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.AISep 15, 2026

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

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

Organizations: Baidu Inc., Beijing, China · Harbin University of Science and Technology, Harbin, China · Shandong University, Jinan, China · East China Normal University, Shanghai, China ·

Abstract

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.

Figures & tables

Figure 1: Overview of PosterVisor . The Evidence Reader extracts traceable paper evidence, and the Orchestrator grounds rubric dimensions to construct a persistent SGC. Content and layout modules generate artifacts under the frozen contract. RCE evaluates assertions as state becomes available, routes violations to bounded repairs, and rechecks affected assertions.

Method	VLM-as-Judge ( $\uparrow$ )									PaperQuiz ( $\uparrow$ )
	Aesthetic				Information				Overall	Raw Accuracy			Density-Augmented
	Elem.	Layout	Engage.	Avg.	Clarity	Content	Logic	Avg.	Overall	Verb.	Interp.	Overall	Verb.	Interp.	Overall
Paper †	4.05	3.89	2.80	3.58	4.00	4.68	3.98	4.22	3.90	84.41	79.69	82.05	92.97	87.75	90.36
GT Poster †	4.07	3.90	2.70	3.56	4.09	3.96	3.89	3.98	3.77	68.25	74.73	71.49	127.75	140.29	134.02
PosterAgent ‡	3.92	3.05	2.98	3.32	3.98	3.97	3.55	3.83	3.58	57.81	76.39	67.10	111.40	147.29	129.34
PosterGen ⋆	3.59	2.98	2.81	3.13	4.14	3.93	3.32	3.80	3.46	43.90	73.15	58.53	87.71	146.19	116.95

Table 1: Results on the 100-paper Paper2Poster benchmark.

\dagger

marks values reported by Paper2Poster,

\ddagger

marks our LaTeX PosterAgent reproduction, and

\star

marks our other reproductions. Paper and GT Poster are reference inputs excluded from automatic-method ranking. Bold and underline indicate the best and second-best automatic results; rounded ties share a mark.

Figure 2: Qualitative comparison for one Paper2Poster paper across the author poster, seven automatic baselines, and the two PosterVisor realizations.

Configuration	VLM-as-Judge ( $\uparrow$ )			Raw PaperQuiz ( $\uparrow$ )
Configuration	Aes.	Info.	All	Verb.	Interp.	All
P2P (base)	3.500	3.913	3.707	55.67	76.60	66.14
SGC	3.530	3.900	3.715	40.42	77.78	59.10
$+$ C/S	3.562	3.938	3.729	45.28	76.38	60.83
$+$ C/S/G	3.657	3.980	3.818	49.74	69.76	59.75
Full	3.380	4.263	3.822	58.23	75.40	66.82

Table 2: HTML ablations on 100 papers. C, S, and G denote content-, structure-, and geometry-level RCE; “

+

” indicates cumulative addition to SGC. Each configuration is generated independently; the base and full-system results match Table 1 .

Configuration	VLM	QA	Vis.	Struct.
	Overall	Overall	demot.	repair
Full system	3.818	64.47	27	217
w/o Evidence Grounding	3.537	64.60	18	162
w/o Visual Constraint	3.622	66.40	33	179
w/o Geometry Verifier	3.573	66.50	0	200
w/o Repair Guard	3.496	67.30	29	6

Table 3: Diagnostic leave-one-component-out ablations for the PPTX realization. VLM and QA are Overall scores; event columns report totals over each run.

Case	Violation and scoped repair	Recheck
H01	Three required visuals missing $\rightarrow$ restore affected placeholders	Content and global checks pass
H02	Required visual lacks two-column span $\rightarrow$ local layout repair and rerender	Violations: $2\!\to\!1\!\to\!0$
U01	Four-column contract, three-column output $\rightarrow$ two bounded local retries	Unresolved and logged

Table 4: Selected auditable HTML repair trails. These examples illustrate RCE behavior rather than aggregate repair reliability.

