{"id":"75e7fc27-08d4-4bba-9bf4-050c5d9d70b3","arxiv_id":"2504.18738","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"This is a review that organizes traditional and vision-language-model-based 3D object detection methods into comparison tables and discusses strengths, trade-offs, and open challenges.","lead":"This paper surveys about one hundred papers on detecting 3D objects with vision-language models, comparing traditional geometry-based detectors with language-guided approaches. A smart generalist might read it to learn the landscape of this emerging robotics and computer vision area, but it contains no new experiments or results.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of a comprehensive, first-of-its-kind review rests on an unverifiable and demonstrably error-prone corpus; without the full list of 105 papers and corrected attributions, the review's map of the field cannot be trusted.","rationale":"The reader's weakest assumption identifies the corpus as the load-bearing premise, and the manuscript text itself provides multiple grounds for doubting it. The strongest claim is that the paper is a reliable, first-of-its-kind map of 105 state-of-the-art papers on VLM-based 3D object detection. To support that, the corpus must be enumerable, representative, and accurately described. None of these conditions is met: the 105 papers are never listed, the AI-assisted search protocol lacks dates and validation, and the tables and prose contain repeated citation and attribution errors. I independently checked the manuscript and found specific, checkable failures: duplicate MV3D rows, misattributed PointNet++ and VoxelNet citations, a malformed FusionTransformer citation, an OWL-ViT placeholder, a mismatch between OmniVLM3D and its cited reference, and Cube R-CNN attributed to the OMNI3D benchmark paper rather than the actual model paper. The uniqueness claim is also weakened by the authors' own cited surveys of overlapping topics. These are not stylistic flaws; they directly affect whether a reader can rely on the review's comparisons and conclusions. I agree with the reader's REJECT verdict: the paper could become a useful survey after a full enumeration and correction of the corpus and after dropping the unsupported 'first' claim, but as submitted the central claim is unsupported.","tokens_in":36885,"tokens_out":4829,"duration_ms":46843,"concrete_test":"Request from the authors, or extract from the associated GitHub repository, the complete enumerated list of the 105 included papers together with search dates, query strings, and inclusion/exclusion decisions. Independently verify every Table 1 and Table 2 entry and all in-text attributions by locating the original paper and checking (a) title and venue match, (b) whether the work actually addresses 3D object detection with vision-language models, and (c) whether the table row describes the method correctly. Specifically audit PointNet++, VoxelNet, the duplicate MV3D rows, FusionTransformer, OWL-ViT, OmniVLM3D, and Cube R-CNN.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's contribution is a 'groundbreaking and comprehensive review, the first of its kind' of 3D object detection with VLMs, built on 105 curated papers. That claim stands or falls on the representativeness and accurate attribution of the corpus. This premise is not secure. The full list of 105 papers is never provided; search dates, query strings, inclusion/exclusion decisions, and any validation of AI-engine outputs are absent. The screening protocol (459 to 208 to 134 to 105, using ChatGPT, Perplexity, Grok, and DeepSeek) is not reproducible as described. Concrete errors confirm the risk: Table 1 lists MV3D twice and cites PointNet++ via a 2020 workshop paper rather than Qi et al. 2017; Table 2 contains an uncited 'OWL-ViT Link to paper' placeholder and a raw 'citefent2024dpft' token; 'OmniVLM3D' is cited to Chen et al. 2024a, which is actually a token-compressed 2D VLM paper; Cube R-CNN is attributed to Brazil et al. (the OMNI3D dataset paper) rather than to Piekenbrinck et al. 2024; and the Objectives section mis-cites VoxelNet as Chen et al. 2023b, which is VoxelNeXt. The 'first' claim is also contradicted by the authors' own references: Ma et al. 2024 surveys 3D tasks via multi-modal LLMs, and Sapkotaa and Karkeea 2025 reviews object detection with multimodal LVLMs. Sections 3.2 and 3.3 describe generic VLM architectures (LLaVA, BLIP-2, Qwen2-VL) and non-detection tasks such as segmentation, captioning, and VQA, so the taxonomies and trade-off tables do not clearly apply to 3D object detection specifically. Because the review's value is precisely that it organizes