acceptodds
Under review as a conference paper at ICLR 2027

VASP: Visual-anchored Agent for Scientific Poster Generation

10 views

Abstract

Scientific poster generation condenses a paper into a single page, where visual evidence must clearly communicate its central claims. Existing approaches exhibit two mismatches: a layout mismatch, in which text arrangements or templates constrain visual evidence regardless of its scientific importance, and an execution mismatch, in which models estimate geometry that programs can compute directly. Consequently, key evidence can be compressed or incompletely extracted, while refinement focused on space utilization leaves its association with the claims it supports unchecked. We address both mismatches with VASP, a visual-first framework that combines semantic reasoning with precise execution. Vision-Centric Anchoring makes key visual evidence the basis of layout construction, with its scientific role guiding spatial allocation and the arrangement of related claims. Semantic-Geometric Decoupling assigns scientific interpretation and design decisions to LLMs, while programs compute figure boundaries and check spatial constraints derived from explicitly declared claim–evidence relations. This division supports intact evidence extraction and iterative refinement guided by the intended scientific organization. We further introduce two criterion-level benchmarks that assess visual presentation and author-informed key-information coverage. On the 100-paper Paper2Poster evaluation set, VASP achieves the highest visual-presentation score (94.3 vs. 88.1 for the strongest existing approach) and Paper2Poster VLM-judge score (3.71 vs. 3.52). It also achieves 94.3 mAP in figure extraction, compared with 84.1 for ResearchStudio's agent loop, with complete crops for all 100 sampled figures, and preserves more author-informed key information than existing fixed-canvas approaches.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Scientific Poster Generation, Multimodal Agents, Visual-First Design, Semantic-Geometric Decoupling, Fine-Grained Evaluation

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

VASP is a framework for generating scientific posters with two components: SemParser and VisComposer. SemParser combines language-model judgments about document semantics with deterministic PDF geometry to extract complete figures. VisComposer organizes content around visual evidence using explicit claim–evidence associations, geometric checks, and iterative refinement. The evaluation compares VASP with six existing methods on 100 papers and examines figure extraction, visual presentation, information coverage, and runtime costs.

Critical strengths
  1. The design addresses practical problems, and the figure extraction results are well supported.
  2. The experiments demonstrate a useful trade-off between quality and cost.
Critical weaknesses
  1. The ablations do not fully isolate the proposed design choices, and the gains in scientific content organization need a more careful interpretation. In Table 6, replacing visually anchored composition with a fixed template changes both visual prioritization and layout flexibility. The drop from 91.4 to 80.8 therefore cannot be attributed specifically to the visual-first strategy. A stronger control would retain the same assets, flexible layout, and refinement budget while changing only the planning priority. Removing semantic verification reduces the overall score from 91.4 to 81.1, but Assoc. remains at 100, with most of the decline coming from Space and Density. This supports the usefulness of programmatic layout feedback, but does not isolate the contribution of claim–evidence checks from ordinary geometric checks. Since the ablation covers only 20 papers, possible saturation of the association metric also deserves discussion. Content gains under the same backbone are modest. PaperQuiz improves from the strongest baseline’s 69.25 to 69.96. Overall coverage reaches 82.0 versus PosterGen’s 79.7, but Results coverage is lower, at 71.7 versus 74.7, as is Full coverage, at 20% versus 22%. These results support some average improvement, while broader claims about scientific communication need qualification. The stronger content scores with Opus-5 should also be interpreted alongside baselines using that backbone.
  2. The new evaluation has a specific scoring concern and lacks independent calibration of poster quality. Appendix D.2 assigns N/A to EN05 when a poster has no substantive branding or decoration. However, the formula in D.3 retains a fixed denominator and awards credit only for PASS. Under this definition, a poster without these elements loses approximately 6.67 overall points. This may penalize a reasonable design choice. The authors should report N/A frequencies by method and examine whether rankings and margins remain stable when genuinely inapplicable criteria are excluded. The new evaluations rely primarily on GPT-5.6-Sol. The manual crop checks do not calibrate judgments of poster quality or information coverage. Extracting seven reference units from the author poster, auditing them against the paper, and freezing them before evaluation is a useful procedure. Still, this measures retention of selected information. Independent human assessment on a subset would help establish whether the visual scores and coverage judgments agree with readers’ assessments of readability and content.
Questions for authors
  1. Can visual prioritization, claim–evidence checks, and ordinary geometric checks be isolated while keeping assets, layout flexibility, and refinement budgets fixed?
  2. How frequently is N/A assigned to EN05 and other criteria for each method? Are the main rankings and margins stable under appropriate N/A handling?
  3. Can the authors provide human assessment on a subset, or agreement results using an independent evaluation model?
  4. Can the smaller PaperQuiz and coverage gains be accompanied by confidence intervals paired by paper? If the Opus-5 results are central to the claims, can a representative baseline also be evaluated with that backbone?
9 comments
AuthorsOpenReview ↗

Response to Reviewer qPYK

We sincerely thank Reviewer qPYK for the detailed and technically demanding review. We are glad the reviewer finds that our design addresses practical problems with well-supported figure extraction results, and that our experiments demonstrate a useful trade-off between quality and cost. We address every weakness and question below.

W.1 The ablations do not fully isolate the proposed design choices, and the gains in scientific content organization need a more careful interpretation.

W.1.1 Isolating visual-first priority and claim-evidence vs. geometric checks

In Table 6, replacing visually anchored composition with a fixed template changed both planning priority and layout flexibility. On the 20-paper subset, we built the exact control the reviewer suggests: a text-first variant that keeps the same extracted assets, template-free flexible CSS grid, and two-round refinement budget as VASP (full), changing only the planning priority so section prose is drafted and budgeted first and figures are placed into the remaining space.

