{"id":"bf542e3a-6cb5-45b1-bd92-8b83492a3ab1","arxiv_id":"2607.07105","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"CompoVista introduces a four-layer Composition Graph and canvas-based visual analytics system enabling structure-aware retrieval, cohort comparison, and evidence inspection for Traditional Chinese Paintings.","lead":"The paper builds a visual analytics system, CompoVista, that lets art historians query and compare the composition of Traditional Chinese Paintings at scale using a four-layer scene graph. A generalist might read it to see how domain-specific visual structure can be made computationally searchable for cultural heritage analysis.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Matching function (Eq. 3) is the computational foundation of the entire system, yet its weights, component scoring, and threshold are unspecified with no quantitative retrieval evaluation — case study findings cannot be independently verified.","rationale":"The reader correctly identified the unspecified weights in Eq. (3) as the weakest assumption. I sharpen this to note that it is not only the weights but the entire scoring pipeline — component functions S_c, threshold θ, and absence of any quantitative retrieval evaluation — that constitutes the load-bearing gap. The paper's evaluation (user study, case studies, expert interviews) assesses the system's usability and perceived usefulness, which is appropriate for a VA/HCI paper. However, the paper also makes a technical claim about structure-aware retrieval (Sec. V-B) and presents case study findings that depend on retrieval correctness. These findings cannot be verified without parameter specification or quantitative metrics. This does not invalidate the contribution — the Composition Graph representation, the visual design, and the workflow are legitimate contributions — but it does mean the retrieval-dependent claims (especially the case study pattern discoveries) are not fully supported. The verdict should remain CONDITIONAL: specify matching parameters, provide quantitative retrieval evaluation against expert ground truth, and ideally release code/data for independent verification. The paper's own limitations section (Sec. VIII-C) acknowledges corpus size and relation sparsity but does not address the scoring pipeline opacity, suggesting the authors may not recognize this as a gap.","tokens_in":21978,"tokens_out":2231,"duration_ms":67187,"concrete_test":"Construct 20–30 queries with ground-truth relevance judgments from the two art historians (A1, A2). For each query, compute precision@10 and recall@100 under at least three weight configurations: uniform (λ_c = 0.25 each), entity-dominant (λ_entity = 0.7), and relation-dominant (λ_relation = 0.5). If precision@10 varies by more than 20 percentage points across configurations for the same queries, the case study findings are not robust to weight choices and the retrieval pipeline needs explicit justification or quantitative evaluation before the central claim is fully supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that CompoVista supports composition-oriented cohort construction, pattern discovery, and evidence inspection. Every downstream visualization — Distribution View, Relations View, Comparison View, Detail Node — depends on which paintings the matching function S(Q, G_p) retrieves into a cohort. Eq. (3) defines this as a weighted sum of component scores S_c over entity, relation, layout-region, and void-space constraints, with weights λ_c summing to 1. However: (1) the weights λ_c are never specified — not whether they are uniform, hand-tuned, or learned; (2) the component scoring functions S_c ∈ [0,1] are never defined — it is unclear what similarity metric is used for entity matching, relation matching, or void-space overlap; (3) the threshold θ for cohort membership is never given. Both case studies depend on retrieval working correctly: Case 1's 'Ma Yuan corner formula' structural query and Case 2's 'sitting on boat' vs 'standing on boat' relation queries produce cohorts whose compositional patterns are then interpreted as art-historical findings. Without knowing the scoring parameters or having any quantitative retrieval evaluation (precision, recall, baseline comparison), these results cannot be reproduced or distinguished from artifacts of parameter tuning. The paper acknowledges no quantitative metrics were collected for retrieval quality.