this literature, an unverifiable, internally inconsistent corpus undermines the central claim rather than merely requiring cosmetic citation fixes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to be the first comprehensive review of 3D object detection with vision-language models (VLMs). It reports a hybrid search strategy combining academic databases and AI engines that reduced 459 candidate papers to 105 selected works, and it organizes the field into traditional point-cloud/voxel methods versus VLM-based approaches, with comparison tables, architecture overviews, benchmark discussions, and a list of challenges and future directions. The core value proposition is a reliable, curated map of this emerging area, together with a comparative trade-off analysis.","tokens_in":37276,"tokens_out":3679,"duration_ms":35823,"significance":"If the corpus were accurate and representative, a systematic survey of VLM-based 3D object detection would be timely and useful to researchers in robotics, autonomous driving, and embodied AI. The paper does assemble a broad set of references and organizes them into useful categories, and the accompanying GitHub project is a potentially helpful resource. The authors also correctly identify real open problems such as spatial grounding, annotation cost, and real-time latency. However, the scientific value of a review stands or falls on the correctness and completeness of its curation and attribution, and the manuscript's pervasive citation errors, unverifiable corpus, and failure to engage prior surveys undermine that value in its current form.","major_comments":[{"comment":"The central claim of a 'comprehensive' review rests on the 105-paper corpus, but the manuscript never provides the list of included papers, the search dates, the query strings, or the inclusion/exclusion decisions for the filtering cascade (459 to 208 to 134 to 105). The protocol as described is not reproducible: the role of AI engines such as ChatGPT, Perplexity, Grok, and DeepSeek in retrieving and screening papers is not specified, and no validation of the AI-generated outputs is given. Without a verifiable corpus, every aggregate comparison and trend claim in the review is unsupported.","section":"Section 2.1, Figures 3-4"},{"comment":"Several foundational papers are misattributed in ways that corrupt the baseline of the review. PointNet++ is attributed to Sheshappanavar and Kambhamettu 2020, which is a different workshop paper, instead of Qi et al. 2017; VoxelNet is cited in Section 1 (Objective) as [Chen et al. 2023b; Sindagi et al. 2019], both of which are different works (VoxelNeXt and MVX-Net respectively), although Table 1 correctly cites Zhou and Tuzel 2018; and MV3D appears twice in Table 1. These errors are not typographical, because they misassign foundational contributions and appear in the paper's own structured comparison tables.","section":"Table 1 and Section 3.1.1 (Objective)"},{"comment":"The VLM-based comparison table includes models that are not 3D object detectors and citations that point to unrelated papers. For example, 'OmniVLM3D' is cited as [Chen et al. 2024a], but the reference list entry for Chen et al. 2024a is 'OmniVLM: A Token-Compressed, Sub-Billion-Parameter Vision-Language Model for Efficient On-Device Inference,' a 2D efficient VLM with no 3D detection component. Similarly, 'Cube R-CNN' is attributed to Brazil et al. 2023 (the OMNI3D dataset paper) rather than to Piekenbrinck et al. 2024, which is the actual source of the Cube R-CNN architecture and is also in the reference list. These misattributions mean the paper's map of the VLM-based 3D detection field is unreliable at the level of individual models.","section":"Table 2 and Section 3.3.1"},{"comment":"The paper repeatedly calls itself 'the first of its kind' focused on 3D object detection with VLMs, but this claim is contradicted by the authors' own reference list: Ma et al. 2024 surveys 3D tasks via multi-modal LLMs, and Sapkotaa and Karkeea 2025 is an in-depth review of object detection with multimodal large vision-language models. The 'first' claim is therefore either inaccurate or requires a specific delimitation of scope that the manuscript does not provide; the lack of any discussion of these prior surveys weakens the novelty assertion that is central to the paper's framing.","section":"Introduction (first paragraph) and Section 2.2"},{"comment":"The review's scope is broader than its title, and this dilutes the claimed focus on 3D object detection. The VLM section discusses general-purpose models (LLaVA-1.5, BLIP-2, Qwen2-VL, InternVL) and non-detection tasks such as segmentation, captioning, and