For verification in Table 6, two factors obscured the separate roles of semantic and geometric checks. First, Assoc. (AL05) is a layout-proximity check, not a semantic one: on the rendered poster, it tests whether labels such as headings, captions, and legends sit visually next to the elements they belong to, rather than floating apart. Removing SV changes what a section's text says, not where headings or captions are placed, so Assoc. stays at 95%–100% even when the body text never mentions the figure next to it. To directly measure text-to-figure semantic alignment, we evaluate Claim-Evidence Grounding (CEG), a blind VLM check on the rendered poster testing whether each displayed figure or table is substantively explained by its section's body text. Second, in the submitted Table 6, the w/o semantic verification row disabled the revision controller—turning off both iterative SV and ordinary geometric/whitespace verification (GV), while retaining initial (round-0) SV planning. Below we vary Composition Priority, Initial SV, Iterative SV, and GV separately on the 20-paper subset:

\hline & & \text{Figure (\\%)} & & \text{Layout (\\%)} & & & \text{Area (\\%)} & \text{Semantics (\\%)} & \\\\ \hline \text{Variant} & \text{Overall}\uparrow & \text{Fit}\uparrow & \text{Text}\uparrow & \text{Assoc.}\uparrow & \text{Space}\uparrow & \text{Density}\uparrow & \text{Largest}\uparrow & \text{CEG}\uparrow & \text{Coverage}\uparrow \\\\ \hline \text{Composition priority (with full SV and GV enabled)} & & & & & & & & & \\\\ \mathbf{VASP}\\ \text{(full: visual-first)} & \mathbf{91.4} & \mathbf{80} & \mathbf{65} & \mathbf{100} & \underline{80} & \mathbf{90} & \underline{7.3} & 82.3 & 75.7 \\\\ \text{w/o anchored comp. (template)} & 80.8 & 30 & 30 & \mathbf{100} & \mathbf{95} & 65 & 4.5 & 75.8 & \underline{81.4} \\\\ \text{w/o anchored comp. (text-first)} & 72.8 & 25 & 15 & \mathbf{100} & 75 & 30 & 2.7 & \underline{87.0} & \mathbf{89.3} \\\\ \hline \text{Verification decoupling (with visual-first composition fixed)} & & & & & & & & & \\\\ \text{w/o iterative SV (init. SV + GV)} & \underline{90.6} & \mathbf{80} & \mathbf{65} & \mathbf{100} & 75 & \underline{85} & 6.9 & 78.2 & 77.1 \\\\ \text{w/o all SV (GV only)} & 89.0 & 60 & \underline{60} & \mathbf{100} & \mathbf{95} & \mathbf{90} & \mathbf{8.9} & 25.0 & 67.9 \\\\ \text{w/o GV (init. + iter. SV)} & 80.1 & 65 & 50 & 95 & 15 & 65 & 7.0 & \mathbf{88.1} & 80.0 \\\\ \text{w/o iter. SV \\& GV (Table 6)}^{\dagger} & 81.1 & \underline{70} & \underline{60} & \mathbf{100} & 10 & 65 & 7.0 & 72.1 & 77.1 \\\\ \hline \end{array}$$ *$^{\dagger}$Submitted Table 6 row, which retained initial SV planning while disabling iterative SV and GV during revisions.* Every factor separates cleanly. **Visual-first priority governs figure scale and layout balance**: with assets, flexible layout, and refinement budget fixed, switching from visual-first to text-first drops Overall from 91.4 to 72.8, as unconstrained prose crowds figures into leftover space (*Largest* figure area 7.3% → 2.7%, *Fit* 80% → 25%, in-figure *Text* 65% → 15%, *Density* 90% → 30%). **GV governs whitespace**: disabling GV drops *Space* to 15% and 10%, whereas keeping GV holds *Space* at 75%–95%, explaining the 80% to 10% drop in Table 6. **Initial SV establishes claim-evidence grounding**: with GV fixed, removing initial SV as well collapses CEG from 78.2% to 25.0% and Coverage from 77.1% to 67.9%. **Iterative SV preserves grounding across revisions**: with GV enabled, removing iterative SV lets geometry-only edits trim figure-explaining prose (82.3% to 78.2% CEG), whereas without GV, adding iterative SV lifts CEG from 72.1% to 88.1%. We will expand these decoupled ablations to all 100 papers in the revision.
AuthorsOpenReview ↗

Response to Reviewer qPYK (continued)

W.1.2 Interpreting content gains and qualifying claims on scientific communication

Trade-off on Table 4. On a fixed single-page canvas, packing more text trades off directly against visual legibility. Under GPT-5.6-Sol, PosterGen scores slightly higher on Results coverage (74.7% vs. 71.7%) and Full coverage (22% vs. 20%), but its Visual Presentation drops to 65.6% compared with 94.3% for VASP (Table 3). Conversely, ResearchStudio keeps Visual Presentation high (88.1%) by omitting content, dropping Overall Coverage to 73.0% compared with 82.0% for VASP. VASP is the only fixed-canvas method that remains strong on both visual presentation and content coverage.

Modest margins on PaperQuiz. A single-page poster has limited space and is meant to help readers grasp the core work quickly, rather than retain every detail of the paper. PaperQuiz instead samples questions across the entire paper. As a result, even author-designed official posters score only 67.55% on PaperQuiz (despite reaching 96.4% on Table 4), below text-heavy Paper2Poster (69.25%) and unbounded-height P2P (85.20%). Paper-paired confidence intervals for PaperQuiz and Table 4 are provided under Q.4.

Qualifying claims on scientific communication. We agree that our benchmarks do not directly test human reader comprehension. Instead, they measure visual legibility, figure preservation, claim-evidence layout, and retention of key takeaways from the author poster. We will update the manuscript as follows:

  • Abstract and Introduction. Scope "communicating central claims" to visual legibility, claim-evidence layout, and key-information retention.
  • Results (§ 5, Table 4). Explicitly note PosterGen's higher Results (74.7% vs. 71.7%) and Full coverage (22% vs. 20%) alongside its 28.7-point drop in Visual Presentation.
  • Limitations (Appendix). Note that direct reader-comprehension user studies are not evaluated.

