems and evaluation metrics.

Critical weaknesses
  1. Insufficient motivation. The pipeline design is underjustified and appears closely aligned with Academic50. It is unclear why jointly targeting Experts, Novices, and Adopters is preferable to generating audience-specific presentations, while in practice a presentation is usually adapted for a specific audience. Moreover, the optimization objectives closely match the benchmark metrics, raising concerns about generalizability beyond these settings.

  2. Unmatched baseline settings. As acknowledged in Lines 348–349, audience briefs, slide counts, and talk durations are not matched across systems. Comparisons under consistent settings would strengthen the results.

  3. Lack of human evaluation. The evaluation relies primarily on automated metrics and LLM judges. Human evaluation of presentation quality and audience comprehension would provide more convincing evidence.

  4. Limited ablation studies. The contributions of core components, such as the transduction graph and PrefixEval, are not isolated, making their effectiveness difficult to assess.

  5. Missing efficiency analysis. API costs, token consumption, and generation latency should be reported and compared with the baselines.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Author Response: Motivation, Evaluation and Ablations

Thank you for the constructive feedback. We clarify the intended setting and report additional baseline comparisons and component ablations.

Q1. Motivation and generalizability

Audience-specific decks are useful when listeners receive separate presentations. We address a single conference talk or seminar attended by people with different backgrounds and goals. Because everyone follows one sequence within a fixed time budget, the deck must explain prerequisites while retaining technical evidence and practical implications. Expert, Novice, and Adopter capture these complementary information needs, rather than fixed expertise levels.

The contract specifies source constraints and audience priorities, while the graph ties content transformations to their evidence. PrefixEval checks for potentially misleading inferences from the information visible so far. These choices follow from the demands of sequential communication. We also evaluate on ArcBench, whose narrative, visual, and quiz criteria provide a test beyond the Academic50 protocol.

Q2. Comparable baseline settings

The original runs used a 20-minute researcher brief on ArcBench and default settings on Academic50. These settings did not fully align audience conditioning across systems. We therefore added ArcDeck-Joint with the same Expert/Novice/Adopter prompt and a 20-slide, 20-minute target on all 100 ArcBench and 50 Academic50 papers. The table reports mean slide counts and Academic50 duration estimates, following X+Slides:

Estimated duration = 0.25 × slide count + visible word count / 130.

SystemArcBench slidesAcademic50 slidesAcademic50 minutes
ArcDeck14.816.512.2
DeepPresenter12.613.612.9
DeepSlides18.014.813.4
EvoPresent6.315.014.6
MemSlides14.713.914.4
SlideGen25.743.724.5
ArcDeck-Joint18.819.616.2
AudienceDeck14.415.314.2

On ArcBench, the joint specification increases ArcDeck's Judge/Quiz scores from 60.9/72.0 to 67.6/82.2, compared with 86.6/83.8 for AudienceDeck. Their actual output lengths still differ, as shown above.

Q3. Human evaluation

Our metrics assess information availability, source support, and question-based recovery as reproducible proxies for conditions supporting comprehension. X+Slides includes human and agent evaluation of probe fidelity, audience-relevant importance, and duplication (Fig. 2, Secs. 4.3, 6.1). ArcDeck/ArcBench includes a human evaluation in which 25 participants ranked slide decks by content quality and narrative coherence (Appendix A.1).

Q4. Transduction graph and PrefixEval

In a separate ablation of updated variants on 28 ArcBench papers, we held the source inventory, audience contract, narrative strategies, feasibility gates, backbone (GPT-5.5), and renderer constant.

VariantRepresentation and selectionJudgeQuiz
O-TStructured outline + terminal selection86.6484.46
G-TTransduction graph + terminal selection87.9292.75
G-PTransduction graph + prefix selection87.1492.79

O-T replaces explicit source–decision–slide bindings with a structured outline. G-T and G-P use the same candidate graphs: G-T maximizes final utility, while G-P also considers prefix risk.