VQA, while the trade-off analysis in Section 3.3.2 makes quantitative claims such as '2–5x more GPU resources' and '8 FPS' versus '50 FPS' without providing a source or a derivation, and Section 5 cites '15–20% lower FPS' without a citation. These generic and unsupported comparisons do not constitute a rigorous comparative analysis of VLM-based 3D object detection specifically.","section":"Sections 3.2, 3.3, and Tables 3-4"}],"minor_comments":[{"comment":"The entry 'FusionTransformer citefent2024dpft' contains a raw LaTeX citation token that should be replaced with the proper citation to Fent et al. 2024.","section":"Table 1"},{"comment":"The entry 'OWL-ViT Link to paper' contains a placeholder URL text, not a proper citation, and no year or author information is given.","section":"Table 2"},{"comment":"There are typographical errors such as 'bby' in the Figure 9 caption and 'Likwise' in Section 3.2.2; these should be corrected.","section":"Figure 9 caption and Section 3.2.2"},{"comment":"The subsection numbering '4.0.1' and '4.0.2' is nonstandard; these should be numbered 4.1 and 4.2, and the references to 'Table 4' appear before the table is introduced and are not clearly tied to the enumerated tradeoffs.","section":"Section 4.0.1-4.0.2"},{"comment":"The text in the 'Voxelization Bottlenecks' bullet cites VoxelNet as [Sindagi et al. 2019] while Table 1 correctly cites [Zhou and Tuzel 2018]; these internal inconsistencies should be systematically fixed across the manuscript.","section":"Section 3.1.3"}],"recommendation":"reject","confidential_remarks":"The manuscript has an unusually high density of citation and metadata errors for a review whose primary contribution is curation. The missing list of the 105 reviewed papers and the non-reproducible AI-based search protocol make the central empirical claim unverifiable. The 'first of its kind' framing also appears to overlook prior surveys cited in the paper itself. The heavy reliance on the first author's own recent preprints for framing and novelty claims is a further concern. I would not consider acceptance in any form until the corpus is fully enumerated, every citation in the tables is verified, and the scope of the review is honestly repositioned relative to existing surveys."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This survey has the right raw material on the wrong footing. The topic—3D object detection with vision-language models—is timely, and the authors have pulled together a broad set of relevant works across traditional and VLM-based detectors, with useful table-driven comparisons of architectures, benchmarks, and trade-offs. The GitHub page is a helpful addition. For a reader who simply wants names and pointers to start a literature search, this has real value.\n\nThe paper's central claim, however, is 'groundbreaking...first of its kind...curated and analyzed over 100 state-of-the-art papers,' and that claim is not supported as submitted. The corpus of 105 papers is never enumerated: no list, no search dates, no query strings, no reproducible inclusion/exclusion protocol. The hybrid search via ChatGPT, Perplexity, Grok, and DeepSeek may be a reasonable way to generate candidates, but treating chatbot outputs as validated review input without transparency is not a systematic method. The 'first' claim is also contradicted by surveys the paper itself cites—Ma et al. 2024 covers 3D tasks via MLLMs, and the authors' own Sapkotaa and Karkeea 2025 covers object detection with large VLMs—even if those are not exactly scoped to 3D detection.\n\nThe concrete citation errors are too many to wave off as typos: Table 1 lists MV3D twice and contains the raw token 'citefent2024dpft' for FusionTransformer; Table 2 has 'OWL-ViT Link to paper' as a placeholder; the Objectives section cites PointNet++ to Sheshappanavar and Kambhamettu 2020 and VoxelNet to Chen et al. 2023b (VoxelNeXt), even though Table 1 gets both right; OmniVLM3D is cited to Chen et al. 2024a, which is a token-compressed 2D VLM, not the 3D model; and Table 3 attributes Cube R-CNN to Brazil et al. 2023 rather than Piekenbrinck et al. 2024. A survey is only as good as its curation, and these are curation failures in a load-bearing part of the paper.