PaperlayerView exact revision on arXiv

Source-linked AI summary

From Transient Prompts to Persistent Control: Scientific Poster Generation via Recursive Semantic-Geometric Contracts

Runze Li, Yukun Zhao, Can Xu, Yucheng Shen, Shuaiqiang Wang, Jianmin Wu, Lingyong Yan, Dawei Yin

arXiv:2609.17326v1cs.AI

TL;DR

Scientific poster generation lacks persistent, paper-specific control over semantic, evidential, and spatial requirements across stages and repairs. PosterVisor addresses this gap with Semantic–Geometric Contracts and Recursive Contract Enforcement, and its implementations outperform matched baselines, with PosterVisor-PPT preferred in 72.5% of non-tied comparisons.

  • Problem

    Existing poster-generation pipelines bind plans and checks to individual stages, so semantic and spatial requirements can drift and repairs can silently invalidate earlier checks.

  • Method

    PosterVisor uses an Orchestrator to compile paper-grounded requirements into a Semantic–Geometric Contract and applies Recursive Contract Enforcement when evidence emerges and after affected repairs.

  • Results

    PosterVisor-PPT improves Raw PaperQuiz Overall from 58.53 to 64.47 over PosterGen and receives 72.5% of non-tied human preferences on the 100-paper benchmark.

  • Takeaways & Limitations

    The results support persistent contracts, scoped repair, and rechecking as practical control mechanisms for the two poster-generation pipelines studied.

  • Takeaways & Limitations

    Regression detection is limited by the declared dependencies, and the HTML human-evaluation intervals include 50%, making that result descriptive.

Abstract

from arXiv · show

Scientific poster generation distills a multimodal paper into a single-page visual artifact, forcing strict trade-offs between informational coverage and readability under a fixed spatial budget. Existing methods pass plans as transient prompts and validate individual stages in isolation. This strategy causes requirements to drift across content and layout modules, and previous checks to be silently invalidated. We introduce PosterVisor, a control framework that shifts poster generation from transient prompts to persistent control. An Orchestrator grounds rubrics in the paper and visual assets, compiling them into a Semantic-Geometric Contract (SGC) that binds claims and sources to required visuals, budgets, and spatial commitments. Only fully instantiated records become executable assertions; other usable requirements remain soft guidance. Recursive Contract Enforcement (RCE) dynamically triggers checks across stages as evidence emerges. Crucially, during repairs, RCE rechecks affected checkpoint states, preventing repair-induced regressions from propagating silently. We instantiate PosterVisor in HTML/CSS and editable PPTX generators. On the 100-paper Paper2Poster benchmark, PosterVisor-PPT improves observed mean poster-grounded QA accuracy over PosterGen (64.47% vs. 58.53%) and is preferred by human judges in 72.5% of non-tied pairwise comparisons (95% CI, 61.6-83.4%). A secondary 30-paper study also yields higher VLM Overall and PaperQuiz means. These results support rubric-compiled contracts and stage-conditioned enforcement for controllable poster synthesis.

Introduction

PosterVisor addresses requirement drift in scientific poster generation by replacing stage-local plans and checks with persistent, paper-specific contracts and repair-aware enforcement. Its HTML/CSS and editable-PPTX implementations improve benchmark quality and human preference over matched baselines.

  • Introduction: Poster generation must jointly select claims, preserve evidence, and coordinate text, visuals, and layout under a fixed spatial budget.Unlike conventional text summarization, the task requires a legible single-page narrative with coordinated semantic and visual commitments.
  • Introduction: Existing pipelines use stage-specific plans and checks, allowing semantic and spatial requirements to weaken and earlier validations to become invalid after later repairs.Without tracking repair scope and dependencies, systems cannot determine which verification results remain valid.
  • Introduction: PosterVisor compiles paper-grounded criteria into a Semantic–Geometric Contract that binds claims and evidence to visuals, budgets, and spatial commitments.Fully instantiated requirements become executable assertions, while other useful requirements remain soft guidance.
  • Introduction: Recursive Contract Enforcement checks assertions when their evidence becomes available, then revalidates affected assertions after scoped, bounded repairs.Together, SGC and RCE provide persistent, repair-aware control across content, layout, and rendering stages.
  • Introduction: PosterVisor improves poster quality across HTML/CSS and editable-PPTX generators, with PosterVisor-PPT preferred in 72.5% of non-tied human comparisons.On 100 papers, PosterVisor-PPT improves Raw PaperQuiz from 58.53 to 64.47 and both implementations obtain higher observed VLM Overall and Raw PaperQuiz means than matched baselines.

Problem Formulation

PosterVisor formulates poster generation as maintaining paper-specific semantic and spatial requirements across editable and rendered artifacts. Its SGC compiles grounded requirements into assertions or guidance, while RCE checks available states and rechecks affected assertions after repairs.

