{"id":"437bee79-6c45-4147-ac4e-9d978f4ba981","arxiv_id":"2607.28073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A curriculum-plus-RL-plus-repair SVG agent, trained on a new 100K office-diagram dataset, beats strong LLM baselines on layout-correct logical diagrams from long professional text.","lead":"GVR-Coder turns long meeting notes and documents into editable SVG diagrams using curriculum fine-tuning, dual visual rewards, and a generate-verify-repair loop. It ships a 100K office-diagram dataset and reports higher acceptance rates than larger closed-source models on a 320-sample test set.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Headline Ar gains vs Gemini rest on Gemini as data factory, reward labeler, verify agent, and primary judge; independence is only partially stress-tested.","rationale":"The reader correctly isolates the Gemini-centric stack as the weakest assumption behind the strongest claim (Table 1 Ar dominance). The paper is a solid systems contribution—dataset, CRSFT curriculum, dual-reward GRPO, and VRL are clearly specified, with stage-wise ablations and some cross-checks—but the headline “outperforms… reliably” is not fully externalized while the same proprietary VLM defines data, rewards, repair, and primary scores. Cross-VLM Table 2 and human Table 4/Fig.4 reduce but do not eliminate circularity risk. No stronger independent flaw (e.g. broken math or non-reproducible core method) displaces this. Verdict stays CONDITIONAL; confidence remains moderate until a Gemini-free human/open re-eval confirms the gap. Concrete test above would settle it.","tokens_in":21939,"tokens_out":647,"duration_ms":13419,"concrete_test":"On the fixed 320-test set, freeze GVR-Coder-32B-VRL outputs and re-score Ar/As with a held-out protocol that never entered training or VRL: (i) fully blind human panel (new designers, forced ranking vs Gemini-3-pro/GPT-5.5 on the same captions) and (ii) an open VLM judge (e.g. Qwen2.5-VL) using only the paper’s six-defect + factual rubrics, without Gemini in the loop. If human preference for GVR over Gemini falls below ~10pp or open-judge Ar gap vs Table 1 collapses >15pp, the outperformance claim is judge-dependent.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that GVR-Coder (esp. 32B-VRL) “outperforms competitive baselines” and reaches Ar up to 100%/82–90% above Gemini-3-pro (Table 1) is load-bearing on Gemini-3-Pro being an independent quality oracle. Per §3, Gemini generates the two SVG candidates and drives Judge/Repair for DocMeetSVG-100K; §4.2 Eq.1 uses Gemini as the high-tier VLM for R_visual (six defect types); §4.3 uses Gemini as Verify Agent in VRL; §5.1 uses Gemini as primary VLM-as-Judge for As/Ar. Thus training targets, repair policy, and headline metrics share one preference model. Table 2 (cross-VLM) and Fig.4/human checks mitigate but do not fully break the loop: other judges still score outputs already optimized under Gemini feedback, and human n is modest (200 agreement; 10 designers). If gains largely track Gemini-likeness rather than external diagram quality, the “reliably produces… above SOTA” claim weakens even if internal ablations (Table 3) remain valid.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces DocMeetSVG-100K, a 100K Text-to-SVG dataset for document/meeting diagrams, and GVR-Coder, a three-stage framework (curriculum rejection-sampling SFT with layout priors; GRPO with dual visual/complexity rendering rewards; and a generate–verify–repair agent loop) that turns long professional text into structured SVG. On a 320-sample held-out suite spanning Arena dialogues, meeting covers/topics, and rule-constrained docs, staged 14B/32B variants report large gains in average aesthetic score and perfect acceptance rate over traditional SVG models and strong open/closed LLMs, with ablations, cross-VLM re-scoring, designer ratings, a small user ranking study, and latency–quality analysis offered as support.","tokens_in":22274,"tokens_out":1802,"duration_ms":53890,"significance":"If the quality claims hold under more independent evaluation, the work is a solid systems contribution to multimodal code generation for a practically important office setting that existing icon-centric SVG datasets and methods underserve. Concrete strengths include a large, difficulty-stratified domain dataset with type-specific pipelines and partial human audit; an explicit anti-degeneration complexity reward alongside aesthetics; staged ablations (Table 3) and repair-transition analysis (Figure 8); cross-evaluator checks (Table 2); human designer scores (Table 4); and a promised code/data release. The combination of layout-constraint injection, dual rendering RL, and fine-grained repair is a coherent response to the three