G-T versus O-T isolates the graph: Judge improves by 1.28 and Quiz by 8.29.

Process evidence supports gate-guided revision. Of 236 initial drafts, 133 were flagged for conclusions preceding necessary conditions; 94 passed all gates after one revision. A blinded model evaluation covered 38 initial–revised pairs from 23 papers. The judge saw only visible text from the first several slides, without version labels or gate decisions. The proportion of judgments indicating conclusions preceding necessary conditions showed a paper-paired reduction of 2.18 percentage points (95% interval [0.94, 3.46]).

Thanks for providing the additional experimental results and clarifications. I will raise my score to 3.

Reviewer_AkwfOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AudienceDeck, a framework for generating scientific presentations for audiences with different knowledge levels and goals. The paper formulates this task as joint-audience semantic transduction, where the content and presentation order can be adapted to different audiences while preserving the original scientific meaning and evidence. Experiments demonstrate strong performance in presentation quality, audience coverage, and factual correctness.

Critical strengths
  • The paper addresses a meaningful problem of generating a single scientific presentation for audiences with different knowledge levels and information needs.

  • The proposed executable transduction graphs provide an interesting approach to connecting source evidence, audience goals, and presentation decisions while maintaining scientific fidelity.

  • The proposed method demonstrates strong performance across two benchmarks, outperforming existing baselines on several presentation quality, coverage, and correctness metrics.

Critical weaknesses
  • The choice of joint audience types (Expert, Novice, and Adopter) is not clearly motivated. The paper would benefit from a more detailed explanation of why these particular audience types were selected and whether they adequately represent the diverse audiences of scientific presentations.

  • The paper lacks a human user study to evaluate whether the generated presentations actually improve understanding across different audience types. The current evaluation relies on automated metrics, which may not fully reflect real audience comprehension.

  • The baseline comparisons could be made fairer by considering the joint-audience setting. Since ArcDeck supports a global audience specification, the authors could provide it with a combined description of Expert, Novice, and Adopter audiences rather than relying only on its native settings. This would allow for a fairer comparison of multi-audience presentation generation.

Questions for authors
  • Could the authors further explain the motivation behind selecting Expert, Novice, and Adopter as the three audience types? How would the framework generalize to other audience compositions?

  • Could the authors consider conducting a human user study involving participants with different expertise levels to evaluate actual audience comprehension and the effectiveness of the proposed audience-aware generation?

  • Could the authors provide an additional comparison with ArcDeck by specifying a combined multi-audience description in its input prompt? This would help clarify whether the improvements come from the proposed joint-audience semantic transduction framework rather than differences in audience specifications.

3 comments
AuthorsOpenReview ↗

Author Response: Audience Design, Human Evaluation, and Baseline Alignment

Thank you for recognizing the value of joint-audience presentation generation and executable transduction graphs. We clarify the audience design, discuss the human involvement in the benchmarks, and report the requested ArcDeck comparison.

Q1. Audience selection and generalization

We study one talk with a shared deck for listeners with different prior knowledge and goals, as in a conference presentation or departmental seminar. Within a fixed time budget, the same narrative must explain prerequisites, present technical evidence, and convey applicability conditions. Separate tailored decks are useful when audiences can receive separate presentations; our task addresses a shared presentation.

Prior knowledge affects what needs explanation, while audience goals affect what deserves emphasis. Paris (1988) shows that expertise affects information selection as well as explanatory detail (paper); X+Slides similarly distinguishes content depth from information needs (Sec. 4.1). We therefore use Expert, Novice, and Adopter to represent three complementary needs: assessing technical contributions and evidence, understanding concepts and prerequisites, and evaluating conditions for practical use.

These perspectives overlap and do not form an exhaustive audience taxonomy: an expert may also be an adopter. Other combinations can be specified through prior knowledge, goals, potential misconceptions, and goal weights in the audience contract. The graph links presentation decisions to source evidence, while prefix checks examine whether the information presented so far supports the intended interpretation. This design allows the audience specification to change without changing the underlying planning framework.