  • Problem Formulation: Poster generation maintains a paper-specific requirement set across generated content, editable geometry, and rendered output, with requirements becoming verifiable at different stages.The formulation includes an editable artifact X, rendered poster Y, and audit log L.
  • Overview: The Evidence Reader extracts indexed text and a provenance-preserving visual inventory, after which the Orchestrator grounds rubric criteria and freezes a shared SGC.Content and layout modules then generate artifacts under the frozen contract.
  • Semantic–Geometric Contract Construction: The SGC stores thesis, panels, claims, source anchors, required visuals, budgets, assignments, and spatial commitments derived from paper evidence and format constraints.It is represented as K = (Kc, Kl, E, S).
  • Semantic–Geometric Contract Construction: Grounded rubric records contain criteria, typed target queries, required values or policies, and provenance, while normalization resolves conflicts before contract freezing.Unresolved conflicts remain plan violations.
  • Semantic–Geometric Contract Construction: Only records with resolved targets and expectations plus compatible predicate templates become executable assertions; usable unresolved requirements remain soft guidance, while unusable ones stay unresolved in the audit log.A policy key alone cannot make a requirement executable.
  • Semantic–Geometric Contract Construction: Compiled predicates return pass, fail, or unavailable, and unavailable observations are never treated as passes.This preserves a distinction between failed verification and missing checker evidence.
  • Recursive Contract Enforcement: RCE evaluates assertions at content, editable-layout, and rendered-output checkpoints, using deterministic validators or task-scoped model predicates as appropriate.Violations are routed to allowed repair operators, and affected assertions are rechecked.
  • Recursive Contract Enforcement: Repair rechecking closes over declared and observed changes, dependencies, geometry, and global guards; if differences are unavailable, all applicable checkpoint assertions are rerun.This makes repair-induced regressions detectable, though detection remains limited by dependencies, observed differences, checkers, and unmodeled properties.

Experiments

PosterVisor is evaluated on matched HTML and editable-PPTX poster generators using benchmark, cross-model, ablation, failure-signal, repair, and human-preference analyses. Across the primary and secondary evaluations, the systems improve key observed quality and paper-grounded QA measures while exposing format-specific trade-offs and added generation overhead.

  • Main Results: 64.47% versus 58.53% Raw PaperQuiz Overall for PosterVisor-PPT versus PosterGen, with 72.5% of non-tied human preferences.The preference interval is 61.6–83.4%.
  • Main Results: PosterVisor-HTML raises VLM Overall from 3.707 to 3.822 and Raw PaperQuiz Overall from 66.14 to 66.82, while PosterVisor-PPT rises from 3.460 to 3.818 and from 58.53 to 64.47.Both realizations round to the highest displayed automatic VLM Overall of 3.82; HTML leads clarity, logical flow, and aggregate Information, while PPT leads engagement and improves most strongly in Aesthetic.
  • Cross-Model Evaluation: PosterVisor-HTML improves VLM Overall and Raw PaperQuiz in all four additional model-backbone settings, while PosterVisor-PPT improves each objective in three settings.HTML gains range from 0.019–0.127 VLM points and 1.56–3.90 QA points, whereas PPT shows stronger model-dependent content–presentation trade-offs.
  • Secondary Evaluation: On the 30-paper secondary set, PosterVisor-PPT raises VLM Overall from 3.91 to 4.01 and Raw PaperQuiz Overall from 78.9 to 85.5, ranking first on both.The secondary comparison uses a separate 30-paper evaluation set and reproduces the Paper2Poster question protocol.
  • Ablation Study: The full HTML system is the only ablation exceeding P2P on both aggregate objectives, while removing any PPTX component lowers VLM Overall from 3.818 to 3.496–3.622.The ablations test cumulative HTML content, structure, and geometry enforcement and leave-one-component-out PPTX variants.
  • Failure-Mode Analysis: Failure signals fall from 31% to 8% for HTML overflow and 27% to 5% for missing figures, while PPTX severe text crowding falls from 18% to 0% and illegible figures from 14% to 2%.These are output-level failure signals measured across 100 matched cases for each format.
  • Repair Analysis: RCE rechecks affected states after repairs: eight pass-to-fail assertion transitions were observed, five eliminated before finalization, and three recorded as residual violations after the repair budget.The analysis covers 41 HTML repair stages with subsequent full checks and 75 code-level assertion transitions.
  • Generation Overhead: PosterVisor-PPT increases calls from 8.05 to 19.15 and latency from 160.32 to 505.49 seconds, but reduces total tokens from 104,251 to 67,066 and estimated token cost from $0.297 to $0.228.The comparison is per poster.

