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
cs.AI
Quick verdict
PosterVisor adds frozen rubric contracts and repair-aware rechecking to LLM poster generation; gains are real but modest, with ablation trade-offs.
Summary
The paper presents PosterVisor, a control framework for automatic scientific poster generation built on two mechanisms. A Semantic-Geometric Contract (SGC) compiles paper-grounded rubric criteria into a frozen specification that binds claims and source evidence to required visuals, text budgets, and spatial commitments; only fully instantiated, checkable records become executable assertions, while the rest remain soft guidance. Recursive Contract Enforcement (RCE) evaluates these assertions at content, layout, and rendering checkpoints, routes violations to bounded repairs, and via a dependency closure rechecks assertions affected by each repair, recording repair-induced regressions. The authors instantiate the framework twice: an HTML/CSS generator derived from P2P and an editable-PPTX generator derived from PosterGen. On the 100-paper Paper2Poster benchmark with GPT-4o, PosterVisor-PPT raises raw PaperQuiz accuracy from 58.53 to 64.47 and VLM Overall from 3.46 to 3.82, and is preferred in 72.5% of non-tied blind human pairwise comparisons; PosterVisor-HTML marginally improves over P2P and ties for the top VLM Overall (3.82). A 30-paper transfer study and four additional backbone models support generalization. Limitations: automatic gains are small and reported without significance tests; density-augmented QA ranks the PPT variant below several baselines; ablations are non-monotonic (SGC alone hurts QA); no code release is stated; and PPTX latency triples to about 505 seconds per poster.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Grounding a fixed rubric catalog into a frozen Semantic-Geometric Contract limits drift of semantic and spatial commitments across the content, layout, and rendering stages.
- Recursive Contract Enforcement detects regressions introduced by repairs through dependency-closure rechecking, evidenced by 8 detected pass-to-fail transitions and 1.0 recheck coverage in archived PPTX logs.
- On the 100-paper Paper2Poster benchmark, PosterVisor-PPT improves raw PaperQuiz accuracy over PosterGen (64.47 vs 58.53) and wins 72.5% of non-tied human pairwise comparisons, while PosterVisor-HTML improves over P2P on both automatic objectives; both tie for the highest automatic VLM Overall (3.82).
- The control abstraction transfers across four additional model backbones and a 30-paper secondary benchmark without GPT-4o in the generation loop.
Strengths
The contract formulation cleanly separates enforceable from aspirational requirements
The compiler emits an executable assertion only when target, expectation, and a predicate template from the format policy all resolve; everything else stays as soft guidance or is logged as unresolved (Method, SGC construction; worked archived instance in Table S1). This prevents uncheckable aesthetic goals from being counted as compliance.
Repair-aware revalidation is demonstrated with archived evidence, not just asserted
All 159 deterministic HTML repairs are followed by rule-based rechecks; the regression audit finds 8 pass-to-fail transitions across 8 posters, 5 resolved within budget and 3 recorded as residual violations. The PPTX audit reports 1.0 recheck coverage (29/29 critic-loop trajectories; 422 control events in Table S2), and Table 4 shows auditable repair trails.
Evaluation rigor is above the norm for this subfield
Transparent mechanism accounting and honest trade-off disclosure
Unusually detailed, checkable implementation specification
Weaknesses
Automatic-metric gains are small and uncertainty is unreported
Density-augmented results undercut the PPT accuracy headline
Ablations are non-monotonic and component attribution is weak
The regression-safety claim is fully verified on only a subset of repairs
Human evaluation omits the strongest recent baselines and the HTML result is a tie
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PosterVisor and what problem does it solve?
PosterVisor is a framework for automatically turning a scientific paper into a presentation poster. Instead of handing each pipeline stage a one-off prompt, it compiles paper-grounded quality criteria into a persistent contract that binds claims, evidence, figures, word budgets, and space, and it re-validates earlier checks whenever a later repair changes the poster, so fixes do not silently break previously satisfied requirements.
How much better is PosterVisor than existing poster generators?
On the 100-paper Paper2Poster benchmark using GPT-4o, the PPTX version raises factual question-answering accuracy from 58.53% to 64.47% over its base system PosterGen and is preferred in 72.5% of non-tied blind human comparisons; the HTML version slightly beats its base P2P, and both versions tie for the best judged overall quality (3.82 of 5). The HTML version's human-preference win over P2P was a statistical tie (54.7%).
What do SGC and RCE mean in this paper?
SGC (Semantic-Geometric Contract) is a frozen specification recording which figures a poster must include, text budgets, and spatial commitments; only requirements that can actually be measured become enforced assertions, while the rest remain soft guidance. RCE (Recursive Contract Enforcement) checks those assertions at each pipeline stage, routes failures to bounded repairs, and re-checks anything a repair could have broken, recording any regressions it finds.
What are the main limitations of PosterVisor?
Automatic-metric gains are modest and reported without significance tests; after text-density correction the PPTX version's question-answering score trails several baselines; removing some components actually improves accuracy, so gains come from the whole stack rather than single parts; generation latency roughly triples (about 505 seconds per poster for the PPTX version); and the paper states no code release. Factual recoverability gains are strongest for the PPTX variant.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 and S1-S4 are images not analyzable from the extracted text; qualitative evidence was assessed from captions and surrounding prose only.
- Tables were extracted as dense inline text; numbers were cross-checked against the prose, but minor transcription risk remains.
- No code or data availability statement appears in the extracted text; if one exists outside the extracted sections, codeAvailable may be understated.
- The exact wording of the rubric catalog R and of the judge prompts is not present in the text, limiting assessment of metric validity.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.