{"id":"23be2370-a8c8-42cf-a99a-fce9ae2abf0c","arxiv_id":"2611.01860","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified taxonomy and automated survey system organizes the field of LLM-based agentic computer graphics, enabling cross-task comparison and trend analysis.","lead":"Advances in Agentic Computer Graphics surveys the growing literature on LLM agents that automate graphics tasks. It proposes a four-dimensional taxonomy and an agentic literature survey system, analyzing over 300 papers to reveal design patterns and trends.","discovery_kind":"review","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Trend analyses in Sec 4.2.2 rely on automated taxonomy labels whose correctness is not verified; the reported percentages may be artifacts of classification noise.","rationale":"The reader identified the reliability of automated classifications as a weakest assumption. I agree that this is a critical vulnerability and sharpen it to a specific, testable deficiency: the quantitative trend analyses in Sec 4.2.2 are unvalidated. The taxonomy itself is a valuable conceptual contribution, and the automated system is presented as assistive, so the paper does not fail entirely. However, the paper makes explicit empirical claims (e.g., percentages and trends) that are not supported by evidence of classification accuracy. A conditional accept that requires validation of a random sample (or explicit softening of the claims) would strengthen the paper's empirical grounding without rejecting its main organizational contribution.","tokens_in":49026,"tokens_out":2108,"duration_ms":24385,"concrete_test":"Randomly sample 50 papers from the corpus; have two domain experts independently assign all relevant taxonomy labels (dimensions in Sec 3). Compute per-category Cohen's kappa between experts and between each expert and the automated system. Report precision and recall for the automated labels using the majority expert vote as ground truth. If mean per-category agreement (expert vs automated) falls below 75% for any category appearing in more than 10% of the sample, the trend claims in Sec 4.2.2 should be caveated as preliminary and the paper should release the validation data.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest empirical contribution—trend analyses showing, e.g., local patching rising from 50.0% to 83.0% (Fig 5b.1)—is derived entirely from taxonomy labels assigned by the automated literature survey system. The paper explicitly states that the system's curation checks \"do not verify the correctness of the classifications\" (sentence b238). Manual verification was limited to paper relevance (Sec 4.2.1), not to the taxonomy labels themselves. Without ground-truth validation, classification errors of unknown magnitude propagate into every quantitative claim in Sec 4.2.2. For example, if the system systematically mislabels 'local patch' vs 'regeneration' for a subset of papers, the observed temporal trend could be spurious. The paper's central assertion—that the taxonomy enables systematic cross-task comparison and trend discovery—depends on the reliability of these automatically generated numbers. The absence of any validation protocol (e.g., human inter-annotator agreement, per-category precision/recall) is a significant gap that weakens the empirical support for this claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper presents a survey of agentic computer graphics systems, organized around a unified taxonomy with four dimensions: specification, base model, harness, and graphics interface and tooling. To keep pace with the rapidly growing literature, the authors also develop an agentic literature survey system that automatically discovers papers, assigns taxonomy labels, and curates structured reports. Using this system on a corpus of 350 papers, the paper reports quantitative trends (e.g., update strategies, resource usage) and synthesizes design patterns across 2D, static 3D, and dynamic 3D tasks. The survey concludes by discussing open problems including multi-agent collaboration, action-space design, and evaluation challenges.","tokens_in":49401,"tokens_out":2275,"duration_ms":26148,"significance":"The taxonomy provides a well-structured vocabulary that can unify descriptions across the fragmented agentic graphics literature. The automated survey system is a novel methodological contribution that, if made reliable, could lower the barrier to keeping surveys current. The qualitative synthesis of design patterns (Secs. 6–8) is valuable for researchers entering the field. If the quantitative trend claims can be validated, the paper would also offer a data-driven view of the field's evolution.","major_comments":[{"comment":"The quantitative trend analyses (e.g., 'local patching rising from 50.0% to 83.0%') are derived entirely from taxonomy labels assigned by the automated literature survey system. The paper explicitly states that curation checks 'do not verify the correctness of the classifications' (sentence b238). Manual verification was limited to paper relevance (Sec. 4.2.1), not to taxonomy labels. Without any validation of label accuracy (e.g., human inter-annotator agreement, per-category precision/recall), the reported percentages may reflect classification noise rather than genuine trends. This is a load-bearing issue because the trend analysis is a central empirical contribution of the paper. The authors should either (a) provide a validation study on a representative sample, or (b) clearly reframe these figures as exploratory and remove the quantitative precision (e.g., replace exact percentages with qualitative ranges).","section":"Sec. 4.2.2, Fig. 5","authors_acknowledge":"We note that while these checks identify missing or inconsistent entries, they do not verify the correctness of the classifications.","authors_acknowledge_p":0.77},{"comment":"The taxonomy is treated as a complete design space covering all agentic graphics systems. While the paper notes in Sec. 10 that categories can be expanded, the analysis in Sec. 4.2.2 relies on the current fixed set of dimensions and options. The paper would benefit from a more explicit discussion of potential blind spots—e.g., whether the four dimensions adequately capture systems that involve mixed-initiative human–agent collaboration or that use non-LLM decision-making components alongside LLMs. This could be addressed by adding a limitations paragraph to Sec. 3.","section":"Sec. 3, overall taxonomy","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The agentic literature survey system is described at a high level, but key details needed for reproducibility are omitted: the exact prompt template for taxonomy annotation, the model configuration (e.g., temperature, number of reasoning steps), the tools used for paper reading and evidence extraction, and the criteria for the 'insufficient supporting evidence' flag. The paper states that the system will be released publicly, but until then the annotation process is not verifiable. The authors should provide a more detailed technical appendix or release the system as part of the supplementary material.","section":"Sec. 4.