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces CompoVista, a visual analytics system for compositional analysis of Traditional Chinese Paintings (TCPs). The core contribution is the Composition Graph, a four-layer structured representation (entities, relations, void space, context) built on scene-graph foundations. The system supports format-aware cohort construction through visual queries, cohort-level summarization (Distribution View, Relations View), cross-cohort comparison (Comparison View), and painting-level evidence inspection. The work is grounded in a formative study with two art historians and evaluated through a task-based user study (12 participants), two case studies, and expert interviews. The case studies demonstrate the system's use in discovering compositional patterns, such as Ma Yuan's corner formula and fisherman motif variations across dynasties.","tokens_in":22372,"tokens_out":1437,"duration_ms":167284,"significance":"The paper addresses a genuine gap in computational art history: existing TCP computational methods focus on style, attribution, or object presence rather than spatial composition, and existing scene-graph methods are designed for natural images without TCP-specific concepts like void space and format-aware placement. The Composition Graph representation and the integrated VA workflow connecting structured retrieval, cohort comparison, and evidence inspection are well-motivated by the formative study. The system design is thoughtful, with domain-appropriate visual encodings (calligraphic entity labels, paper-like textures, perceptually grounded color palettes). The evaluation with 12 domain-aware participants provides reasonable qualitative evidence for usability. However, the computational foundation of the retrieval pipeline is underspecified, which limits the reproducibility and verifiability of the demonstrated findings.","major_comments":[{"comment":"§V-B, Eq. (3): The matching function S(Q, G_p) is the computational foundation of the entire system—all downstream visualizations depend on which paintings are retrieved into a cohort. However, the component scoring functions S_c(Q, G_p) ∈ [0,1] are never defined. It is unclear what similarity metric is used for entity matching (bounding-box IoU? center distance?), relation matching (exact match? partial credit?), or void-space overlap. Without these definitions, the retrieval pipeline cannot be reproduced or independently verified. The paper should specify each S_c and provide at least a minimal quantitative evaluation of retrieval quality (e.g., precision@k against expert-judged relevance for a small set of queries).","section":null},{"comment":"§V-B, Eq. (3): The weights λ_c are stated to sum to 1 but are never specified as uniform, hand-tuned, or learned. Both case studies depend on retrieval working correctly: Case 1's 'Ma Yuan corner formula' structural query (§VII-B) and Case 2's 'sitting on boat' vs 'standing on boat' relation queries produce cohorts whose patterns are then interpreted as art-historical findings. If the weights were tuned post-hoc to produce desirable results, the case study findings could be artifacts of the weighting rather than robust compositional matches. The paper should disclose the weight settings and justify them, or demonstrate that findings are robust to weight perturbation.","section":null},{"comment":"§V-A2 and §VIII-C: The relation evidence is extremely sparse—1,197 relation instances across 960 paintings, averaging approximately 1.25 relations per painting. This raises a correctness-risk concern for the Relations View and relation-level comparison, which are central to Case 2's findings about fisherman postures. Some relations may be absent because they were not annotated rather than because they have no visual meaning. The paper acknowledges this in §VIII-C but does not address how it affects the reliability of the case study interpretations. The authors should either (a) report the relation density within the specific cohorts used in the case studies, or (b) add a caveat that relation-level findings are preliminary given annotation sparsity.","section":null},{"comment":"§VII-A, Fig. 8B: The user study reports that only 5/12 participants agreed or strongly agreed that the system helped compare compositional tendencies across cohorts (Q4), and only 6/12 for turning findings into research clues (Q8). Cross-cohort comparison is a central claimed contribution (T3, R2). The paper should discuss whether this indicates a fundamental limitation of the Comparison View's design or a usability issue that could be addressed, and whether the comparison claim should be scoped more narrowly given these results.","section":null}],"minor_comments":[{"comment":"§V-B: The threshold θ for cohort membership is mentioned but never given a value or a method for selecting it. Please specify.","section":null},{"comment":"§V-A2, Table I: The 'Other/Unclassified' theme category (79 paintings) and 'Other or unspecified' format category (93 paintings) are non-trivial fractions of the corpus. Please clarify whether these are included in analysis or excluded.","section":null},{"comment":"§V-A2, Fig. 4: The void-space extraction procedure uses ink-density and texture-density cues, but the specific thresholds or parameters for selecting candidate regions are not given. Please specify or cite the method.","section":null},{"comment":"§VI-B: The Distribution View uses font size to encode average spatial size and opacity to encode frequency. This is an unconventional encoding that may be confusing—please justify why this is preferable to the standard word-cloud convention or provide evidence