Conclusion

PosterVisor coordinates scientific content, visual evidence, and geometry through persistent SGC and RCE, with stronger observed benchmark outcomes and human preference for PosterVisor-PPT over PosterGen.

  • PosterVisor coordinates scientific content, visual evidence, and geometry through a persistent SGC and RCE.
  • Both realizations exceed their matched baselines on observed VLM Overall and Raw PaperQuiz Overall, tying for the highest displayed VLM Overall of 3.82.
  • 5.94 Raw PaperQuiz points and 72.5% of non-tied preferences over PosterGen favor PosterVisor-PPT on the primary benchmark.
  • The results support persistent contracts, scoped repair, and rechecking as practical control mechanisms for the two studied pipelines.

Supplementary Material for From Transient Prompts to Persistent Control:

The supplementary material accompanies the work titled “Scientific Poster Generation via Recursive Semantic–Geometric Contracts.”

  • The supplementary material concerns scientific poster generation.
  • Its stated approach centers on recursive semantic–geometric contracts.
  • The supplementary material is associated with the title “Scientific Poster Generation via Recursive Semantic–Geometric Contracts.”

1 Operationalizing the SGC and RCE

The supplementary operationalization freezes paper-specific contracts and demonstrates executable checks, deterministic repairs, and residual rechecks across HTML checkpoints.

  • Format-specific states, validators, and bounded repair operators specialize the frozen contract defined in the main paper.
  • An archived HTML SGC binds a warm-start claim to a required source visual, planned placement, survival checks, and readability guidance.
  • Deterministic repair restored an omitted visual reference, after which the residual recheck passed.
  • A bounded layout retry restored the frozen column target after COLUMN_COUNT_DRIFT, and the final checkpointwide recheck passed.
  • The case distinguishes executable constraints in E from soft preferences in S rather than testing every stored plan field as compliance.

2 Format-Specific Instantiations

PosterVisor instantiates the SGC and RCE in HTML/CSS and editable PPTX generators, compiling paper-specific content, geometry, and visual requirements into bounded verification and repair policies.

  • 2 Format-Specific Instantiations: HTML and PPTX plans become frozen SGCs only after normalization and assertion compilation instantiate Kc, Kl, E, and S.
  • 2.1 PosterVisor-HTML: PosterVisor-HTML uses a 650-word budget, balanced density, and at least three required visual assets on the 100-paper benchmark.
  • 2.1 PosterVisor-HTML: The HTML contract links sections, sources, budgets, positions, visual roles, spans, size constraints, readability guidance, and grounded rubric records.
  • 2.1 PosterVisor-HTML: HTML verification checks required sections, visual placeholders, word budgets, geometry, rendered bounding boxes, overlap, overflow, distortion, widths, coverage, whitespace, and figure size.
  • 2.1 PosterVisor-HTML: HTML content and editable-layout checks allow two repair rounds, while the layout–render–check loop permits three attempts; unresolved executable assertions keep compliance false.
  • 2.2 PosterVisor-PPT: The PPTX pipeline verifies normalized content, editable geometry, typography, and rendered output using deterministic and task-scoped predicates compiled into E.
  • 2.2 PosterVisor-PPT: PPTX normalization constructs five-to-eight-panel blueprints, assigns one top-middle anchor panel to the key visual, and binds claims, evidence, budgets, and layout priorities.
  • 2.2 PosterVisor-PPT: PPTX geometry uses near-equal column profiles, a 0.75–1.10 coverage band, large-visual allocation rules, and bounded actions that preserve frozen assignments.

3 Mechanism Validation

The mechanism audit finds complete post-repair recheck coverage across content, editable-layout, and rendered-output repair records, while distinguishing pre-freeze geometry mutations from post-freeze repairs. Immediate regression checks recorded all eight detected pass-to-fail transitions, resolving five before finalization and leaving three residual violations.

  • Regression Detection: Eight detected pass-to-fail regression transitions were recorded as failures immediately after repair; five were resolved and three remained at finalization.The residual violations remained after the repair budget was exhausted.
  • Scope Boundaries: The audit separates 21 pre-freeze blueprint-geometry mutations from post-freeze RCE repair records because they occur before the SGC is frozen.The 21 mutations span 18 posters and are classified as SGC construction rather than post-freeze repair.
  • Recheck Coverage: 1.0 recheck coverage was achieved for content, editable-layout, post-typography, and all 29 rendered-output critic-loop repair trajectories.Each critic-loop trajectory completed detection, repair writing, application, rerendering, and full-poster rechecking.
  • Scope Boundaries: Repair counts use different scopes: 159 HTML local repairs, 422 PPTX control records, and 217 category-specific structural repairs are not alternative totals.The counts differ in denominator and record definition, so they should not be directly compared.