W.1.3 Baselines under Claude Opus-5 and Qwen3.8-Flash-Next

To interpret the Opus-5 scores alongside baselines using the same backbone, we evaluated VASP with ResearchStudio (strongest visual baseline), PosterGen (strongest coverage baseline), and Paper2Poster across all 100 papers under both Claude Opus-5 and Qwen3.8-Flash-Next, keeping the evaluation protocol unchanged:

\hline \text{Method} & \text{VLM-as-Judge} & \text{PaperQuiz} & \text{Visual Presentation} & \text{Key-Information} \\\\ & \text{(Table 1)}\uparrow & \text{(\\%)}\uparrow & \text{(\\%)}\uparrow & \text{Coverage (\\%)}\uparrow \\\\ \hline \text{Claude Opus-5 backbone} & & & & \\\\ \mathbf{VASP} & \mathbf{3.70} & \mathbf{78.08} & \mathbf{92.5} & \mathbf{88.7} \\\\ \text{ResearchStudio} & \underline{3.58} & 71.77 & \underline{88.7} & 77.9 \\\\ \text{PosterGen} & 3.00 & \underline{76.05} & 65.9 & \underline{86.6} \\\\ \text{Paper2Poster} & 2.76 & 74.55 & 44.0 & 78.0 \\\\ \hline \text{Qwen3.8-Flash-Next backbone} & & & & \\\\ \mathbf{VASP} & \mathbf{3.58} & \mathbf{81.35} & \mathbf{89.2} & \mathbf{89.3} \\\\ \text{ResearchStudio} & \underline{3.54} & \underline{73.68} & \mathbf{89.2} & 80.4 \\\\ \text{PosterGen} & 3.04 & 73.28 & 65.7 & \underline{82.6} \\\\ \text{Paper2Poster} & 2.42 & 73.16 & 46.7 & 74.7 \\\\ \hline \text{Human reference} & & & & \\\\ \text{Official posters} & 3.12 & 67.55 & 83.0 & 96.4 \\\\ \hline \end{array}$$ When baselines use Claude Opus-5 as well, VASP leads on all four benchmarks at once: it exceeds PosterGen on both PaperQuiz (78.08% vs. 76.05%) and Coverage (88.7% vs. 86.6%) with a 26.6-point margin on Visual Presentation (92.5% vs. 65.9%), and exceeds ResearchStudio by 6.31 points on PaperQuiz and 10.8 points on Coverage. The same ordering holds under Qwen3.8-Flash-Next.
AuthorsOpenReview ↗

Response to Reviewer qPYK

W.2 Scoring of N/A on EN05 under a fixed denominator, and lack of independent calibration of poster quality.

W.2.1 N/A frequency by method and ranking stability excluding N/A

Across all 900 posters (9 methods × 100 papers × 12 criteria = 10,800 judgments), 7 of the 12 criteria have zero N/A verdicts on every poster. Almost all N/A verdicts occur on EN05 (subordinate branding/decoration in Visual Guidance, triggered on minimalist posters without decorative elements), with the remaining 0%–0.8% occurring on the three Figure Presentation criteria when a poster displays zero figures (plus 2 official posters without a conclusion block on EN07). Under Excl. N/A, for each poster p and group g, we replace the fixed group size ∣𝒬g∣∈3,5,4\lvert𝒬_{g}\rvert \in \\{3, 5, 4\\} in Eq. 6 (Appendix D.3) with the number of applicable (non-N/A) criteria ∣𝒬g,papp∣\lvert𝒬_{g,p}^{\mathrm{app}}\rvert on that poster displays no figures at all), and average the three groups equally:

\hline & \text{N/A Rate (\\%)} & & & \text{Visual Guidance} & & & \text{Overall} & \\\\ \hline \text{Method} & \text{EN05} & \text{Total} & \text{Figure Pres.}\uparrow & \text{Paper}\uparrow & \text{Excl. N/A}\uparrow & \text{Layout}\uparrow & \text{Paper}\uparrow & \text{Excl. N/A}\uparrow \\\\ \hline \text{Ours} & & & & & & & & \\\\ \mathbf{VASP}\\ \text{(Opus-5)} & 5.0 & 0.7 & \underline{81.3} & \underline{98.4} & \underline{99.4} & \mathbf{97.8} & \underline{92.5} & \underline{92.8} \\\\ \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & 0.0 & 0.0 & \mathbf{88.0} & \mathbf{100.0} & \mathbf{100.0} & \underline{95.0} & \mathbf{94.3} & \mathbf{94.3} \\\\ \hline \text{Baseline methods} & & & & & & & & \\\\ \text{Paper2Poster} & 84.0 & 7.8 & 35.0 & 61.6 & 73.9 & 43.0 & 46.5 & 50.6 \\\\ \text{PosterMELD} & 40.0 & 3.8 & 71.7 & 80.6 & 87.6 & 46.5 & 66.3 & 68.6 \\\\ \text{PosterForest} & 91.0 & 8.3 & 38.0 & 56.4 & 68.8 & 42.0 & 45.5 & 49.6 \\\\ \text{PosterGen} & 28.0 & 2.8 & 53.0 & 75.4 & 79.7 & 68.3 & 65.6 & 67.0 \\\\ \text{P2P} & 2.0 & 0.9 & 11.0 & 76.4 & 76.7 & 61.3 & 49.6 & 49.6 \\\\ \text{ResearchStudio} & 0.0 & 0.0 & 73.7 & 97.0 & 97.0 & 93.5 & 88.1 & 88.1 \\\\ \hline \text{Human reference} & & & & & & & & \\\\ \text{Official posters} & 3.0 & 0.8 & 72.7 & 86.4 & 87.4 & 90.0 & 83.0 & 83.5 \\\\ \hline \end{array}$$ Excluding *N/A* raises minimalist baselines such as Paper2Poster ($+4.1$) and PosterForest ($+4.1$) as expected, while leaving the top-6 ranking unchanged and VASP ahead of ResearchStudio (94.3 and 92.8 vs. 88.1) and PosterMELD/PosterGen ($>24$ points). Dropping *EN05* altogether gives virtually identical Overall scores (94.3, 92.9, 88.3, 82.6, 68.3, 65.9). We will adopt the *Excl. N/A* denominator in Eq. 6 in the revision.
Reviewer_P6veOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • Existing poster generation methods often prioritize text and layout, leaving insufficient space for important figures. VASP proposes a visual-first approach that organizes the poster around key figures.

  • SemParser uses an LLM to classify the semantic roles of PDF text, while a program leverages PDF geometry to extract figures accurately.

  • VisComposer connects key claims with supporting evidence and determines figure sizes and text placement based on their importance.

  • Semantic-Geometric Verification separates semantic reasoning by the LLM from geometric checks by programs, including figure sizes, claim-evidence proximity, and reading order. The layout is iteratively refined based on this feedback.