that users interpreted it correctly.","section":null},{"comment":"§VII-B, Case 1: P3's finding that Ma Yuan's corner formula 'could carry a small courtly scene' is presented as a hypothesis. The paper should be clearer that this is an interpretive suggestion from the participant, not a system-verified finding.","section":null},{"comment":"§II-C: The related work on visual analytics for cultural heritage could cite more recent work on interactive cohort construction and provenance management in VA systems.","section":null},{"comment":"§VIII-C: The exclusion of handscrolls from normalized spatial analysis is acknowledged, but 44 handscrolls are in the corpus (Table I). Please clarify whether these paintings appear in any analysis or are entirely excluded from spatial views.","section":null},{"comment":"Fig. 6 and Fig. 7: The visual encoding descriptions are detailed but the figures themselves are difficult to parse at the resolution provided. Consider adding annotated close-ups for key interactions.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the matching function is well-founded and is the primary reason for the major revision recommendation. The paper's contributions are primarily in the VA design and representation layers, and the qualitative evaluation is reasonable for a VA paper. However, the complete absence of specification for the scoring function, weights, and threshold—combined with no quantitative retrieval evaluation—means the case study findings cannot be independently verified. This is fixable: the authors need to specify the scoring components, disclose parameter settings, and provide at least a small-scale quantitative evaluation. The relation sparsity issue is also worth flagging to the editor as it affects the reliability of one of the two case studies. The paper is otherwise well-written and well-motivated, and I would expect it to be acceptable after revision."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"This paper builds a visual analytics system for compositional analysis of Traditional Chinese Paintings, and the core idea is genuinely good. The Composition Graph extends scene graphs with void space and context layers specific to TCP art theory, and the integrated workflow—query, cohort construction, comparison, evidence inspection—is well-matched to how art historians actually work. The two case studies (Ma Yuan's corner formula, fisherman postures) are illustrative and show the system can surface real compositional patterns. The user study with 12 domain-aware participants is thorough for this venue. The visual design choices—format-normalized coordinate space, calligraphic entity labels, paper texture—are thoughtful and domain-appropriate. Credit earned here: the representation and workflow are the real contribution, and they hold up.","headline":"Solid VA system for TCP compositional analysis; the matching pipeline is a black box that needs opening","tokens_in":22704,"tokens_out":216,"would_cite":false,"duration_ms":52094,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Making Chinese painting composition computationally searchable","keywords":["visual analytics","Traditional Chinese Painting","composition analysis","scene graph","cohort comparison","digital humanities","void space","structure-aware retrieval"],"falsifier":"If the matching function's weights are varied across a reasonable range and the resulting cohorts change substantially — meaning different weightings produce different art-historical patterns — then the case-study findings are artifacts of parameter choice rather than discoveries about TCP composition.","tokens_in":22072,"feed_emoji":"🖼️","tokens_out":1644,"duration_ms":86016,"temperature":0.7,"pith_summary":"The paper claims that the composition of a Traditional Chinese Painting can be captured as a four-layer graph — entities (painted objects with positions), relations (spatial and semantic connections between them), void space (deliberately reserved blank regions), and context (artist, period, format, medium) — and that this representation, called the Composition Graph, makes it possible to retrieve, summarize, and compare paintings by their compositional structure rather than by metadata labels or overall visual similarity. Built on this representation, the CompoVista system lets art historians draw a visual query (placing entities, drawing relation edges, marking voids, filtering by context), retrieve a cohort of structurally matching paintings, inspect where entities tend to sit and which relations recur, compare cohorts side by side, and drill back into individual paintings to verify that an aggregate pattern holds at the level of a single work. The paper evaluates this through a task-based study with twelve domain-aware participants, two case studies (one tracing a compositional shift from Northern to Southern Song landscapes and the Ma Yuan corner formula; another separating fisherman scenes by posture into different narrative moods), and expert interviews. The central claim is that composition, long treated as