4 Ablation Study and Visual Evidence

The progressive PPTX study builds PosterVisor incrementally from a PosterGen baseline through grounded contracts, pre-render checks, guarded repairs, and rendered-output enforcement. Its configurations are independently generated, while the visual ablation compares cumulative HTML controls and diagnostic PPTX leave-one-component-out variants.

  • Progressive Study: PosterVisor-PPT progresses from PosterGen through a blueprint scaffold, grounded SGC, pre-render verification and repair controls, then rendered-output RCE.The grounded blueprint is normalized, assertion-compiled, and frozen before later controls are activated.
  • Visual Evidence: Figure S1 compares cumulative HTML configurations with diagnostic PPTX leave-one-component-out configurations using same-paper visual outputs.The figure is used to inspect visual effects alongside the quantitative ablations.

P2P SGC

The visual and cross-model studies examine RCE components, presentation failures, and robustness under alternative model backbones. Their protocols define which controls, outputs, and model assignments are compared without supplying additional outcome values here.

  • Visual Ablations: Figure S1’s full configuration adds content, structure, and rendered-geometry RCE to SGC, plus visual analysis, readability-aware figure sizing, and bounded layout–render–check retries.Repairs include layout compaction and remain under the frozen SGC.
  • Cross-Model Robustness: The cross-model study uses one named model for every model-based generation and internal-control call, including planning, figure analysis, critics, and repair writing.The rule applies to Qwen3.7-Plus, GPT-5.5, Gemini-3.1-Pro-Preview, and Claude-Opus-4.8; GPT-4o is evaluation-only.
  • Output Diagnostics: HTML output comparisons review final-poster labels for overflow, missing or illegible figures, and reading-order disruption, with categories allowed to overlap.Labels originate from poster-level VLM judgments and are manually checked against final rendered posters.
  • Output Diagnostics: PPTX output comparisons apply fixed rules to paired PosterGen and PosterVisor-PPT VLM judgments for severe crowding, content gaps, and illegible figures.The rules require score thresholds plus explanations identifying the relevant defect.
  • Transfer Study: The 30-paper transfer study freezes 50 verbatim and 50 interpretive questions per paper because its source set lacks PaperQuiz questions.The transfer benchmark uses GPT-4.1 for generation and a shared evaluation setup described separately.

6 Evaluation and Human-Study Protocols

The evaluation combines final-image automatic metrics, human rankings and pairwise preferences, progressive controls, ablations, cross-model tests, and transfer results. Protocols fix model roles, questions, scoring, and clustering procedures, while the HTML preference interval remains descriptive.

  • Automatic Evaluation: Each system produces one poster per paper, with HTML and PPTX rendered by Playwright and LibreOffice and automatic metrics computed only from final poster images.The primary protocol uses the 100-paper Paper2Poster benchmark.
  • Automatic Evaluation: PaperQuiz averages image-only scores from GPT-4o, GPT-4o-mini, and o3 over 50 frozen verbatim and 50 frozen interpretive questions per paper.Missing, invalid, or unsupported answers are counted as incorrect.
  • Human Evaluation: Fifteen trained annotators completed seven-way rankings, and 13 also completed paired comparisons with concealed identities and randomized order.The seven-source subset was fixed when the study was run.
  • Statistical Analysis: Within-pipeline preference intervals use two-way cluster-robust standard errors, with crossed-cluster bootstrap sensitivity intervals and Holm correction across two tests.The clustering accounts for repeated observations of annotators and papers.
  • Ablation and Robustness: The progressive PPTX study reports control switches and aggregate VLM quality and Raw PaperQuiz Accuracy for its configurations, while the leave-one-component-out table records control events.These are separate diagnostic and result tables rather than a single evaluation measure.
  • Limitations: The HTML preference intervals include 50%, so that preference result remains descriptive rather than conventionally supported by the reported interval.Leave-one-annotator-out analysis does not attribute the PPTX preference to one annotator.

7 Extended Qualitative Results

Figures S2 and S3 compare HTML- and PPTX-oriented methods on four papers, enabling qualitative inspection of density, evidence use, readability, and spatial organization. Figure S4 additionally shows PosterHarness outputs on the same paper set.

  • 7 Extended Qualitative Results: Figures S2 and S3 compare HTML- and PPTX-oriented methods on the same four papers.The comparisons support direct qualitative inspection rather than aggregate performance estimation or component-level causal attribution.
  • 7 Extended Qualitative Results: The figures support direct inspection of density, evidence use, readability, and spatial organization.
  • 7 Extended Qualitative Results: Figure S4 additionally shows PosterHarness outputs on the same paper set, while light-blue headers identify PosterVisor outputs.
Loading 2609.17326v1…