{"id":"a96ce1a1-ffe4-45ae-8637-f3a836b86aae","arxiv_id":"2507.11261","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ViewSRD improves 3D visual grounding by decomposing multi-anchor language queries and adding learned view tokens to align text and point clouds across perspectives.","lead":"ViewSRD is a new framework for 3D visual grounding that decomposes complex multi-anchor queries into simpler single-anchor statements and aligns text with scenes across multiple viewpoints using shared view tokens. The authors report state-of-the-art accuracy on Nr3D, Sr3D, and ScanRefer, with the largest gains on complex spatial queries.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ScanRefer inconsistency undermines SOTA claim: full ViewSRD (45.4 Acc@0.25) trails BUTD-DETR+SRD (57.9) by 12.5 points, so Table 3 suggests the complete framework is worse than a baseline plus one component.","rationale":"The reader's weakest_assumption identifies the SRD decomposition losing overlapping relationships, and the paper itself acknowledges this limitation. That is a legitimate concern, but it is a stated limitation about language understanding rather than a contradiction in the empirical record. The ScanRefer inversion in Table 3 is more serious because it is an internal inconsistency: under the architecture described, the full ViewSRD should outperform BUTD-DETR+SRD, not trail it by 12.5 points. This discrepancy directly affects the strongest_claim because state-of-the-art is claimed across 3DVG datasets, and ScanRefer is one of the two benchmarks used. The reader's rationale already noticed this ('the full ViewSRD model achieves 45.4 ... while the SRD-augmented BUTD-DETR reaches 57.9'), so my agreement is partial rather than complete: we flag the same table, but I elevate it to the primary load-bearing issue. A controlled reproduction is the minimal check that would settle whether the discrepancy is a reporting artifact or a genuine failure of the framework. If the numbers hold, the paper's SOTA claim must be narrowed and the role of Multi-TSI on ScanRefer clarified. I therefore keep the reader's CONDITIONAL verdict; the condition should explicitly include resolving Table 3 before acceptance.","tokens_in":17286,"tokens_out":9482,"duration_ms":106828,"concrete_test":"Reproduce Table 3 with one controlled protocol on ScanRefer: identical split, same detection backbone, same number of views, and same training budget for ViewSRD and for BUTD-DETR+SRD, and report per-split (Unique/Multiple) Acc@0.25/0.5. If BUTD-DETR+SRD still exceeds ViewSRD by more than 10 points, the paper must either restrict its SOTA claim to Nr3D/Sr3D or supply a documented reason for the gap (e.g., ViewSRD was not trained on ScanRefer).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest evidence for the paper's central claim is internally inconsistent. In Table 3, the full ViewSRD system, which contains both the SRD module and the Multi-TSI module, reports 45.4% overall Acc@0.25 on ScanRefer. The same table reports that adding only the SRD text-decomposition module to the existing BUTD-DETR baseline gives 57.9% overall Acc@0.25, a 12.5-point advantage over the full model. Since SRD is described as model-agnostic preprocessing and Multi-TSI is described as a further improvement, the full model should dominate a baseline augmented with just SRD. The paper gives no protocol explanation (e.g., different backbones, training data, evaluation splits, or number of views) for this inversion. As reported, either the Table 3 numbers are not comparable, in which case the 'SRD enhances other methods' claim is not established, or the ViewSRD architecture is substantially worse than a baseline that received only one of its components, contradicting the abstract's claim that ViewSRD 'significantly outperforms state-of-the-art methods' on 3D visual grounding datasets. This is more load-bearing than the decomposition limitation because it does not rely on an assumption about language; it is a direct inconsistency among the paper's own reported results. The ablation in Table 4 shows that removing SRD from ViewSRD costs only 1.3 points on Nr3D, which makes the 14.9% relative ScanRefer gain from SRD on BUTD-DETR particularly difficult to reconcile.