qualitative and unscalable, becomes a queryable, comparable, and inspectable analytical object once the four-layer graph is in place.","feed_headline":"Graphing the blank space in Chinese painting","feed_subtitle":"A four-layer scene-graph representation makes compositional structure in Traditional Chinese Paintings queryable, comparable, and traceable,","key_machinery":"Composition Graph G_p = (V_p, E_p, Z_p, M_p) with entity, relation, void-space, and context layers; weighted matching function S(Q, G_p) scoring query components; format-aware coordinate normalization for spatial comparison; canvas-based node interface linking query, cohort, comparison, and detail views","core_discovery":"The Composition Graph extends scene-graph representation to TCP-specific compositional structure by adding two layers that generic scene graphs lack: a void space layer that treats reserved blank regions as first-class spatial primitives, and a context layer that anchors spatial and relational patterns to art-historical metadata. This four-layer structure, combined with a coarse-to-fine weighted matching function, enables structure-aware retrieval where a user can search for paintings sharing a compositional pattern (e.g., dense elements in one corner paired with a large void opposite) rather than a keyword or a visual look-alike. The system then normalizes retrieved cohorts into format-comb","pith_inferences":["The four-layer decomposition is testable for inter-annotator reliability: if different annotators produce substantially different void-space or relation layers for the same paintings, the representation's stability as an analytical object is uncertain, and cohort comparisons may reflect annotation conventions rather than compositional fact.","The matching function's sensitivity to weight settings could be stress-tested directly: if varying the lambda weights shifts cohort membership substantially, the case-study patterns (e.g., Ma Yuan's corner formula retrieval) may be artifacts of a particular weighting rather than robust structural matches.","The exclusion of handscrolls from normalized spatial comparison suggests that sequential-viewing formats may require a fundamentally different compositional representation — perhaps a fifth layer encoding viewing sequence or temporal progression — which the paper acknowledges but does not resolve."],"forward_implications":["Composition-specific structured representations could be ported to other painting traditions where spatial arrangement carries art-historical meaning, such as Renaissance altarpieces or Japanese screen paintings, by redefining the entity vocabulary and relation types while keeping the four-layer architecture.","Treating void space as a queryable primitive rather than mere background could shift how computational art history models negative space across all visual traditions, making absence a searchable structural feature.","The cohort-as-provisional-evidence-space concept — where a working set is formed, inspected, revised, and dissolved during analysis — could generalize to other humanities domains where scholars iterate between distant viewing and close reading.","If the matching function proves robust to weight variation, the approach could support automated discovery of compositional formulas across a collection, surfacing structural repetitions that no scholar has yet named."],"fun_headline_variants":["Void space becomes queryable in Chinese painting analysis","Scene graphs model blank space in Traditional Chinese Paintings","Four-layer graph makes painting composition searchable","Treating blank space as structure in Chinese painting analysis","Composition graph enables pattern search across painting collections"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The retrieval pipeline depends on a weighted matching function whose component weights are not specified as fitted, uniform, or hand-tuned, so the demonstrated cohort patterns may reflect a particular weighting choice rather than robust compositional similarity.","fun_headline_variants_meta":{"raw":{"variants":["Void space becomes queryable in Chinese painting analysis","Scene graphs model blank space in Traditional Chinese Paintings","Four-layer graph makes painting composition searchable","Treating blank space as structure in Chinese painting analysis","Composition graph enables pattern search across painting collections"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":684,"prompt_tokens":631,"completion_tokens":53,"prompt_tokens_details":null},"tokens_in":631,"tokens_out":53,"duration_ms":58253,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T19:59:08.983884+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the matching function's weights are varied across a reasonable range and the resulting cohorts change substantially — meaning different weightings produce different art-historical patterns — then the case-study findings are artifacts of parameter choice rather than discoveries about TCP composition.","supporting_citations":[],"review_version":1}