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The font in Fig. 2b is very small; consider enlarging it or using a legend that lists the terms in a table for readability.","section":"Sec. 2, Fig. 2b","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'These types differ in how much of an operation the agent must specify and how much is handled by the execution environment' is a helpful summary, but it could be made more precise by adding a short example contrasting code/script vs. tool/function call.","section":"Sec. 3.4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The tables are dense and use many icons. Consider adding a one-sentence caption that guides the reader to the most important pattern in each table (e.g., 'most methods use off-the-shelf VLMs with local patch updates').","section":"Tables 2, 3, 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The subfigures (a.1, a.2, a.3) are labeled only in the caption; label them on the figure itself for easier reference.","section":"Sec. 4.2.2, Fig. 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The discussion of 'creativity and attribution' is welcome but brief. Consider adding a reference to recent work on human–AI co-creation to strengthen the context.","section":"Sec. 9","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The phrase 'these generators fall outside our scope' is clear, but it could be more explicit about why: they do not use an LLM to decide actions, as stated earlier. Repetition here is fine, but the justification should be consistent.","section":"Sec. 2.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hi colleague,\n\nHere's my take on the agentic graphics survey (2611.01860).\n\nThe paper's main contribution is a clean four-dimensional taxonomy—specification, base model, harness, graphics interface and tooling—that makes it possible to compare systems that use different terminology. The survey covers a huge range of recent work (350 papers) and the application sections (2D, static 3D, dynamic 3D) are thorough. The open problems discussion is thoughtful. The paper is clearly written.\n\nWhat's good: the taxonomy is the lasting contribution. The authors also built an automated literature survey system that classifies papers using their taxonomy. As a way to keep up with a fast-growing field, the idea is sound. The manual verification of paper relevance is a step toward quality control.\n\nWhat's soft: the trend analysis in Section 4.2.2—the main empirical claim—rests entirely on taxonomy labels assigned by the automated system. The paper explicitly says (sentence b238) that the curation checks \"do not verify the correctness of the classifications.\" Manual verification was limited to relevance, not the labels themselves. So numbers like local patching rising from 50% to 83% could reflect classification noise. The stress-test note on this is accurate. Without inter-annotator agreement or per-category precision/recall, these percentages are suggestive but not robust.\n\nThat said, the taxonomy and qualitative comparisons survive this issue. The paper acknowledges the limitation. If the authors want the trend analysis to carry weight, they need to validate the labels. For now, I'd treat those numbers as rough indicators.\n\nWho this is for: anyone entering agentic graphics, or anyone who wants a structured map of the field. The taxonomy is worth citing. I'd accept peer review, but would ask for either label validation or a significant softening of the quantitative claims.\n\nBest,\n[Your Name]","headline":"A well-organized taxonomy for agentic graphics, but the trend analysis relies on unverified automatic labels and should be treated as suggestive, not conclusive.","tokens_in":49851,"tokens_out":1695,"duration_ms":21775},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A unified four-dimensional taxonomy organizes the rapidly growing field of agentic computer graphics, supported by an automated literature survey system that scales with the literature.","keywords":["agentic computer graphics","LLM agents","taxonomy","design space","literature survey","automated annotation","2D and 3D graphics","feedback mechanisms"],"falsifier":"If a concrete agentic graphics system (e.g., a diffusion-based agent that edits 3D scenes without any explicit harness or memory) cannot be placed into the taxonomy without stretching the definitions beyond recognition, the taxonomy's coverage would be shown incomplete. Likewise, if two systems that designers consider fundamentally different receive identical taxonomy labels, the granularity would need revision.","tokens_in":50570,"feed_emoji":"📐","tokens_out":2330,"duration_ms":27251,"temperature":0.7,"pith_summary":"This survey paper argues that research on LLM-based agentic graphics systems—where a language model writes code, runs it in a graphics environment, and revises based on feedback—has grown so fast and with such inconsistent terminology that manual understanding is no longer sufficient. The authors propose a unified taxonomy that classifies any such system along four dimensions: specification, base model, harness, and graphics interface and tooling. To demonstrate and deploy this taxonomy, they build an agentic literature system that automatically discovers relevant papers, reads them, assigns taxonomy labels, and curates structured reports. Using this system on 350 papers, they identify several design patterns, such as the dominance of visual feedback (77 % of systems) and a shift