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ViewSRD, a framework for 3D visual grounding that decomposes complex multi-anchor queries into simpler single-anchor sentences via an LLM-based Simple Relation Decoupling (SRD) module, and then fuses multi-view text and scene features using Cross-modal Consistent View Tokens (CCVTs) in a Multi-view Textual-Scene Interaction (Multi-TSI) module. Experiments on Nr3D, Sr3D, and ScanRefer report strong results on Nr3D and Sr3D (e.g., 69.9% overall on Nr3D, a 5.2-point gain over CoT3DRef), and the authors claim that the SRD module improves existing methods (MVT, BUTD-DETR, EDA) on ScanRefer. The paper also includes ablations on components, view counts, LLM decouplers, and loss weights, and releases code.","tokens_in":17592,"tokens_out":4282,"duration_ms":43834,"significance":"If the reported results are correct, ViewSRD would be a meaningful advance in multi-view 3D visual grounding, particularly for complex multi-anchor queries, and the SRD module as a model-agnostic preprocessing step would be a useful contribution. The paper is clearly motivated, provides code, and includes a reasonable set of ablations, including a t-SNE visualization of CCVTs. However, the ScanRefer results in Table 3 contain internal inconsistencies and arithmetic errors that directly affect the central claim that SRD enhances other methods and that ViewSRD is state-of-the-art across datasets. These issues must be resolved before the findings can be relied upon.","major_comments":[{"comment":"The reported overall Acc@0.25 numbers in Table 3 are internally inconsistent with the stated Unique (19%) / Multiple (81%) split and the sub-scores. For BUTD-DETR+SRD, the weighted average of 85.0 (Unique) and 45.3 (Multiple) is approximately 52.8, not the reported 57.9; similarly, for BUTD-DETR the weighted average is approximately 51.9, not 50.4. Since the paper uses these overall numbers to claim a 14.9% relative improvement from SRD, the table must be corrected or the calculation explained.","section":"Sec. 4.4, Table 3"},{"comment":"The full ViewSRD model reports 45.4 overall Acc@0.25 on ScanRefer, which is 12.5 points lower than BUTD-DETR+SRD (57.9) and even lower than the plain BUTD-DETR baseline (50.4). Because SRD is a component of ViewSRD, this inversion is unexplained and contradicts the claim that SRD improves grounding and that ViewSRD is state-of-the-art on ScanRefer. The authors need to either provide a protocol explanation (e.g., different backbone, training data, or evaluation splits) or correct the numbers.","section":"Sec. 4.4, Table 3"},{"comment":"The ablations on Nr3D (Table 4) show that removing SRD from ViewSRD costs only 1.3 points (68.6 vs. 69.9), whereas Table 3 claims SRD adds 7.5 points to BUTD-DETR on ScanRefer (50.4 to 57.9). Even accounting for dataset differences, this large discrepancy in the measured contribution of SRD is not discussed and raises questions about whether the ScanRefer improvement is replicable or confounded.","section":"Sec. 4.5, Table 4 vs. Sec. 4.4, Table 3"},{"comment":"The abstract states that ViewSRD 'significantly outperforms state-of-the-art methods' on 3D visual grounding datasets without qualification, but Table 3 shows that on ScanRefer the full ViewSRD is outperformed by BUTD-DETR+SRD and, at Acc@0.25, is comparable to or worse than several baselines. The claim should be restricted to the datasets where it holds (Nr3D and Sr3D) or otherwise reconciled with Table 3.","section":"Abstract and Sec. 4.4"}],"minor_comments":[{"comment":"The text uses 'DeepSeek-R1' but reference [27] is the DeepSeek-V3 technical report; please cite the appropriate R1 technical report or clarify the model version.","section":"Sec. 3.1 and References"},{"comment":"The appendix text says the best loss weights are 'λObj = 1.0, λRef = 0.5, λSent = 0.5', but the table columns are ordered λRef, λObj, λSent, making the row '1.0 0.5 0.5' correspond to λRef=1.0, λObj=0.5. Please fix this mismatch.","section":"Appendix G"},{"comment":"There is a typo in the caption: 'Simple Queris' should be 'Simple Queries'.","section":"Figure 4 caption"},{"comment":"All results appear to be from a single run with no error bars or variance reporting; given the small differences in Table 6 (e.g., 69.6 vs. 69.9), please state whether results are averaged over multiple seeds or report standard deviations.","section":"Sec. 4.1 and Tables 1-6"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistencies in Table 3 are concerning, especially the arithmetic