stated bottlenecks (data, priors, visual feedback).","major_comments":[{"comment":"Load-bearing independence of the Gemini-3-Pro oracle is not adequately established for the headline claim that GVR-Coder “outperforms competitive baselines” (Abstract; §5.2; Table 1). Gemini is used to generate and repair DocMeetSVG-100K candidates (§3), to supply R_visual via six defect types in RLDRF (§4.2, Eq. 1), as the Verify Agent in VRL (§4.3, Eq. 4), and as the primary VLM-as-Judge for As/Ar (§5.1). Table 2 and Fig. 4/human checks reduce pure self-grading but still largely re-score outputs already optimized under Gemini feedback; they do not show that training targets and rewards were built without that preference model. Please either (i) retrain or re-label a substantial reward/data slice with a held-out judge family and re-report Table 1, or (ii) clearly scope the claim as “improves under Gemini-aligned visual criteria” and lead with human and non-Gemini metrics.","section":"§3, §4.2 Eq. (1), §4.3, §5.1–5.2, Tables 1–2"},{"comment":"The strongest Table 1 numbers are for *-VRL variants (e.g., 32B-VRL Ar 100% on Meeting_cover_100, 88.3% on Svg_with_rules_120) compared to single-pass Gemini-3-pro and other LLMs. VRL’s Verify Agent is Gemini-3-Pro and may run up to three repair rounds (§4.3), so this is a multi-agent system with Gemini in the inference loop versus single-pass baselines—not a matched generator comparison. RSFT/RLDRF rows are the fair single-model evidence; VRL should be reported as a system result, with ablations that freeze or replace the verifier (open VLM / rule-only / human feedback) and a same-budget multi-sample or self-refine baseline for Gemini/GPT. Without that, “outperforms Gemini” overstates model quality relative to inference scaffolding.","section":"§4.3, §5.2 Table 1, Fig. 5"},{"comment":"R_complex (§4.2, Eq. 2) defines complexity as a count of structural tags (path, circle, line, …) matched to ≥0.8 of a reference SVG. This is a reasonable anti-simplification regularizer (and Fig. 14–15 help), but it is not shown to track logical/topological richness required by the paper’s own motivation (hierarchies, flow connectivity, factual structure). Tag inflation can raise R_complex without improving diagram logic. Please justify the 0.8 threshold, report sensitivity, and add at least one structure-aware check (e.g., node/edge counts by diagram type, connectivity validity, or correlation with factuality) so the dual-reward claim is not carried only by aesthetic VLM scores plus raw tag density.","section":"§4.2 Eqs. (2)–(3); Appendix Fig. 14–15"},{"comment":"Reliability claims rest on a 320-example test set and modest human studies (200 for judge agreement in Fig. 4; 10 designers in Table 4; 20 users on 50 texts). Given six primary categories and four named benchmarks, please report per-benchmark confidence intervals or bootstrap uncertainty for Ar/As, break down failures by defect type and diagram category, and clarify how much of the test distribution overlaps Gemini’s stylistic priors. The factuality protocol is binary (0/1 on hallucination/omission/logic) and is secondary in the narrative to aesthetic Ar; for “logically coherent” office diagrams, factuality should be tabulated with the same prominence as As/Ar for all main systems, not mainly via the radar summary (Fig. 7).","section":"§5.1–5.2, Fig. 4, Fig. 7, Table 4"}],"minor_comments":[{"comment":"Several bibliography entries and in-text model names appear contemporaneous or inconsistently versioned (e.g., GPT-5.x, Gemini-3, GLM-5, Kimi-k2.5). Ensure public, citable snapshots and dates so results are reproducible after model updates.","section":"References; Table 1"},{"comment":"Figure 9 and appendix qualitative figures note that high-resolution legible versions are deferred to the supplement; for the main camera-ready, at least one full-resolution case per benchmark should be readable in-page (text in nodes is part of the claim).","section":"§5.5 Fig. 9; Appendix H"},{"comment":"Curriculum mixture ratios (100% simple; 85/15; 85/15), rejection width of two candidates, VRL cap of three iterations, and unweighted R_total = R_visual + R_complex are free parameters; a short sensitivity subsection or appendix table would strengthen §5.4.","section":"§4.1–4.3; Table 3"},{"comment":"SVG Validity is listed as N/A for VRL rows in Table 1; briefly explain (post-repair validity assumed?) to avoid implying the metric was dropped when convenient.","section":"Table 1"},{"comment":"Minor prose/typos: “protoc al” → protocol (contributions); “A verage” spacing in Table 1 caption; “Staged Curriculum” figure text is dense—consider enlarging reward definitions in Fig. 3.","section":"§1 contributions; Table 1; Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The