Figures as Programs: Recursive Generation of Editable Scientific Figures
Yepeng Liu, Dasen Dai, Chengzhi Liu, Yiren Song, Hai Ci, Yu Zhang, Qi Zhang, Mike Zheng Shou, et al.
cs.AI · cs.GR
Quick verdict
FigTree recasts paper-to-figure as recursive SVG code via multi-agent LLMs, topping PaperBananaBench; promising but judge-dependent, statistically untested.
Summary
The paper recasts scientific methodology-figure generation from raster image synthesis into recursive construction of editable SVG programs. FigTree is a multi-agent LLM pipeline: a grounder extracts a conservative content plan and global style from the paper text; generator agents recursively decompose a layout tree (two to four children per level, depth capped at 10), with each leaf drawing a short SVG fragment exposing named boundary ports; parent nodes merge fragments and exclusively draw cross-boundary connections; and a render-critic loop inspects both the rendered PNG and the SVG code, issuing element-anchored repairs within a three-round budget, guided by verification rules distilled offline from AutoFigure-Edit Bench. On PaperBananaBench (292 NeurIPS 2025 methodology figures), FigTree leads the reference-based evaluation on all four dimensions with 59.59% overall, 9.59 points above the strongest agentic baseline (Crafter), and tops an eight-dimension diagnostic rubric at 8.39/10. Ablations attribute the largest contribution to verification memory (-3.95) and paper grounding (-2.47), and an eight-round editing study shows the vector branch recovering quality better than a raster branch. Main limitations: all three evaluations rely on undisclosed VLM judges with no human study; GPT-Image-2 ranks near-worst (1.37%) under the reference protocol yet second-best (8.28/10) under the rubric, an unaddressed judge-sensitivity signal; no variance or significance testing; the ablation subset is unspecified; and coverage is limited to ML methodology figures with ML-specific hardcoded aesthetics.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Scientific figure generation is better formulated as recursive SVG program construction, decomposing one long-horizon generation problem into short, locally verifiable subprograms.
- FigTree achieves the best results across all four dimensions of the reference-based evaluation on PaperBananaBench, with 59.59% overall, 9.59 points above the strongest baseline Crafter (Table 1).
- Recursive decomposition, paper grounding, and render-critic repair each causally contribute to quality, with verification memory the largest factor (-3.95 when removed, Table 3).
- Vector (SVG) representation enables more effective multi-round editing than raster representation under the same editing protocol (Figure 2, eight rounds).
- The approach particularly controls hallucination, achieving 8.59/10 hallucination control versus 8.41 semantic coverage (Table 2).
Strengths
Well-posed representation shift with concrete mechanisms
Recasting figure generation as recursive SVG construction (Sections 3.1-3.3) turns one long-horizon program into short fragments; boundary ports plus parent-owned cross-boundary connections give each inter-fragment edge exactly one owner, preventing duplicate edges - a real design insight, not just prompting.
Backbone-controlled baseline
Direct-SVG shares FigTree's Claude Sonnet 4.6 backbone but omits recursion and critic-guided repair, isolating the system's contribution (Table 1: 3.77 vs 59.59 overall); a five-component ablation (Table 3) further decomposes the effect.
Honest failure accounting
Gains concentrate exactly where raster methods fail
Full prompt transparency
Weaknesses
Undisclosed judges, no human study
Unexplained protocol disagreement
No statistical rigor
Benchmark-coupled verification memory
Narrow domain with hardcoded aesthetics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: Figures as Programs (FigTree, arXiv 2609.01006)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- 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 does FigTree do?
FigTree is a multi-agent LLM system that reads a scientific paper and automatically produces a publication-style methodology figure as an editable SVG (vector) program rather than a raster image. It grounds content in the paper text, recursively decomposes the figure into short SVG fragments, and uses a render-critic loop that inspects both the rendered image and the code to repair defects element by element.
How much better is FigTree than existing methods?
On PaperBananaBench (292 methodology figures from NeurIPS 2025 papers), FigTree scores 59.59% overall in the reference-based comparison, 9.59 points above the strongest baseline (Crafter), leading all four dimensions: faithfulness 46.23, conciseness 64.38, readability 51.37, aesthetics 66.78. It also tops an eight-dimension diagnostic rubric at 8.39/10. Note that even so, author-drawn figures still win the majority of faithfulness comparisons (46.23 < 50).
Why generate figures as SVG code instead of images?
In an SVG program every visual element is an addressable code statement, enabling precise local edits and resolution independence. The paper's eight-round editing experiment shows that a vector branch recovers figure quality better than a raster branch edited under the same protocol, and defects found visually can be traced back to specific code statements for repair.
Is the code and evaluation setup available?
The paper states code is available at github.com/yepengliu/FigTree, and Appendices B-C publish all evaluation and pipeline prompts including scoring anchors. However, the judge models used for scoring are not named, and the ablation experiments use an unspecified subset of the benchmark, which limits full independent verification.
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 2 and 3 are plots: per-round editing scores and per-method cost/quality coordinates are not recoverable from text extraction beyond axis labels and sparse annotations (e.g., original figure quality 8.51 in Figure 2).
- Rendered outputs (Figures 1 and 4) cannot be visually assessed from the extracted text; figure-quality judgments here rely solely on the reported tables.
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.