gap between the weighted averages and the reported overall numbers. This goes beyond presentation and affects the credibility of the 'SRD enhances other methods' claim. I would like the editor to request the authors to provide the raw per-split numbers and a clear protocol statement for the ScanRefer experiments, and to re-verify all reported numbers before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper to know about: ViewSRD is a 3D visual grounding framework that combines LLM-based query decomposition with multi-view transformers using shared, learnable view tokens. The headline number is 69.9% overall on Nr3D, 5.5 points above prior SOTA. The architecture is straightforward and mostly reproduces known pieces: the SRD module is a prompt-based decomposition similar in spirit to EDA and CoT3DRef, the view tokens are essentially learned positional embeddings, and the multi-view interaction follows MVT/ViewRefer. The genuine novelty is in the specific packaging, plus the parallel referential loss that supervises anchors and target together. The paper is clearly written and ships abundant ablations, including a useful view-count analysis and an LLM replacement study. The t-SNE of the view tokens is nice evidence that the tokens capture viewpoint structure.\n\nThe soft spot is in the ScanRefer table. The full ViewSRD scores 45.4 Acc@0.25, while BUTD-DETR+SRD, which is just the baseline with the SRD preprocessing, scores 57.9. The paper gives no protocol explanation. As reported, the full model is 12.5 points worse than a baseline with one of its own components. That directly contradicts the abstract's claim of strong superiority and makes the SRD-enhances-others claim hard to interpret. Either the numbers come from different evaluation setups, which the paper doesn't disclose, or ViewSRD's multi-view machinery hurts on ScanRefer. Either way, the central claim is not established as written. The ablation doesn't help: removing SRD from ViewSRD costs only 1.3 points on Nr3D, so the module that supposedly drives the 14.9% relative gain on BUTD-DETR is almost dispensable within the full model. The loss-weight reporting also conflicts between the main text (λRef=1.0, λObj=0.5, λSent=0.5) and the appendix (λObj=1.0, λRef=0.5). And a strong recent baseline, G3-LQ, is missing from the comparison tables.\n\nThat said, the Nr3D gain is large and the multi-anchor analysis is interesting. If the ScanRefer numbers are reconciled and the missing baseline included, this could be a solid contribution. The decomposition assumption is honestly stated as a limitation in the conclusion; that's a point in the authors' favor.\n\nMy recommendation: send this to peer review. The desk reject rate is low, but the referees need to pressure-test the ScanRefer discrepancy and the ablation attribution. It deserves a serious referee, but not acceptance in current form.","headline":"ViewSRD's Nr3D gain is real and the method is clean, but the ScanRefer table contradicts the SOTA claim, so the paper needs revision before the numbers can be trusted.","tokens_in":18205,"tokens_out":2890,"would_cite":false,"duration_ms":30957,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that 3D visual grounding improves substantially when complex multi-anchor referring expressions are first decomposed into independent single-anchor statements and processed with explicit cross-modal viewpoint tokens.","keywords":["3D visual grounding","multi-view decomposition","multi-anchor query","cross-modal view tokens","point cloud","large language models","ScanRefer","Nr3D"],"falsifier":"Take a held-out set of Nr3D queries where the target's location depends on a relation between two anchors (for example, 'the chair between the table and the door that is to the left of the window') and compare ViewSRD's accuracy on those overlapping-relation queries against non-overlapping multi-anchor queries. If accuracy on overlapping queries does not improve over a baseline that keeps the original sentence, or if removing the SRD module yields equal or better accuracy on that subset, the central claim fails.","tokens_in":17022,"feed_emoji":"🎯","tokens_out":6955,"duration_ms":78822,"temperature":0.7,"pith_summary":"The paper sets out to show that 3D visual grounding—finding the object in a 3D scene that a sentence names—fails for two structural reasons, and that both can be addressed before any