(a) Ranking over seven poster sources
Method	Top-1	Top-2	Top-3	Score ( $\uparrow$ )	Rank ( $\downarrow$ )
Author Poster	341	198	186	1605	3.18
P2P	164	225	209	1151	3.73
PosterGen	95	128	199	740	4.30
PosterAgent	59	98	131	504	4.68
PPTAgent	64	62	55	371	4.93

Table 5: Human evaluation: seven-source ranking (top) and matched-pipeline pairwise outcomes (bottom), with cluster-robust 95% CIs.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Grounded requirement	Saved contract realization	Operational class	Validator and archived evidence
Warm-start claim and source visual	$K_{c}$ : claim-bearing section and required source visual; $K_{l}$ : assigned column and multi-column span	$K_{c}/K_{l}$ contract fields; the linked requirement is executable only with a validator binding	Fields are retained in the frozen contract and supplied to content and layout generation
Required visual survives writing	The required visual must appear as a valid numeric Markdown placeholder after content generation	$\mathcal{E}$ : executable content assertion	MISSING_FIGURE_PLACEHOLDER ; failed before repair and passed after deterministic reference restoration
Required visual survives layout	Editable HTML must retain the corresponding image reference before source-asset embedding	$\mathcal{E}$ : executable layout assertion	MISSING_VISUAL_REF ; passed at the archived layout checkpoint
Readable, logically placed visual	Keep trends and annotations visible; place figures logically to support the storyline	$\mathcal{S}$ : soft guidance	Retained as generation guidance unless compilation produces a supported measurable predicate

Appendix

Table S1: One archived paper-specific SGC instance. The

\mathcal{E}/\mathcal{S}

assignments and their validator bindings are read from the saved frozen-contract record; checkpoint artifacts provide the observed pass/fail evidence.

Control stage	Events	Mechanism role
Candidate-blueprint normalization	21	Pre-freeze SGC construction
Content-checkpoint RCE	153	Content repair and recheck
Editable-layout RCE	121	Geometry repair and recheck
Post-typography RCE	24	Styled-geometry repair and recheck
Rendered-output RCE	103	Guarded repair, rerendering, and recheck
Total	422

Appendix

Table S2: PPTX control events observed in the main-benchmark GPT-4o PosterVisor -PPT runs over 100 papers. Counts are per-event repair records, not violation counts or state transitions; one record may bundle multiple violation codes. The 21 pre-freeze blueprint-geometry mutations are excluded from the 401 post-freeze records by definition.

Layer	Added	EG	VC	GV	RG	RR
L1 PosterGen	baseline	–	–	–	–	–
L2 blueprint scaffold	scaffold	$\times$	$\times$	$\times$	$\times$	$\times$
L3 grounded SGC	EG	✓	$\times$	$\times$	$\times$	$\times$
L4 pre-render	VC/GV/RG	✓	✓	✓	✓	$\times$
L5 full	rendered RCE	✓	✓	✓	✓	✓

Appendix

Table S3: Switches in the progressive PosterVisor -PPT study. EG: Evidence Grounding; VC: Visual Constraint; GV: Geometry Verifier, including pre-render layout verification; RG: Repair Guard; RR: rendered-output RCE.

Configuration	VLM-as-Judge ( $\uparrow$ )			Raw PaperQuiz Accuracy ( $\uparrow$ )
Configuration	Aesth.	Info.	Overall	Verbatim	Interpretive	Overall
L1 PosterGen	3.130	3.800	3.460	43.90	73.15	58.53
L2 blueprint scaffold	3.040	4.100	3.570	45.17	76.10	60.64
L3 + grounded SGC	3.060	4.030	3.540	48.91	75.49	62.20
L4 + pre-render controls	3.100	4.050	3.580	47.09	74.70	60.89
L5 full PosterVisor -PPT	3.607	4.030	3.818	53.02	75.91	64.47

Appendix

Table S4: Complete progressive PosterVisor -PPT results on the main GPT-4o benchmark. Because configurations are independently generated, adjacent rows describe configuration-level transitions rather than isolated marginal effects.