Critical strengths
  • The paper is clearly written and well organized, with informative figures that make the proposed method easy to understand.
  • The visual-first approach is well motivated, as it aligns with how human poster designers may prioritize key visual evidence when creating scientific posters.
  • The method effectively decouples LLM-based semantic reasoning from programmatic geometric computation, achieving strong performance. Programmatic geometric verification also contributes to improved efficiency.
  • The quality of figure extraction is further validated through human annotations and manual completeness checks.
Critical weaknesses
  • The qualitative results show that the generated posters have different visual themes. It would be helpful to investigate whether these stylistic differences influence the Visual Presentation Scores assigned by the VLM judge.

  • The contribution of Semantic Verification (SV) to preserving claim-evidence relationships in the actual layout is not directly evaluated. In Table 6, removing SV leaves the Spatial Association score unchanged at 100%, while the largest performance drop occurs in Whitespace (80% to 10%). Additional (or more direct) experiments are needed to verify whether SV effectively improves claim-evidence alignment.

  • The proposed Visual Presentation Benchmark lacks validation against human judgments. Notably, the official human-designed posters score lower than VASP (83.0 vs. 94.3). Evaluating the correlation between the proposed benchmark and human preferences would help establish its reliability.

  • Using GPT-5.6-Sol consistently across all baselines provides a fair comparison. However, it would be valuable to investigate whether VASP maintains its relative performance advantage when other backbone models, such as Opus-5 or Qwen 3, are applied to the baselines.

Questions for authors

I would appreciate it if the authors could address the concerns raised in the weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer P6ve

We sincerely thank Reviewer P6ve for the clear and constructive review. We are glad the reviewer finds our paper well organized, our visual-first approach well motivated by how human designers prioritize key visual evidence, our decoupling of semantic reasoning from programmatic geometry effective in both quality and efficiency, and our figure extraction validated by human annotations. We address each weakness below.

W.1 The generated posters have different visual themes; do stylistic differences influence Visual Presentation scores?

VASP maintains its mean score advantage across all nine tested themes, indicating that this advantage does not stem from VASP's default visual theme. On the 20-paper ablation subset, we froze VASP (GPT-5.6-Sol)'s HTML text, extracted figures, and grid layout, and varied only the CSS stylesheet across six dimensions: canvas palette, typography, header banner, section title bars, container borders and card surfaces, and conclusion callouts. Nine themes were tested: VASP's default (T0), six themes matching the styles of the six baselines (T1–T6), a zero-color monochrome wireframe with 1px square borders (T7), and a classic LaTeX Beamer serif theme (T8). Everything else is unchanged:

\hline \text{Visual Theme / Method} & \text{Figure Presentation}\uparrow & \text{Visual Guidance}\uparrow & \text{Layout}\uparrow & \text{Overall}\uparrow \\\\ \hline \text{VASP (GPT-5.6-Sol) with fixed layout across 9 visual themes} & & & & \\\\ \text{T0: VASP Default} & 80.0 & \mathbf{100.0} & \mathbf{91.2} & \mathbf{90.4} \\\\ \text{T1: Paper2Poster Style} & 73.3 & \mathbf{100.0} & \underline{90.0} & 87.8 \\\\ \text{T2: PosterForest Style} & \underline{83.3} & 99.0 & 87.5 & \underline{89.9} \\\\ \text{T3: PosterMELD Style} & 75.0 & \mathbf{100.0} & 87.5 & 87.5 \\\\ \text{T4: PosterGen Style} & 70.8 & \mathbf{100.0} & \mathbf{91.2} & 87.4 \\\\ \text{T5: ResearchStudio Style} & 73.3 & \mathbf{100.0} & 88.8 & 87.4 \\\\ \text{T6: P2P Style} & 72.5 & \mathbf{100.0} & \underline{90.0} & 87.5 \\\\ \text{T7: Monochrome Wireframe} & 74.2 & \mathbf{100.0} & 88.8 & 87.6 \\\\ \text{T8: LaTeX Beamer Classic} & \mathbf{84.2} & \mathbf{100.0} & 82.5 & 88.9 \\\\ \hline \text{Mean }\pm\text{ SD across 9 themes} & 76.3\pm 5.0 & 99.9\pm 0.3 & 88.6\pm 2.7 & 88.3\pm 1.2 \\\\ \hline \text{Baseline and human references on the same 20 papers} & & & & \\\\ \text{ResearchStudio} & 69.2 & 95.0 & 87.5 & 83.9 \\\\ \text{Official posters} & 67.5 & 89.0 & 86.2 & 80.9 \\\\ \hline \end{array}$$ Across all nine stylesheets, Overall stays within a narrow band of 87.4 to 90.4 (88.3 ± 1.2). Even stripped to a monochrome wireframe (T7, 87.6) or rendered in ResearchStudio's minimal style (T5, 87.4), VASP remains above ResearchStudio's generated posters (83.9) and the official posters (80.9). Dropping the only two style-related criteria in Table 3 (COL01 color harmony and EN05 decoration) across all 100 papers likewise leaves VASP's 10-criterion score unchanged at 94.3 (GPT-5.6-Sol) and 92.8 (Opus-5), against 88.1 for ResearchStudio. Although surface styling introduces modest variation (87.4–90.4), VASP outperforms the baselines under every tested theme, confirming that its lead is not an artifact of default styling.
AuthorsOpenReview ↗

Response to Reviewer P6ve

W.2 Removing Semantic Verification (SV) leaves Spatial Association at 100% while Whitespace drops from 80% to 10%; direct evaluation of claim-evidence alignment is needed.

Two factors produced that pattern in Table 6. First, Assoc. (AL05) is a layout-proximity check, not a semantic one: on the rendered poster, it tests whether labels such as headings, captions, and legends sit visually next to the elements they belong to, rather than floating apart. Removing SV changes what a section's text says, not where headings or captions are placed, so Assoc. stays at 95%–100% even when the body text never mentions the figure next to it. To directly measure text-to-figure semantic alignment, we evaluate Claim-Evidence Grounding (CEG), a blind VLM check on the rendered poster testing whether each displayed figure or table is substantively explained by its section's body text.