technical direction and dataset gap are real and suitable for a multimedia/ML systems venue. My recommendation is major_revision rather than reject because RSFT/RLDRF gains and human trends may survive de-circularized evaluation; the blocker is that the paper’s sharpest SOTA comparisons currently entangle data, reward, judge, and repair with one proprietary VLM. If the authors cannot re-label or swap judges, they should be required to deflate the claim language and foreground single-pass and human results. Mild novelty risk relative to concurrent SVG-LLM + RL work should be handled by sharper positioning against ReasonSVG/SVGen/OmniSVG on layout priors and office logic, not by the Gemini-heavy leaderboard alone."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent systems paper that fills a real gap—long professional prose to editable logic diagrams—and ships a domain dataset the field actually needed. The Gemini-in-the-loop issue is real but not a reason to dismiss the work.\n\nWhat’s new is less any single trick (curriculum rejection-SFT, GRPO, layout prompts, multi-agent repair are all known) and more the package aimed at logic-rich office SVGs: DocMeetSVG-100K, the dual aesthetic+complexity render reward so the model doesn’t win aesthetics by emptying the diagram, and the generate–verify–repair loop with attribute-level feedback. That complexity term is a smart, concrete fix for a failure mode you see constantly in render-reward setups.\n\nThey do the empirical work properly for this venue. Staged ablations (curriculum → knowledge → RLDRF → VRL), four-way test split, SVG validity, factuality, Sankey repair transitions, cross-VLM re-score, designer scores, and a small user ranking study. Gains track the stages in a believable way; first-pass validity >99% after RL is practically useful. Code/data promise helps.\n\nSoft spot, in proportion: Gemini is data factory, visual reward, verify agent, and primary judge, including head-to-heads that claim to beat Gemini. Table 2 and humans blunt pure self-grading, but they don’t fully break preference capture—other judges still score Gemini-optimized outputs, and human n is modest. So treat the 100% / 82–90% Ar numbers as upper bounds under that preference model, not fully external truth. Internal ablations remain informative either way. Free knobs (0.8 complexity threshold, unweighted reward sum, 3 repair rounds) are ordinary engineering, not hidden load-bearing math.\n\nWho it’s for: people building diagram codegen, meeting/document tools, or SVG-LLMs. Not a core-ML reorg. I’d bring it to a multimodal/systems reading group if that is the theme; I’d cite the dataset and the dual-reward idea if I work in this lane. It deserves a serious referee, not a desk reject—ask for stronger judge-independence and clearer open reproduction of the labeling path, not a rewrite of the premise.","headline":"Solid office text-to-SVG systems paper with a real domain dataset; headline wins over Gemini are partly Gemini-shaped, but the stack and ablations still hold up.","tokens_in":22983,"tokens_out":573,"would_cite":true,"duration_ms":16443,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"GVR-Coder turns long professional texts into coherent, editable SVG diagrams by combining layout-aware curriculum training, dual rendering rewards, and a generate-verify-repair loop.","keywords":["Large Language Models","Scalable Vector Graphics","Text-to-SVG","Reinforcement Learning","Visual Feedback","Document Diagrams","Curriculum Learning","Multi-Agent Repair"],"falsifier":"Have independent human designers who never see model names perform a fully blind usability-and-fidelity scoring of the 320-test-set diagrams for GVR-Coder versus the strongest closed-source baseline; if GVR’s acceptance-rate advantage disappears or reverses under pure human judgment, the central claim does not hold.","tokens_in":22759,"feed_emoji":"📊","tokens_out":999,"duration_ms":43260,"temperature":0.7,"pith_summary":"Dense meeting notes and office documents hide their logic under long prose, raising cognitive load for anyone who must extract structure. This paper claims Scalable Vector Graphics are the right automatic target because they stay resolution-independent and editable, yet prior text-to-SVG work fails on real office diagrams: suitable data are scarce, models lack layout priors, and nothing ever looks at the rendered picture to fix defects. The authors release DocMeetSVG-100K, a 100,000-pair dataset of document and meeting diagrams graded by difficulty, and introduce GVR-Coder. The system first teaches simple-to-hard structure with explicit layout constraints, then uses reinforcement learning on dual rewards that jointly push visual cleanliness and