Configuration	VLM-as-Judge			Raw PaperQuiz Accuracy
Configuration	Aesth.	Info.	Overall	Verbatim	Interpretive	Overall
full	3.607	4.030	3.818	53.02	75.91	64.47
w/o Visual Constraint	3.080	4.163	3.622	60.80	72.00	66.40
w/o Geometry Verifier	3.057	4.090	3.573	61.50	71.50	66.50
w/o Evidence Grounding	3.057	4.017	3.537	58.80	70.30	64.60
w/o Repair Guard	3.016	3.975	3.496	62.40	72.20	67.30

Appendix

Table S5: Diagnostic PosterVisor -PPT leave-one-component-out configurations and recorded control events. VLM quality and Raw PaperQuiz Accuracy are aggregate results for each configuration. Whenever the geometry-gate node is instantiated, it emits a report file; when GV is disabled in the leave-one-out ablation, that file is a status="disabled" stub with no actions. Active geometry reports therefore count only status="ok" records, rather than files. In the focal full versus w/o GV comparison, the active gate produced 21 geometry mutations across 18 posters, versus zero mutations when GV was disabled. Visual demotions and structural repairs are event totals rather than poster counts or repair success rates. EG, VC, GV, and RG follow the component names defined in Table S3 .

Panel of Figure 14 — Figure S1: Same-paper visual ablations. In (a), C, S, and G denote content-, structure-, and rendered-geometry-level RCE, respectively. The $+$ C/S/G configuration adds all three RCE levels to SGC; Full additionally enables visual analysis, readability-aware figure-size enforcement, and bounded layout–render–check retries with layout compaction, all performed under the frozen SGC. Callouts mark undersized or unreadable figures, unused space, crowded layouts, column imbalance, and misplaced visual priority across the HTML and PPTX configurations.

Qwen3.7-Plus
Method	VLM-as-Judge ( $\uparrow$ )									PaperQuiz ( $\uparrow$ )
	Aesthetic				Information				Overall	Raw PaperQuiz Accuracy			Density-Augmented PaperQuiz
	Elem.	Layout	Engage.	Avg.	Clarity	Content	Logic	Avg.	Overall	Verbatim	Interpretive	Overall	Verbatim	Interpretive	Overall
GT Poster	4.07	3.90	2.70	3.56	4.09	3.96	3.89	3.98	3.77	68.25	74.73	71.49	127.75	140.29	134.02
PosterAgent	2.92	3.01	1.76	2.56	2.07	3.89	4.90	3.62	3.09	60.53	72.91	66.72	114.21	137.90	126.06
PosterGen	3.53	3.62	2.38	3.18	4.71	3.90	4.94	4.52	3.85	53.53	72.97	63.25	107.07	145.93	126.50

Appendix

Table S6: Cross-model robustness results. In each setting, the named model is used for every model-based call in the generation and internal-control pipeline. GPT-4o does not participate in poster generation, internal checking, or repair; during external evaluation, it serves as the VLM-as-Judge and as one member of the common PaperQuiz reader ensemble. The complete GT Poster row reuses the main-comparison reference values. Bold and underlined values denote the best and second-best automatic methods within each setting; rounded ties share the same mark.

Method	VLM-as-Judge ( $\uparrow$ )									PaperQuiz ( $\uparrow$ )
	Aesthetic				Information				Overall	Raw PaperQuiz Accuracy			Density-Augmented PaperQuiz
	Elem.	Layout	Engage.	Avg.	Clarity	Content	Logic	Avg.	Overall	Verbatim	Interpretive	Overall	Verbatim	Interpretive	Overall
GT Poster	3.97	3.73	3.10	3.60	4.40	4.00	4.37	4.26	3.93	71.27	87.93	79.60	135.67	167.80	151.73
PosterAgent	4.03	3.70	3.13	3.62	3.90	4.00	4.23	4.04	3.83	70.07	90.00	80.03	134.12	172.36	153.24
PosterGen	4.03	3.60	3.07	3.57	4.53	4.00	4.23	4.26	3.91	65.60	92.27	78.93	130.44	183.45	156.94
PPTAgent	2.07	3.40	2.53	2.67	4.33	3.07	3.67	3.69	3.18	24.80	46.87	35.83	49.60	93.73	71.67