toward local patch updates (rising to 83 % over the analyzed period). The paper closes with open challenges around multi-agent coordination, long-horizon planning, human–agent alignment, and reliable evaluation.","feed_headline":"Four dimensions organize the explosive growth of AI-driven graphics agents","feed_subtitle":"A unified taxonomy and an automated literature system let researchers compare dozens of systems across 2D, CAD, and dynamic 3D tasks on…","key_machinery":"The four-dimensional taxonomy (specification, base model, harness, graphics interface and tooling) is the central organizing framework; it provides a common vocabulary for describing and comparing agentic graphics systems across inconsistent terminologies. The agentic literature survey system, built on top of a SWE-agent-style harness with PaperQA2-style evidence extraction, operationalizes the taxonomy by reading full papers, assigning labels, and curating structured records.","core_discovery":"The paper establishes that the design space of agentic graphics systems can be captured by four axes—specification (task, input, output), base model (type and training), harness (planning, feedback, memory), and graphics interface and tooling (action type, software, representation, tools, resources)—each with concrete sub-options. Using this taxonomy, the authors find that most systems rely on off-the-shelf base models, harness designs are highly diverse, code and structured specifications dominate action types, and feedback patterns vary strongly by task domain (e.g., CAD favors full regeneration while scene generation favors local patches). The automated survey system itself is a demonstration that large-scale, taxonomy-driven literature curation can be performed by an agent, though the authors note that classification correctness is not systematically verified.","pith_inferences":["The taxonomy's four dimensions are a design choice, not a theorem; a different partitioning (e.g., separating feedback from planning more sharply) might prove equally useful, and the paper invites such refinements as new approaches emerge.","The automated system's classification accuracy is unverified; a human validation study comparing its labels against expert annotations would be a natural next step before relying on trend numbers.","If the taxonomy is adopted widely, it could enable meta-analyses that cross-task comparisons of architecture choices (e.g., which harness patterns correlate with success on dynamic vs. static tasks) become routine.","The open challenge of evaluating interactive graphics systems is particularly acute; the paper notes stochastic runs and proprietary models make disentangling system design from model capability difficult, implying that a shared evaluation protocol would greatly accelerate progress."],"forward_implications":["Researchers can use the taxonomy to position their own system design explicitly, enabling apples-to-apples comparisons with prior work even when different application terms are used.","The observed shift toward local patch updates suggests that future system designs should prioritize targeted edits over full regeneration to improve efficiency and preserve earlier progress.","The dominance of visual feedback points to an increasing reliance on VLMs as critics; future work may need to address VLM blind spots in geometric and physical reasoning.","The automated survey system, if publicly released as planned, could become a living bibliography that keeps pace with the field's rapid growth, reducing the lag of manual reviews."],"supporting_citations":[{"why":"Provides the general definition of language model agents that the paper adopts and builds upon.","marker":"[XCG*23]"},{"why":"Supplies the ReAct framework for iterative reasoning and acting, a core pattern in agentic graphics harnesses.","marker":"[YZY*23]"},{"why":"Contributes the SWE-agent tool-use harness that the paper's automated literature system uses for reading and annotating papers.","marker":"[YJW*24]"},{"why":"Provides the PaperQA2 evidence-gathering methodology that the literature system adapts for evidence-based annotation.","marker":"[SCL*24]"},{"why":"Exemplifies a fully specified agentic graphics system (Articraft) whose design is used to illustrate the taxonomy's coverage.","marker":"[ZLL*26b]"}],"fun_headline_variants":["A unified taxonomy for agentic computer graphics","Automated literature review of graphics agents","Four axes define AI graphics agent design","Mapping the landscape of AI-driven graphics agents","Agentic systems: a new paradigm for computer graphics"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The four dimensions and their subcategories capture the essential design choices of all agentic graphics systems sufficiently to enable meaningful cross-task comparison; the paper acknowledges this is a design choice that may need expansion as new approaches appear.","fun_headline_variants_meta":{"raw":{"variants":["A unified taxonomy for agentic computer graphics","Automated literature review of graphics agents","Four axes define AI graphics agent design","Mapping the landscape of AI-driven graphics agents","Agentic systems: a new paradigm for computer graphics"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":8.6e-05,"raw_usage":{"cost":8.62155e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":8.62155e-05,"upstream_inference_prompt_cost":4.503198e-05,"upstream_inference_completions_cost":4.118352e-05},"total_tokens":1313,"prompt_tokens":901,"completion_tokens":412,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":345}},"tokens_in":901,"tokens_out":412,"cost_source":"openrouter_usage","duration_ms":5598,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":345,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-04T19:08:18.602700+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If a concrete agentic graphics system (e.g., a diffusion-based agent that edits 3D scenes without any explicit harness or memory) cannot be placed into the taxonomy without stretching the definitions beyond recognition, the taxonomy's coverage would be shown incomplete. Likewise, if two systems that designers consider fundamentally different receive identical taxonomy labels, the granularity would need revision.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}