Second, in the submitted Table 6, the w/o semantic verification row disabled the revision controller—turning off both iterative SV and ordinary geometric/whitespace verification (GV), while retaining initial (round-0) SV planning. Below we vary Initial SV, Iterative SV, and GV orthogonally on the 20-paper subset, keeping extracted assets and visual-first composition unchanged:

\hline \text{Verification} & & & & \text{Figure (\\%)} & & & \text{Layout (\\%)} & & \text{Area (\\%)} & \text{Semantics (\\%)} & \\\\ \hline \text{Init. SV} & \text{Iter. SV} & \text{GV} & \text{Overall}\uparrow & \text{Fit}\uparrow & \text{Text}\uparrow & \text{Assoc.}\uparrow & \text{Space}\uparrow & \text{Density}\uparrow & \text{Largest}\uparrow & \text{CEG}\uparrow & \text{Coverage}\uparrow \\\\ \hline \checkmark & \checkmark & \checkmark & \mathbf{91.4} & \mathbf{80} & \mathbf{65} & \mathbf{100} & \underline{80} & \mathbf{90} & \underline{7.3} & \underline{82.3} & 75.7 \\\\ \checkmark & \text{--} & \checkmark & \underline{90.6} & \mathbf{80} & \mathbf{65} & \mathbf{100} & 75 & \underline{85} & 6.9 & 78.2 & \underline{77.1} \\\\ \text{--} & \text{--} & \checkmark & 89.0 & 60 & \underline{60} & \mathbf{100} & \mathbf{95} & \mathbf{90} & \mathbf{8.9} & 25.0 & 67.9 \\\\ \checkmark & \checkmark & \text{--} & 80.1 & 65 & 50 & 95 & 15 & 65 & 7.0 & \mathbf{88.1} & \mathbf{80.0} \\\\ \checkmark & \text{--} & \text{--}^{\dagger} & 81.1 & \underline{70} & \underline{60} & \mathbf{100} & 10 & 65 & 7.0 & 72.1 & \underline{77.1} \\\\ \hline \end{array}$$ *$^{\dagger}$Submitted *w/o semantic verification* row in Table 6 (retained initial SV planning while disabling iterative SV and GV). Row 1 ($\checkmark, \checkmark, \checkmark$) is VASP (full).* The three components separate cleanly. **GV governs whitespace**: disabling GV (rows 4–5) drops *Space* to 15% and 10%, whereas keeping GV (rows 1–3) holds *Space* at 75%–95%; disabling GV—not SV—caused the 80% to 10% drop in Table 6. **Initial SV establishes claim-evidence grounding**: with GV fixed, removing initial SV as well (row 2 → row 3) collapses CEG from 78.2% to 25.0% and Coverage from 77.1% to 67.9%, as columns fill with generic bullets unrelated to the displayed figures. **Iterative SV preserves grounding across revisions**: when GV runs without iterative SV (row 1 → row 2), geometry-only edits trim figure-explaining prose to fix overflow, lowering CEG from 82.3% to 78.2% (and without GV, adding iterative SV in row 5 → row 4 lifts CEG from 72.1% to 88.1%).
AuthorsOpenReview ↗

Response to Reviewer P6ve

W.3 The Visual Presentation Benchmark lacks validation against human judgments, and official posters score lower than VASP (83.0 vs. 94.3).

Three human evaluators performed independent blind ranking across all 100 papers (N=300N = 300 rankings). On each paper, method identities were anonymized and shuffled across five representative sets—VASP (GPT-5.6-Sol), ResearchStudio, Official posters, PosterGen, and Paper2Poster—and each evaluator ranked the five posters from 1 (best) to 5 on Visual Readability:

\hline & \text{Human Blind Ranking} & & \text{Table 3 Visual Presentation} & \\\\ \hline \text{Method} & \text{Mean Rank}\downarrow & \text{Rank}\downarrow & \text{Score (\\%)}\uparrow & \text{Rank}\downarrow \\\\ \hline \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \mathbf{1.47} & \mathbf{1} & \mathbf{94.3} & \mathbf{1} \\\\ \text{ResearchStudio} & \underline{2.20} & \underline{2} & \underline{88.1} & \underline{2} \\\\ \text{Official posters} & 2.65 & 3 & 83.0 & 3 \\\\ \text{PosterGen} & 4.06 & 4 & 65.6 & 4 \\\\ \text{Paper2Poster} & 4.62 & 5 & 46.5 & 5 \\\\ \hline \end{array}$$ Blind human rankings (1.47, 2.20, 2.65, 4.06, 4.62) produce the exact same 1-to-5 ordering as Table 3 Visual Presentation (94.3, 88.1, 83.0, 65.6, 46.5). Independent 12-criterion human annotation on the 20-paper subset (Response to Reviewer qNaT, W.1) likewise yields the identical 1-to-5 ordering (95.9, 92.9, 86.9, 75.2, 60.6), with 83.1% human–GPT-5.6-Sol agreement against 83.3% pairwise human–human agreement. **Why official posters score below VASP on Table 3.** Human evaluators themselves rank VASP 1st (1.47) and ResearchStudio 2nd (2.20) ahead of official posters (2.65), and score official posters 86.9% vs. 95.9% for VASP on the 20-paper human annotation above—confirming that Table 3 reflects genuine readability differences rather than benchmark bias. While many author-designed posters are clear and well crafted, official posters are prepared by researchers rather than professional designers and vary widely in layout quality: a substantial portion pack too many figures or overly dense text onto a single board (reaching 96.4% Coverage in Table 4, but dropping Figure Presentation to 72.7% and Visual Guidance to 86.4% in Table 3), which sharply reduces readability and pulls down the 100-paper average. Helping researchers produce consistently legible scientific posters without manual layout bottlenecks is directly what VASP is designed for. ### **W.4 Does VASP maintain its relative advantage when other backbones, such as Opus-5 or Qwen 3, are applied to the baselines?** We ran VASP alongside the top three representative baselines—ResearchStudio (strongest visual baseline), PosterGen (strongest coverage baseline), and Paper2Poster—across all 100 papers under both Claude Opus-5 and Qwen3.8-Flash-Next. The evaluation protocol is unchanged: $$\begin{array}{lcccc} \hline \text{Method} & \text{VLM-as-Judge} & \text{PaperQuiz} & \text{Visual Presentation} & \text{Key-Information} \\\\ & \text{(Table 1)}\uparrow & \text{(\\%)}\uparrow & \text{(\\%)}\uparrow & \text{Coverage (\\%)}\uparrow \\\\ \hline \text{Claude Opus-5 backbone} & & & & \\\\ \mathbf{VASP} & \mathbf{3.70} & \mathbf{78.08} & \mathbf{92.5} & \mathbf{88.7} \\\\ \text{ResearchStudio} & \underline{3.58} & 71.77 & \underline{88.7} & 77.9 \\\\ \text{PosterGen} & 3.00 & \underline{76.05} & 65.9 & \underline{86.6} \\\\ \text{Paper2Poster} & 2.76 & 74.55 & 44.0 & 78.0 \\\\ \hline \text{Qwen3.8-Flash-Next backbone} & & & & \\\\ \mathbf{VASP} & \mathbf{3.58} & \mathbf{81.35} & \mathbf{89.2} & \mathbf{89.3} \\\\ \text{ResearchStudio} & \underline{3.54} & \underline{73.68} & \mathbf{89.2} & 80.4 \\\\ \text{PosterGen} & 3.04 & 73.28 & 65.7 & \underline{82.6} \\\\ \text{Paper2Poster} & 2.42 & 73.16 & 46.7 & 74.7 \\\\ \hline \text{Human reference} & & & & \\\\ \text{Official posters} & 3.12 & 67.55 & 83.0 & 96.4 \\\\ \hline \end{array}$$ VASP ranks 1st (or tied-1st) on every benchmark under both backbones. ResearchStudio matches or approaches VASP on Visual Presentation (88.7% and 89.2%) only by cutting content, trailing VASP by 6.3 to 7.7 points on PaperQuiz and by 8.9 to 10.8 points on Key-Information Coverage. Stronger backbones raise text recall and coverage in PosterGen and Paper2Poster, but without visual-first anchoring and geometric verification, the added text crowds the canvas and leaves Visual Presentation at 65.7%–65.9% and 44.0%–46.7%. VASP is the only method that stays strong on visual presentation and content retention at once across backbones. We again thank Reviewer P6ve for the review. We hope this response has addressed all the concerns raised, and we are happy to provide further detail on any point.
Reviewer_qNaTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper identifies two challenges in the automatic generation of posters from scientific papers: “layout mismatch,” in which important figures and tables are squeezed into existing layouts or templates and become too small; and “execution mismatch,” in which semantic interpretation and geometric computation are handled by the same model. The paper proposes VASP, a system that constructs posters based on important visual evidence. VASP consists of two stages: SemParser, in which an LLM determines the semantic roles of each element and a program accurately extracts the boundaries of figures from the geometric information in the PDF; and VisComposer, which performs layout and iterative revisions without templates while explicitly indicating the relationship between claims and evidence. In an evaluation using 100 papers, VASP outperforms existing fixed-canvas methods in terms of visual presentation quality, figure extraction accuracy, and preservation of author-emphasized key information.