Q2. Human evaluation

Our metrics assess information availability, source support, and question-based recovery as reproducible proxies for conditions supporting comprehension. X+Slides includes human and agent evaluation of probe fidelity, audience-relevant importance, and duplication (Fig. 2, p. 5). ArcDeck/ArcBench includes 25 participants ranking content quality and narrative coherence (Appendix A.1).

These human components inform probe review and the evaluation of presentation quality. We appreciate the suggestion and will pursue a user study with participants of different expertise to assess comprehension and perceived effectiveness.

Q3. Joint-audience ArcDeck comparison

We completed the suggested experiment. ArcDeck-Joint uses ArcDeck's native global audience interface with the same joint Expert/Novice/Adopter description and a 20-slide, 20-minute target. Generation and evaluation cover all 100 ArcBench papers and all 50 Academic50 papers.

MetricArcDeck-Joint: new resultsAudienceDeck: submission results
Academic50: Coverage70.72%78.11%
Academic50: Correctness95.83%97.44%
Academic50: Safe E/time1.76491.9604
ArcBench: Mean Judge67.686.6
ArcBench: Mean Quiz82.283.8

With this configuration, ArcDeck's ArcBench Judge/Quiz scores increase from 60.9/72.0 to 67.6/82.2. AudienceDeck remains higher on these aggregate metrics, with a 19.0-point Judge lead.

The actual outputs still differ in length: ArcDeck-Joint averages 18.8 slides on ArcBench and 19.6 on Academic50, compared with 14.4 and 15.3 for AudienceDeck.

Thank you once more for your time and thoughtful review. With the discussion period nearing its end, we would like to check whether our response has resolved your concerns. If anything is still unclear, we would be glad to elaborate.

I thank the authors for their response, the experiments resolve my concerns and I increase my score to 3.

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

This paper presents AudienceDeck, an automatic generation framework for converting academic papers into presentation slides targeted at mixed audiences. Unlike conventional approaches that generate PPTs directly, the authors formulate this task as Joint-Audience Semantic Transduction, which aims to accommodate the comprehension needs of diverse audiences simultaneously. Experiments are conducted on two datasets, ArcBench and Academic50, and the proposed framework outperforms several existing methods overall.

Critical strengths

This work explicitly links the scientific semantics of the source paper, content transformation decisions, audience requirements, and specific narrative positions within the slides. It bridges planning and downstream generation via an executable Transduction Graph. Compared with direct generation methods or approaches that only evaluate final outputs, this structured design enables clearer tracing of the content transformation pipeline, and achieves strong results on both ArcBench and Academic50.

Critical weaknesses
  1. The two core designs of this work are the Executable Transduction Graph and Prefix-Sensitive Selection. However, the existing ablation experiments focus primarily on candidate selection and coverage-guided revision, without directly verifying the standalone contributions of these two key components. Since they constitute the main methodological innovations of the paper, the lack of dedicated ablation studies undermines the validation of the novelty and necessity of the proposed approach.

  2. As stated in Section 4.1, all baselines are run with their original configurations; their audience prompts, slide counts and presentation durations are not strictly aligned. Consequently, the overall performance gaps may be confounded by varying generation budgets and system setups, rather than solely reflecting the inherent advantages of AudienceDeck.

Questions for authors
  1. Regarding Weakness 1, could you add a direct ablation experiment comparing Prefix-Sensitive Selection against Terminal-Only Selection, which optimizes solely for the final audience utility of the complete presentation?
  2. Regarding Weakness 2, could you share more detailed information on audience prompts, slide counts and presentation durations for each baseline, to better contextualize the advantages of AudienceDeck?
2 comments
AuthorsOpenReview ↗

Author Response: Component Ablations and Baseline Settings

Thank you for recognizing the value of linking source semantics, audience needs, and presentation decisions. We added direct component ablations, a process evaluation of revision, and a joint-audience ArcDeck comparison.