fusion happens. Complex descriptions often name several anchor objects (for example, 'the nightstand to the right of the bed, closest to the wall with the framed picture'), and the same spatial phrase changes meaning when the observer moves. ViewSRD's claim is that rewriting such a query into a set of independent single-anchor statements, each paired with an explicit viewpoint token in both text and scene branches, removes most of the ambiguity. On Nr3D the method reports 69.9% overall accuracy, 5.2 points above the previous best under identical settings, with the largest gains on view-dependent and multi-anchor queries; on ScanRefer, the decomposition module alone improves three existing methods when added as preprocessing. If correct, the paper shows that query restructuring and viewpoint-aligned representation, rather than a better fusion head, are the levers that move accuracy.","feed_headline":"Split tangled queries, and 3D grounding jumps to 69.9%","feed_subtitle":"Decomposing multi-anchor descriptions plus viewpoint-aligned tokens adds 5.2 points over prior best on Nr3D.","key_machinery":"The load-bearing machinery is the pair (SRD, CCVT). SRD uses a large language model with prompt templates plus a sentence-matching filter to rewrite each multi-anchor query into one simplified single-anchor sentence per anchor, preserving the original sentence as an extra reference; this carries the disentanglement of target-anchor relations. CCVTs are learnable vectors, one per viewpoint, shared by the textual and scene modules; the text module reweights them by how well each sentence matches each viewpoint, and the scene module concatenates them to the object tokens before transformer layers. The two together make viewpoint information a common, learned currency across modalities.","core_discovery":"The central discovery, stated in the paper's own terms, is that 3D visual grounding can be formulated as a structured multi-view decomposition process: first split the language, then align every piece of language with every viewpoint, then let a reasoning module vote. The Simple Relation Decoupling (SRD) module turns a multi-anchor query into one simplified sentence per anchor while keeping the original sentence as a reference; textual aggregation randomly mixes these sentences during training so the model learns from both raw and decomposed forms. The Multi-view Textual-Scene Interaction (Multi-TSI) module injects shared learnable Cross-modal Consistent View Tokens (CCVTs) into both pretrained text features and point-cloud scene features, so a phrase like 'to the right of the bed' is interpreted relative to the same viewpoint in both modalities. The final Textual-Scene Reasoning module fuses per-view predictions. Ablations show every component contributes, and removing CCVTs hurts most.","pith_inferences":["If decomposition really is the active ingredient, then language-conditioned tasks beyond grounding—such as instruction following in 3D scenes or referring-expression generation—could benefit from the same rewrite-then-align pattern; this is an extension the paper does not test.","A direct test would replace the LLM decoupler with a rule-based or learned relation splitter; if accuracy holds, the gain is from restructuring itself, not from LLM paraphrase quality, which changes how transferable the result is.","The plateau at eight views suggests uniform multi-view attention dilutes the strongest viewpoint; an adaptive view-selection mechanism, which the paper names as future work, would likely beat simply adding more cameras."],"forward_implications":["Multi-anchor queries stop being a liability: after decomposition, the paper reports higher accuracy on multi-anchor examples (71.5%) than on single-anchor ones (69.5%) on Nr3D, the opposite of the usual trend.","View-dependent descriptions benefit more than view-independent ones, so viewpoint-aligned language and scene tokens are a direct cure for perspective-induced spatial ambiguity.","The SRD module is a drop-in text preprocessor: adding it to MVT, BUTD-DETR, and EDA improves ScanRefer accuracy without changing their architectures.","Better LLM decouplers push accuracy higher (70.7% with the strongest tested), so gains are tied to decomposition quality and not to the specific model.","Testing with more views than the four used in training continues to improve accuracy, while training with