Critical strengths
  1. This paper presents a clear design principle called “Vision-Centric Anchoring,” which uses key figures as the starting point for layout, making it easy to understand how it differs from conventional text-first or template-first approaches. This method explicitly handles the relationship between claims and evidence. Furthermore, an ablation study shows that removing the anchored composition causes the visual-presentation score to drop from 91.4 to 80.8, thereby demonstrating the effectiveness of the proposed method.

  2. This paper proposes “Semantic-Geometric Decoupling,” which divides the task between an LLM for semantic judgment and a program for geometric computation and verification. SemParser entrusts the understanding of figures to the LLM while calculating actual boundaries from PDF geometry, thereby reducing figure omissions and incorrect cropping. It achieves a mAP of 94.30 in figure extraction and perfect cropping in all 100 manually verified cases.

Critical weaknesses
  1. The evaluation relies heavily on VLM (GPT-5.6-Sol), raising concerns about the reliability of the evaluation method. In particular, the results obtained when applying the proposed method to GPT-5.6-Sol may exhibit self-preference bias. Independent annotation by two or three human evaluators should be conducted to demonstrate agreement with the VLM judge as well as human–human agreement.

  2. Although the authors have introduced a new evaluation framework, there is “benchmark-method alignment” between it and the proposed method. In other words, there is a fairly direct correspondence between the characteristics that VASP is designed to improve and the evaluation criteria newly defined by the authors. Such an evaluation creates favorable conditions for the proposed method and may introduce bias into the evaluation.

Questions for authors

Please see the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer qNaT

We sincerely thank Reviewer qNaT for the careful reading. We are glad the reviewer finds Vision-Centric Anchoring a clear design principle that handles claim-evidence relationships differently from text-first or template-first approaches, and recognizes Semantic-Geometric Decoupling for achieving 94.30 mAP in figure extraction and complete crops in all 100 manually verified cases. We address each weakness below.

W.1 The evaluation relies heavily on VLM (GPT-5.6-Sol) and may exhibit self-preference bias; independent human annotation is needed.

Two independent evaluations test whether our conclusions depend on GPT-5.6-Sol: a cross-family VLM judge across all 100 papers, and criterion-level annotation by three human evaluators.

Cross-family evaluation with Gemini-4-Argon. We re-ran all three poster-level benchmarks in the main paper—Table 1 (Paper2Poster VLM-as-Judge and PaperQuiz), Table 3 (Visual Presentation), and Table 4 (Key-Information Coverage)—on all 100 papers using Gemini-4-Argon, which belongs to a separate model family and is never used by VASP or any baseline. Prompts, schemas, and scoring rules are unchanged:

Table 1 (Gemini-4-Argon): Paper2Poster VLM-as-Judge and PaperQuiz (%).