Q1. Transduction graph and prefix-sensitive selection

We ran a separate ablation of three updated variants on the same 28 ArcBench papers. They share the source inventory, audience contract, four narrative strategies, feasibility gates, generation model (GPT-5.5), and final deck generation stage. O-T uses a structured outline that retains the same source references and goals but removes explicit source–decision–slide bindings. G-T and G-P use identical graph candidates.

VariantRepresentation and selectionMean JudgeMean Quiz
O-TOutline + terminal selection86.6484.46
G-TTransduction graph + terminal selection87.9292.75
G-PTransduction graph + prefix selection87.1492.79

G-T selects by the final audience utility of the complete presentation. G-P first applies the prefix-risk tolerance and then selects by terminal utility (Eqs. 5–7).

G-T versus O-T isolates the graph: Judge improves by 1.28 and Quiz by 8.29. G-P versus G-T directly tests selection with a fixed candidate pool. Separately, process evidence supports gate-guided revision. Of 236 initial drafts, 133 were flagged for conclusions preceding necessary conditions; 94 passed all gates after one revision. A blinded model evaluation covered 38 initial–revised pairs from 23 papers. The judge saw only visible text from the first several slides, without version labels or gate decisions. The proportion of judgments indicating conclusions preceding necessary conditions showed a paper-paired reduction of 2.18 percentage points (95% interval [0.94, 3.46]).

Q2. Audience instructions, slide counts, and duration

On ArcBench, the six baseline versions received a researcher-oriented brief specifying 20 minutes and 16:9, with no fixed slide count. On Academic50, we imposed no additional audience, duration, or slide-count requirements. Among these versions, ArcDeck exposed a dedicated global audience interface. We used it to add ArcDeck-Joint, supplying the joint Expert/Novice/Adopter description and a 20-slide, 20-minute target, across all 100 ArcBench and all 50 Academic50 papers. The other baselines retained their reproduction settings.

The table reports mean slide counts and estimated durations:

SystemArcBench slidesAcademic50 slidesAcademic50 minutes
ArcDeck14.816.512.2
DeepPresenter12.613.612.9
DeepSlides18.014.813.4
EvoPresent6.315.014.6
MemSlides14.713.914.4
SlideGen25.743.724.5
ArcDeck-Joint18.819.616.2
AudienceDeck14.415.314.2

Estimated duration is 0.25 × slide count + visible word count / 130, rather than measured speaking time. We have not yet aggregated comparable ArcBench duration estimates.

The joint configuration improves ArcDeck's ArcBench Mean Judge/Quiz scores from 60.9/72.0 to 67.6/82.2. The full-benchmark comparison is:

MetricArcDeck-Joint: new resultsAudienceDeck: submission results
Academic50: Coverage70.72%78.11%
Academic50: Correctness95.83%97.44%
Academic50: Safe E/time1.76491.9604
ArcBench: Mean Judge67.686.6
ArcBench: Mean Quiz82.283.8

AudienceDeck remains higher on all five metrics. On ArcBench, the Judge gap narrows from 25.7 to 19.0 points and the Quiz gap from 11.8 to 1.6 points.

Thank you once more for your time and thoughtful review. With the discussion period nearing its end, we would like to check whether our response has resolved your concerns. If anything is still unclear, we would be glad to elaborate.

Comments on the paper

2 comments

Reviewer 5ibK has read the author rebuttal and updates this scores. Now this paper we received three diverse scores, 2 weak acceptance and 1 weak rejection.

Dear reviewers vPHU and Akwf, could you please both see whether your concerns are addressed by author rebuttal and follow-up if you have additional concerns?

Thank you.

AuthorsOpenReview ↗

Summary of Contributions, Key Findings, and Discussion

Dear AC and Reviewers,

We thank all reviewers for their constructive feedback. The discussion helped us clarify the joint-audience setting and strengthen our empirical evaluation. We appreciate the reviewers recognizing the importance of the joint-audience setting (Akwf), the traceability of executable transduction graphs (vPHU), and the breadth of our experiments (5ibK)—as well as Akwf's confirmation that "the experiments resolve my concerns" and 5ibK's acknowledgment of our additional experiments and clarifications. We summarize our core contribution, key experimental findings, points addressed during the discussion, and the remaining open concern below.