\n\nThat said, the paper is not incoherent, and the chapters on VLM architecture, pretraining, and the traditional-vs-VLM trade-offs are clearly written. The flaws are fixable: add the full list of reviewed papers, document the search protocol, correct the misattributions, and drop the 'first/groundbreaking' framing. As submitted I would not cite it as a reliable reference, but I would send it to a serious referee with a request for major revision rather than desk-reject it.","headline":"A well-intentioned but not-yet-reliable survey: the topic is timely and the compilation is broad, but the unverifiable corpus and repeated citation errors sink the 'first-of-its-kind' claim as submitted.","tokens_in":37915,"tokens_out":3911,"would_cite":false,"duration_ms":37228,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review claims that 3D object detection is shifting from fixed-class geometry models to language-guided vision-language systems, and maps 105 papers to support that.","keywords":["3D object detection","vision-language models","open-vocabulary detection","zero-shot 3D perception","point clouds","multimodal fusion","survey","embodied AI"],"falsifier":"Compile the full list of the 105 included papers, rerun the same search queries on the twelve platforms with dated logs, and verify each row of Tables 1 and 2 against its cited source; a material number of missing or misattributed entries would show that the corpus is not the representative map the review claims.","tokens_in":36653,"feed_emoji":"👁️","tokens_out":6744,"duration_ms":60026,"temperature":0.7,"pith_summary":"The paper aims to establish that 3D object detection with vision-language models has become a coherent research area worth its own survey, and that it is the first to provide one. It argues that traditional point-cloud and voxel detectors are accurate but closed-set, annotation-heavy, and sensor-bound, while VLM-driven systems such as CLIP-based and 3D-LLM approaches enable open-vocabulary, zero-shot, instruction-following detection. The review synthesizes 105 selected papers through a hybrid academic-and-AI search, comparing architectures, pretraining strategies, benchmarks, and trade-offs. Its central takeaway is that language-guided 3D perception brings semantic flexibility and generalization at the cost of speed, compute, and spatial precision, and it identifies 3D-language data scarcity as the main bottleneck. If the map is right, researchers get a usable entry point and roadmap for a fast-moving field.","feed_headline":"Language-guided 3D detection mapped across 105 papers","feed_subtitle":"The first systematic review contrasts geometry-only detectors with open-vocabulary vision-language systems and their trade-offs","key_machinery":"The load-bearing structure is the review's two-axis taxonomy plus the generic VLM architecture it describes. The taxonomy separates traditional 3D detectors (point clouds, voxel grids, frustum proposals) from VLM-based detectors (CLIP-style alignment, 3D-LLM injection, instruction tuning) and then compares them across data efficiency, generalization, explainability, compute, and use cases. The architectural machinery is the three-stage VLM pipeline—image encoder, multimodal projector, text decoder—extended to 3D by projecting 2D language-grounded proposals into 3D frustums and refining them with cross-modal attention. The curated corpus of 105 papers, filtered from 459 through three screening rounds, is what carries the claim that this is a representative map rather than a sample of convenience.","core_discovery":"On its own terms, the paper's discovery is organizational: a systematic account of how 3D object detection is moving from geometry-only neural networks to vision-language models. It classifies traditional methods (voxel-wise, point-wise, ROI-wise, attention-based) against VLM-based systems (CLIP-aligned, instruction-tuned, LLM-injected, open-vocabulary), and identifies a recurring pipeline in which a VLM produces 2D proposals, projects them into 3D frustums with depth or LiDAR, aligns hierarchical features cross-modally, and refines the boxes with language feedback. The review reports that VLM methods excel at zero-shot and open-vocabulary detection and semantic grounding, while traditional methods retain advantages in latency and geometric precision, and it flags spatial misalignment, occlusion sensitivity, and limited real-time viability as the pressing open problems.","pith_inferences":["If the taxonomy holds, one testable extension is a living benchmark that scores VLM-based 3D detectors on open-vocabulary queries across indoor, outdoor, and agricultural scenes, since the review's own apple-orchard example suggests deployment targets beyond autonomous driving.","The recurring 2D-to-3D frustum pipeline implies that progress in 2D open-vocabulary detection could transfer almost directly into 3D, so advances in 2D VLMs may compound into 3D gains without new 3D-specific methods.","A reproducible version of the search protocol—with dated queries, inclusion lists, and inter-annotator agreement—would let the community