structural richness, and finally runs a generate-verify-repair agent loop that diagnoses concrete defects and rewrites the code. On held-out meeting, dialogue, and rule-constrained tests, the resulting 14B–32B models reach higher rates of fully usable diagrams than larger closed-source systems while staying factually faithful. A reader who works with long documents cares because the output is something a colleague can skim, edit, and drop into a slide without redrawing.","feed_headline":"Office texts become clean SVG diagrams via visual feedback","feed_subtitle":"Layout rules, dual rewards, and a repair loop let mid-size models beat larger systems on meeting and document figures.","key_machinery":"The GVR-Coder three-stage stack: CRSFT (curriculum rejection-sampling fine-tuning plus explicit layout constraints), RLDRF (GRPO reinforcement learning with a hybrid visual-plus-complexity reward computed on rendered images), and VRL (an inference-time generate-verify-repair multi-agent loop that supplies fine-grained defect attribution and targeted code repair).","core_discovery":"GVR-Coder—curriculum-driven rejection-sampling fine-tuning with injected layout knowledge, reinforcement learning from dual rendering feedback that balances aesthetics against structural complexity, and a generate-verify-repair agent loop—converts lengthy professional text into logically coherent and visually clean SVG diagrams and outperforms competitive baselines, including larger closed-source models, on the authors’ DocMeetSVG benchmarks.","pith_inferences":["The same generate-verify-repair loop could transfer to other vector or declarative diagram formats once a stable renderer and a shared defect taxonomy exist.","Heavy reliance on a single proprietary verifier at train and test time points to a natural next step: distilling a lightweight open verifier so the loop no longer needs an external oracle at inference.","If the reported cognitive-load reduction replicates in larger workplace field studies, the pipeline could become standard post-processing for meeting-transcription and document-authoring tools."],"forward_implications":["Meeting minutes and long documents can be turned automatically into editable flowcharts, covers, and topic maps without manual layout work.","Open models in the 14B–32B range trained this way can match or exceed much larger closed models on structured office-diagram tasks.","Dual rewards that penalize both visual defects and oversimplification keep diagrams both clean and information-dense.","Injecting layout constraints plus iterative visual repair becomes a reusable pattern for other code-that-must-look-right generation problems.","DocMeetSVG-100K and the aesthetic-plus-factuality protocol become a shared benchmark for logic-rich SVG generation."],"fun_headline_variants":["GVR-Coder turns long office text into coherent SVG diagrams","Layout priors and dual render rewards clean up Text-to-SVG","Generate-verify-repair loop yields logical meeting diagrams","Curriculum fine-tuning plus visual feedback beats larger SVG models","DocMeetSVG-trained coder maps verbose docs to editable figures"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The whole pipeline treats one high-tier vision-language model as a reliable oracle for building the training set, labeling aesthetic rewards, scoring final quality, and driving repairs, so that large gains against that same model are taken as genuine diagram quality rather than style matching.","fun_headline_variants_meta":{"raw":{"variants":["GVR-Coder turns long office text into coherent SVG diagrams","Layout priors and dual render rewards clean up Text-to-SVG","Generate-verify-repair loop yields logical meeting diagrams","Curriculum fine-tuning plus visual feedback beats larger SVG models","DocMeetSVG-trained coder maps verbose docs to editable figures"]},"model":"grok-4.5","effort":"low","cost_usd":0.00381,"raw_usage":{"total_tokens":1242,"prompt_tokens":856,"num_sources_used":0,"completion_tokens":86,"cost_in_usd_ticks":38104000,"prompt_tokens_details":{"text_tokens":856,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":300,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":856,"tokens_out":86,"duration_ms":6741,"temperature":1.0,"reasoning_tokens":300,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T18:40:09.087399+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Have independent human designers who never see model names perform a fully blind usability-and-fidelity scoring of the 320-test-set diagrams for GVR-Coder versus the strongest closed-source baseline; if GVR’s acceptance-rate advantage disappears or reverses under pure human judgment, the central claim does not hold.","supporting_citations":[],"review_version":1}