eight views plateaus, indicating a practical sweet spot around four viewpoints."],"supporting_citations":[{"why":"Supplies the Nr3D and Sr3D datasets with target-anchor annotations that define the grounding benchmarks and the anchor set used by SRD.","marker":"[1]"},{"why":"Supplies ScanRefer, the dataset for the SRD plug-in experiments on existing methods.","marker":"[6]"},{"why":"Defines the multi-view transformer baseline that ViewSRD builds on and compares against for view modeling.","marker":"[18]"},{"why":"Provides the CoT3DRef baseline and the parallel referential loss that supervises target and anchor localization together.","marker":"[2]"},{"why":"Introduces the multi-key-anchor scene-aware baseline that motivates explicit target-anchor decoupling.","marker":"[5]"},{"why":"Supplies the point-cloud encoder used to extract per-object scene features for each viewpoint.","marker":"[35]"},{"why":"Supplies the pretrained text encoder used to extract sentence features for original and decomposed queries.","marker":"[11]"},{"why":"Supplies the cross-attention mechanism that carries the view-token interaction and final text-scene reasoning.","marker":"[39]"},{"why":"Supplies the large language model used as the SRD query decoupler in the main experiments.","marker":"[27]"}],"fun_headline_variants":["ViewSRD: Splitting questions and aligning views improves 3D grounding","Decompose, align, fuse: How ViewSRD hits 69.9% on Nr3D","Multi-view tokens and query splitting lift 3D grounding by 5.2 points","Tangled queries? ViewSRD untangles them to grasp 3D scenes better","Better 3D grounding by breaking down queries and syncing views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a complex query can be fully rewritten as independent single-anchor sentences without losing the overlapping relationships among anchors; the paper states this assumption as its limitation, and if the rewrite drops a needed relationship the whole decomposition benefit disappears.","fun_headline_variants_meta":{"raw":{"variants":["ViewSRD: Splitting questions and aligning views improves 3D grounding","Decompose, align, fuse: How ViewSRD hits 69.9% on Nr3D","Multi-view tokens and query splitting lift 3D grounding by 5.2 points","Tangled queries? ViewSRD untangles them to grasp 3D scenes better","Better 3D grounding by breaking down queries and syncing views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001256,"raw_usage":{"total_tokens":5158,"prompt_tokens":968,"completion_tokens":4190,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":4080}},"tokens_in":584,"tokens_out":4190,"duration_ms":34334,"temperature":1.0,"reasoning_tokens":4080,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:12:13.891437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of Nr3D queries where the target's location depends on a relation between two anchors (for example, 'the chair between the table and the door that is to the left of the window') and compare ViewSRD's accuracy on those overlapping-relation queries against non-overlapping multi-anchor queries. If accuracy on overlapping queries does not improve over a baseline that keeps the original sentence, or if removing the SRD module yields equal or better accuracy on that subset, the central claim fails.","supporting_citations":[{"cited_title":"Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the Nr3D and Sr3D datasets with target-anchor annotations that define the grounding benchmarks and the anchor set used by SRD."},{"cited_title":"Scanrefer: 3d object localization in rgb-d scans using natural language","cited_arxiv_id":null,"evidence_quote":"Supplies ScanRefer, the dataset for the SRD plug-in experiments on existing methods."},{"cited_title":"Multi- view transformer for 3d visual grounding","cited_arxiv_id":null,"evidence_quote":"Defines the multi-view transformer baseline that ViewSRD builds on and compares against for view modeling."},{"cited_title":"Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding","cited_arxiv_id":null,"evidence_quote":"Provides the CoT3DRef baseline and the parallel referential loss that supervises target and anchor localization together."},{"cited_title":"Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding","cited_arxiv_id":null,"evidence_quote":"Introduces the multi-key-anchor scene-aware baseline that motivates explicit target-anchor decoupling."}],"review_version":1}