Appendix

Table S7: Transfer results on the 30-paper PosterGen evaluation set. Automatic posters use GPT-4.1 for generation; all VLM-as-Judge scores in this table are produced by GPT-4o using the same six criteria as in the main experiment. PaperQuiz uses the shared three-reader image-only ensemble of GPT-4o, GPT-4o-mini, and o3; scores are averaged across readers per poster and then across papers. We report Raw PaperQuiz Accuracy and per-poster Density-Augmented PaperQuiz scores using the fixed reference

w=1335.0

. The GT Poster is included only as a reference and is excluded from automatic-method ranking. Bold and underlined values denote the best and second-best automatic results; rounded ties share the same mark.

Comparison	Holm $p$	Bootstrap CI	LOO range
PPT vs. PosterGen	.0015	62.5–82.4%	70.6–74.6%
HTML vs. P2P	.278	46.3–63.7%	53.0–57.2%

Appendix

Table S8: Sensitivity analyses for pairwise preference. Holm

p

adjusts both tests; bootstrap CIs cross-cluster by annotator and paper.

Rubric-Guided Generation

Rubric-guided generation uses explicit, task-specific criteria to steer, score, and improve model outputs when quality is hard to measure automatically.

See all 42 papers

PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs
Aug 3, 2026Haojie Hu, Chenhao Dang, Yaojia Liu +3Multi-Agent LLM Systems AI-Assisted Scientific Research
Tsinghua Shenzhen International Graduate School, Tsinghua University · Tongji University · Shanghai Jiao Tong University +2
PosterHarness: Turning Scientific Poster Generation into an Auditable Instruction-Following Benchmark
Jul 3, 2026Tianyi Yang, Dawei Fu, Youpeng Wu +5AI-Assisted Scientific Research Image Generation
School of Physics, Peking University
EfficientPosterGen: Semantic-aware Efficient Poster Generation via Token Compression and Accurate Violation Detection
Feb 25, 2026Wenxin Tang, Jingyu Xiao, Yanpei Gong +6Multimodal Token Compression
Tsinghua University, Beijing, China · The Chinese University of Hong Kong, Hong Kong, China · Harbin Institute of Technology, Harbin, China +4

Aug 3, 2026J/K move · Enter open · S save

PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs

Haojie Hu, Chenhao Dang, Yaojia Liu, Hengrui Kang+2

Tsinghua Shenzhen International Graduate School, Tsinghua University · Tongji University · Shanghai Jiao Tong University+2

Multi-Agent LLM Systems AI-Assisted Scientific Research Multimodal Generation

Open paper AskPDF arXiv

Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. Existing systems hide request-level failures by scoring only completed outputs; direct image generation is not element-editable, while coding-agent workflows are costly. PosterMELD is a template-conditioned multi-agent pipeline: capacity-aware slots guide writing before rendering, and deterministic gates plus vision-language model (VLM) review route failures to bounded repair. Each accepted request exports editable PowerPoint (PPTX) and Portable Network Graphics (PNG) artifacts; explicit design controls yield same-paper variants. Across 621 papers, Print-Ready Rate (PRR) counts requests passing geometric, readability, asset-integrity, and obvious-factual-error checks, with native editability reported separately. A frozen VLM assigns conditional Craftsmanship-Harmony-Expressiveness (CHE) scores to print-ready outputs. PosterMELD attains 81.3% PRR, 3.4 times P2P's rate and 5.2 times PosterGen's, and the highest conditional CHE among generated methods with multiple print-ready outputs. Native editability and explicit design controls are retained at a mean cost of USD 0.38 per request, 3.5% of Codex+Skill's. Code and resources are available at https://github.com/Shannon4Science/PosterMELD.

Abstract

Rubric-Guided Generation

Explore similar work

PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs

PosterHarness: Turning Scientific Poster Generation into an Auditable Instruction-Following Benchmark

EfficientPosterGen: Semantic-aware Efficient Poster Generation via Token Compression and Accurate Violation Detection