verify whether the 105-paper corpus is genuinely representative, and would turn the review's map into a repeatable methodology.","The review's framing suggests that hybrid architectures combining voxel efficiency with language grounding, rather than pure VLM end-to-end systems, may be the fastest route to real-time open-vocabulary 3D detection; this is an inference, since the paper only notes hybrids as an emerging direction."],"forward_implications":["If the review's map is correct, new researchers get a structured entry point: the field splits cleanly into 43 traditional-network papers and 62 VLM-based papers, with the latter organized by architecture and task.","VLM-based 3D detection becomes the default answer for open-vocabulary and instruction-driven robotics, AR, and embodied AI, where users query objects in natural language rather than fixed class labels.","Traditional voxel- and point-based detectors remain the safer choice for latency-critical, closed-set deployments, since the review reports VLM systems run at roughly 8-15 FPS versus 50+ FPS for optimized conventional detectors.","The identified bottlenecks—3D-language data scarcity, cross-modal misalignment, semantic hallucination, and compute—define a concrete agenda: synthetic captioning, 3D scene graphs, and multimodal reinforcement learning are the paper's proposed next steps.","Benchmarks such as OMNI3D, ScanRefer, SUN RGB-D, and ScanNet will likely be the proving grounds, and the review implies that no standard benchmark yet exists specifically for VLM-based 3D detection."],"supporting_citations":[{"why":"Supplies the VoxelNet baseline that anchors the traditional voxel-wise processing category in Table 1.","marker":"[Zhou and Tuzel 2018]"},{"why":"Establishes PointNet as the point-wise learning baseline the review contrasts with VLM approaches.","marker":"[Qi et al. 2017]"},{"why":"Provides MV3D, the early multi-view fusion method used to illustrate ROI-wise traditional detection limits.","marker":"[Chen et al. 2017]"},{"why":"Exemplifies embodied VLM detection, grounding 3D proposals in language queries for robotics.","marker":"[Driess et al. 2023]"},{"why":"Anchors the 3D-LLM category by injecting multi-view rendered point-cloud features into large language models.","marker":"[Hong et al. 2023]"},{"why":"Supplies CoDA, the open-vocabulary 3D detector whose cross-modal alignment and 80% mAP improvement back the VLM advantage claim.","marker":"[Cao et al. 2023]"},{"why":"Supports the zero-shot claim by prompting CLIP and GPT for 3D classification without 3D training data.","marker":"[Zhu et al. 2023]"},{"why":"Introduces OMNI3D, the large-scale benchmark the review uses to argue traditional evaluation is expanding beyond small datasets.","marker":"[Brazil et al. 2023]"},{"why":"Provides 3DVLP, the object-contrastive pretraining framework the review cites for grounding and captioning generalization.","marker":"[Zhang et al. 2024a]"},{"why":"Anchors OpenScene's CLIP-based open-vocabulary scene understanding in the VLM comparison.","marker":"[Peng et al. 2023]"}],"fun_headline_variants":["First systematic review of 3D detection with VLMs","Geometry vs language: 3D detection's new frontier","105 papers map the rise of VLM-based 3D detection","Open-vocabulary 3D detection: what works, what doesn't","From voxels to language: 3D detection evolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusions stand on the assumption that the 105 selected papers form a representative, correctly attributed corpus of the field, yet the paper does not list the included papers, give search dates, or provide a reproducible protocol.","fun_headline_variants_meta":{"raw":{"variants":["First systematic review of 3D detection with VLMs","Geometry vs language: 3D detection's new frontier","105 papers map the rise of VLM-based 3D detection","Open-vocabulary 3D detection: what works, what doesn't","From voxels to language: 3D detection evolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3374,"prompt_tokens":924,"completion_tokens":2450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2363}},"tokens_in":540,"tokens_out":2450,"duration_ms":17155,"temperature":1.0,"reasoning_tokens":2363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:09:51.272847+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compile the full list of the 105 included papers, rerun the same search queries on the twelve platforms with dated logs, and verify each row of Tables 1 and 2 against its cited source; a material number of missing or misattributed entries would show that the corpus is not the representative map the review claims.","supporting_citations":[],"review_version":1}