Core contribution. AudienceDeck formulates paper-to-presentation generation as joint-audience semantic transduction: one shared narrative serves different information needs while preserving source evidence, required conditions, and claim scope.

  • Frozen contract: Fixes source constraints, audience goals, and presentation budgets before planning.
  • Executable transduction graph: Binds each adaptation to its source support, audience purpose, and narrative position.
  • Prefix checks and PrefixEval: A fidelity gate checks that required conditions are visible when each claim appears. PrefixEval compares feasible candidates by prefix risk and final audience goal attainment.
  • Visible verification: Checks that the selected meaning appears in rendered slides, with bounded repair.

Key findings.

  1. The lead persists against the strengthened baseline. We evaluated ArcDeck-Joint with the combined Expert/Novice/Adopter description and a 20-slide, 20-minute target across all 100 ArcBench and 50 Academic50 papers. AudienceDeck's submission results remain higher on all five reported aggregate metrics:
MetricArcDeck-JointAudienceDeck (submission)
ArcBench: Judge67.686.6
ArcBench: Quiz82.283.8
Academic50: Coverage70.72%78.11%
Academic50: Correctness95.83%97.44%
Academic50: Safe E/time1.76491.9604

These 19.0-point Judge and 1.6-point Quiz leads on ArcBench are achieved with fewer generated slides on average: 14.4 vs. 18.8 on ArcBench and 15.3 vs. 19.6 on Academic50.

  1. Explicit graph bindings improve information recovery. In supplementary ablations on the same 28 ArcBench papers—holding the source inventory, audience contract, narrative strategies, feasibility gates, GPT-5.5 backbone, and renderer constant—replacing the structured outline with the transduction graph under terminal selection (G-T vs. O-T) raises Judge from 86.64 to 87.92 (+1.28) and Quiz from 84.46 to 92.75 (+8.29).

  2. Prefix fidelity checks guide revision of premature conclusions. In a blinded model evaluation of 38 initial–revised pairs from 23 papers (where the judge saw only visible text from opening slides without version labels or gate decisions), judgments indicating conclusions preceding necessary conditions showed a paper-paired reduction of 2.18 percentage points after gate-guided revision (95% interval [0.94, 3.46]).

Concerns addressed during the discussion.

  • Audience selection and generalization (Akwf, 5ibK): We clarified the shared-talk setting and the complementary, overlapping needs represented by Expert, Novice, and Adopter. Other audience compositions can be specified via prior knowledge, goals, potential misconceptions, and weights in the contract. ArcBench provides an evaluation beyond the Academic50 protocol.
  • Baseline settings (all reviewers): We added the requested joint-audience ArcDeck comparison and reported audience instructions, target budgets, actual slide counts, and Academic50 duration estimates, distinguishing requested budgets from realized outputs.
  • Component ablations (vPHU, 5ibK): O-T vs. G-T isolates the graph's contribution. G-T vs. G-P holds graph candidates fixed and changes the selection rule: G-P reaches 87.14 Judge / 92.79 Quiz (close to G-T's 87.92 / 92.75) while incorporating prefix risk. The revision study evaluates how prefix fidelity checks guide correction before candidate selection.

Remaining concern and our explanation.

  • Direct human comprehension (Akwf, 5ibK): Our metrics assess information availability, source support, and question-based recovery as proxies for comprehension. We clarified the human and agent involvement in X+Slides probe evaluation (Fig. 2) and the 25-participant ranking study in ArcDeck/ArcBench (Appendix A.1). A direct user study of AudienceDeck across expertise levels remains a priority for future work.

Full results and experimental settings are provided in our individual responses. We will incorporate the joint-audience ArcDeck comparison, the component ablations, and the blinded evaluation of gate-guided revision into the revised manuscript. Thank you again for your time and constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.