\hline & & \text{VLM-as-Judge} & & & \text{PaperQuiz} & \\\\ \hline \text{Method} & \text{Aesthetic}\uparrow & \text{Information}\uparrow & \text{Overall}\uparrow & \text{Verbatim}\uparrow & \text{Interpretive}\uparrow & \text{Overall}\uparrow \\\\ \hline \text{Ours} & & & & & & \\\\ \mathbf{VASP}\\ \text{(Opus-5)} & \underline{3.62} & \mathbf{3.83} & \mathbf{3.72} & \mathbf{60.46} & \mathbf{86.70} & \mathbf{73.58} \\\\ \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \mathbf{3.72} & 3.62 & \underline{3.67} & \underline{45.68} & \underline{84.68} & \underline{65.18} \\\\ \hline \text{Baseline methods} & & & & & & \\\\ \text{Paper2Poster} & 1.72 & 2.81 & 2.27 & 43.96 & 82.28 & 63.12 \\\\ \text{PosterMELD} & 1.82 & 2.69 & 2.26 & 34.72 & 77.74 & 56.23 \\\\ \text{PosterForest} & 2.06 & 2.83 & 2.44 & 41.46 & 81.84 & 61.65 \\\\ \text{PosterGen} & 2.34 & 3.30 & 2.82 & 43.76 & 80.56 & 62.16 \\\\ \text{P2P}^{\dagger} & 2.21 & \underline{3.65} & 2.93 & 66.46 & 91.36 & 78.91 \\\\ \text{ResearchStudio} & 3.30 & 3.26 & 3.28 & 42.94 & 81.38 & 62.16 \\\\ \hline \text{Human reference} & & & & & & \\\\ \text{Official posters} & 3.07 & 3.29 & 3.18 & 39.70 & 78.60 & 59.15 \\\\ \hline \end{array}$$ *$^{\dagger}$P2P renders a fixed-width page of unbounded height and is not ranked on PaperQuiz.* **Table 3 (Gemini-4-Argon):** Visual Presentation (%). $$\begin{array}{lcccc} \hline \text{Method} & \text{Figure Presentation}\uparrow & \text{Visual Guidance}\uparrow & \text{Layout}\uparrow & \text{Overall}\uparrow \\\\ \hline \text{Ours} & & & & \\\\ \mathbf{VASP}\\ \text{(Opus-5)} & \underline{75.3} & \underline{90.0} & \mathbf{94.0} & \mathbf{86.4} \\\\ \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \mathbf{83.7} & 82.8 & \underline{90.3} & \underline{85.6} \\\\ \hline \text{Baseline methods} & & & & \\\\ \text{Paper2Poster} & 12.0 & 35.6 & 36.3 & 28.0 \\\\ \text{PosterMELD} & 50.3 & 58.4 & 44.3 & 51.0 \\\\ \text{PosterForest} & 20.7 & 39.2 & 36.5 & 32.1 \\\\ \text{PosterGen} & 21.0 & 48.2 & 63.5 & 44.2 \\\\ \text{P2P} & 2.0 & 52.8 & 60.3 & 38.4 \\\\ \text{ResearchStudio} & 65.7 & \mathbf{94.6} & 88.3 & 82.8 \\\\ \hline \text{Human reference} & & & & \\\\ \text{Official posters} & 61.0 & 75.6 & 92.0 & 76.2 \\\\ \hline \end{array}$$ **Table 4 (Gemini-4-Argon):** Author-informed Key-Information Coverage (%). $$\begin{array}{lccccc} \hline & & \text{Coverage by information type} & & & \\\\ \hline \text{Method} & \text{Problem}\uparrow & \text{Method}\uparrow & \text{Results}\uparrow & \text{Overall}\uparrow & \text{Full}\uparrow \\\\ \hline \text{Ours} & & & & & \\\\ \mathbf{VASP}\\ \text{(Opus-5)} & \mathbf{94.0} & \mathbf{88.3} & \mathbf{85.4} & \mathbf{88.3} & \mathbf{41} \\\\ \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \underline{87.2} & \underline{82.5} & 68.2 & 79.1 & \underline{21} \\\\ \hline \text{Baseline methods} & & & & & \\\\ \text{Paper2Poster} & 72.5 & 72.9 & 73.2 & 73.1 & 12 \\\\ \text{PosterMELD} & 81.2 & 72.5 & 46.0 & 65.0 & 5 \\\\ \text{PosterForest} & 66.4 & 75.9 & \underline{76.8} & 72.7 & 10 \\\\ \text{PosterGen} & 85.2 & 81.8 & 74.2 & \underline{79.3} & \underline{21} \\\\ \text{P2P}^{\dagger} & 90.6 & 94.5 & 92.4 & 93.3 & 65 \\\\ \text{ResearchStudio} & 74.5 & 74.2 & 68.2 & 71.6 & 14 \\\\ \hline \text{Human reference} & & & & & \\\\ \text{Official posters} & 99.3 & 96.9 & 94.9 & 96.4 & 80 \\\\ \hline \end{array}$$ *Full: share of posters covering all seven units. $^{\dagger}$P2P renders a page of unbounded height and is not ranked.*
AuthorsOpenReview ↗

Response to Reviewer qNaT (continued)

Under Gemini-4-Argon, VASP takes 1st and 2nd place on VLM-as-Judge (3.72 and 3.67 vs. 3.28 for ResearchStudio), PaperQuiz (73.58% and 65.18%), and Visual Presentation (86.4% and 85.6% vs. 82.8% for ResearchStudio and 76.2% for official posters). On Key-Information Coverage, VASP (Opus-5) ranks 1st (88.3%, Full 41), while VASP (GPT-5.6-Sol) matches PosterGen (79.1% vs. 79.3%, Full 21) and exceeds ResearchStudio by 7.5 points. Overall, VASP's advantage holds under an unrelated judge family.

Independent human annotation and agreement. Within the rebuttal timeframe, we prioritized human annotation for Visual Presentation (Table 3), where layout, hierarchy, and legibility rely heavily on perceptual judgment; Key-Information Coverage (Table 4) was cross-checked with Gemini-4-Argon above, reaching 93.0% unit-level agreement with GPT-5.6-Sol. On a randomly sampled 20-paper subset, three human evaluators not involved in this project independently annotated all 12 Visual Presentation criteria across five anonymized, per-paper shuffled poster sets (1,200 judgments per evaluator): VASP (GPT-5.6-Sol), ResearchStudio, Official posters, PosterGen, and Paper2Poster. Computing Table 3 from the human labels gives:

\hline & & & \text{Human Evaluation (3-Evaluator Mean)} & & & \text{GPT-5.6-Sol (Same 20 Papers)} & \\\\ \hline \text{Method} & \text{Figure Pres.}\uparrow & \text{Visual Guid.}\uparrow & \text{Layout}\uparrow & \text{Overall}\uparrow & \text{Rank}\downarrow & \text{Overall}\uparrow & \text{Rank}\downarrow \\\\ \hline \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \mathbf{95.6} & \mathbf{98.7} & 93.3 & \mathbf{95.9} & \mathbf{1} & \mathbf{94.0} & \mathbf{1} \\\\ \text{ResearchStudio} & \underline{87.2} & \underline{97.3} & \mathbf{94.2} & \underline{92.9} & \underline{2} & \underline{89.2} & \underline{2} \\\\ \text{Official posters} & 83.3 & 83.6 & \underline{93.8} & 86.9 & 3 & 79.5 & 3 \\\\ \text{PosterGen} & 72.5 & 79.8 & 73.3 & 75.2 & 4 & 71.3 & 4 \\\\ \text{Paper2Poster} & 59.7 & 66.5 & 55.4 & 60.6 & 5 & 52.4 & 5 \\\\ \hline \end{array}$$ Human evaluators and GPT-5.6-Sol produce the exact same 1-to-5 ordering. All three evaluators independently rank VASP (GPT-5.6-Sol) 1st (98.2, 92.9, and 96.5) and the mean score is higher than GPT-5.6-Sol scores itself (95.9 vs. 94.0). Criterion-level agreement across all 1,200 judgments is below: $$\begin{array}{lcccc} \hline \text{Category} & \text{Evaluator Pair} & \text{Matched (/1,200)}\uparrow & \text{Agreement (\\%)}\uparrow & \text{Mean Agreement (\\%)}\uparrow \\\\ \hline \text{Human–Human} & \text{H1 vs. H2} & 978 & 81.5 & \\\\ & \text{H1 vs. H3} & \text{1,014} & 84.5 & \mathbf{83.3} \\\\ & \text{H2 vs. H3} & \text{1,007} & 83.9 & \\\\ \hline \text{Human–GPT-5.6-Sol} & \text{H1 vs. GPT-5.6-Sol} & 981 & 81.8 & \\\\ & \text{H2 vs. GPT-5.6-Sol} & \text{1,023} & 85.3 & \mathbf{83.1} \\\\ & \text{H3 vs. GPT-5.6-Sol} & 988 & 82.3 & \\\\ \hline \end{array}$$ The close human–VLM agreement (83.1% vs. 83.3% human–human) and identical 1-to-5 ranking support the reliability of the visual-presentation results on this subset. Together with the cross-family evaluation, these findings reduce the concern that VASP's observed advantage is driven by GPT-5.6-Sol judge preference.
AuthorsOpenReview ↗

Response to Reviewer qNaT

W.2 "Benchmark-method alignment" between the new evaluation framework and VASP may introduce bias.

Both the benchmark and VASP follow from observing how existing poster pipelines fail, not from fitting the benchmark to the method. Existing pipelines repeatedly shrink key figures until axis labels become unreadable, separate claims from the figures they refer to, and either overflow boxes or leave large blank areas. Existing metrics provide complementary evidence, but do not directly diagnose specific visual failures or measure retention of author-prioritized takeaways: holistic 1–5 VLM-as-Judge ratings do not localize concrete visual failures, while PaperQuiz queries details sampled across the full paper rather than what a single-page poster should convey to help readers quickly grasp the core work—so even author-designed official posters score only 67.55%, below text-heavy Paper2Poster (69.25%) and unbounded-height P2P (85.20%). This motivated both sides of our work: evaluating criterion-level visual legibility (Table 3) alongside selective, author-prioritized takeaways from the official poster (Table 4), and designing VASP around key visual evidence as human poster designers do (Reviewer P6ve). We will add a failure-case analysis of existing pipelines before introducing the method and benchmark in the revision.

VASP's gains hold after removing criteria related to its internal checks. First, in Visual Presentation (Table 3), five of the 12 criteria evaluate layout geometry that VASP also checks programmatically (§ A.2)—figure size, overlap, spatial association, density, and whitespace—whereas the remaining seven evaluate rendered visual quality, such as in-figure text legibility, 30%-scale clarity, color harmony, and conclusion emphasis. Second, Key-Information Coverage (Table 4) is scored against seven paper-audited reference units frozen from the author's official poster before generation, which VASP never sees. Below we report those seven Table 3 criteria alongside Table 4 across all 100 papers:

\hline \text{Method} & \text{Visual Presentation} & \text{Visual Presentation} & \text{Key-Information} \\\\ & \text{(All 12 Criteria)}\uparrow & \text{(7 Held-Out Criteria)}\uparrow & \text{Coverage (Table 4)}\uparrow \\\\ \hline \text{Ours} & & & \\\\ \mathbf{VASP}\\ \text{(Opus-5)} & \underline{92.5} & \underline{94.4} & \mathbf{88.7} \\\\ \mathbf{VASP}\\ \text{(GPT-5.6-Sol)} & \mathbf{94.3} & \mathbf{97.1} & \underline{82.0} \\\\ \hline \text{Baseline methods} & & & \\\\ \text{Paper2Poster} & 46.5 & 53.7 & 74.1 \\\\ \text{PosterMELD} & 66.3 & 79.6 & 66.9 \\\\ \text{PosterForest} & 45.5 & 52.7 & 73.9 \\\\ \text{PosterGen} & 65.6 & 70.3 & 79.7 \\\\ \text{P2P}^{\dagger} & 49.6 & 58.6 & 93.6 \\\\ \text{ResearchStudio} & 88.1 & 91.6 & 73.0 \\\\ \hline \text{Human reference} & & & \\\\ \text{Official posters} & 83.0 & 82.3 & 96.4 \\\\ \hline \end{array}$$ *$^{\dagger}$P2P renders a page of unbounded height and is not ranked.* Without the five geometric criteria, the ranking is unchanged: VASP still takes 1st and 2nd place (97.1% and 94.4%, vs. 91.6% for ResearchStudio and 82.3% for official posters). **VASP also leads on prior benchmarks and human evaluation.** It ranks 1st on the prior Paper2Poster VLM-as-Judge (3.71 vs. 3.52) and fixed-canvas PaperQuiz in Table 1, reaches 94.30 mAP and 100/100 human-verified crop completeness across 948 annotated figures in Table 2, and ranks 1st under both Gemini-4-Argon and blind human evaluation (W.1). We again thank Reviewer qNaT for the review. We hope this response has addressed all the concerns raised, and we are happy to provide further detail on any point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.