{"id":"2f95f8c3-1cd2-4c92-82c1-d6b3440fb9ba","arxiv_id":"2501.16384","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Mamba-Transformer hybrid cell with adjacency-preserving reordering achieves near-SOTA view-guided point cloud completion at reduced parameter count.","lead":"MambaTron combines a Mamba state-space layer with a block-based transformer to fill in missing parts of 3D point clouds using a reference image as a guide. It reports accuracy on the ShapeNet-ViPC benchmark that edges out prior methods while using about 3.9 million parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark SOTA claim is confounded by ShapeNet55 pretraining absent from baselines and by unreported architecture/missing variance; the 0.012 CD edge may not be attributable to MambaTron.","rationale":"The reader's weakest assumption is that the ShapeNet-ViPC baseline numbers are directly comparable despite missing variance and differing codebases. My concern overlaps but identifies a more specific, more load-bearing confound: MambaTron is pretrained on ShapeNet55 complete point clouds and their image projections before the cross-modal ShapeNet-ViPC stage (§4.2), while the text gives no evidence that EGIInet and XMFnet received equivalent pretraining. If baselines were trained only on ShapeNet-ViPC, the reported 0.012 CD improvement could be an artifact of additional training data, not of the MambaTron cell. This is directly testable and was not addressed in the paper's ablations. The missing architecture details and apparent parameter-count discrepancy (3.92 M in §5.1 vs 13.6 M in Table 3) further prevent verification of the efficiency claim. These are evidence gaps rather than demonstrated errors, so the appropriate disposition is UNVERDICTED: the central claim is plausible but cannot be accepted or rejected on the current evidence. I agree with the reader's broad concern about comparability but not with their precise attribution to noise alone; the pretraining protocol is the more decisive issue.","tokens_in":14318,"tokens_out":5266,"duration_ms":47876,"concrete_test":"Re-run the ShapeNet-ViPC comparison in a controlled protocol: train MambaTron from scratch on ShapeNet-ViPC with no ShapeNet55 unimodal stage, and also train EGIInet and XMFnet with the same ShapeNet55 pretraining, using identical train/test split, input token counts, and at least 5 seeds; report mean±std of CD and F-score. If the EGIInet gap vanishes or reverses when pretraining is removed or matched, the §5.1 claim is not attributable to the MambaTron architecture.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim in §5.1 — 'On average, our model outperforms the latest state-of-the-art technique with only 3.92 M parameters' — is not yet supported because the comparison is protocol-unfair and the model identity is underspecified. §4.2 describes a two-stage training procedure: a unimodal stage on complete ShapeNet55 point clouds with top-down image projections, followed by the cross-modal ShapeNet-ViPC stage. The baselines (XMFnet, EGIInet) are only reported by their published ShapeNet-ViPC numbers; the paper does not state whether they received any equivalent pretraining. If they were trained from scratch on ShapeNet-ViPC only, MambaTron's 0.012 CD average improvement and 0.003 F-score gain may be a pretraining advantage rather than an architectural one. The paper's own ablation (Table 4) does not include a no-pretraining condition, so this confound is untested. In addition, the claimed parameter count (3.92 M) is not derivable from the paper: no layer count, channel width, tokenizer size, or decoder configuration is given, and the same 'Ours' model in Table 3 reports 13.6 M parameters for ScanObjectNN with no reconciliation. The GPU-memory plot (Figure 4) has no methodology. Together these gaps make the headline efficiency/accuracy claim impossible to verify from the manuscript as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaTron, a hybrid cell that combines a bidirectional Mamba state-space layer with a block-transformer layer, and an Adjacency-Preserving Reordering (APR) scheme, and assembles these into a two-stage network for view-guided point cloud completion and unimodal point cloud pretraining. The authors report state-of-the-art or near-state-of-the-art results on ShapeNet-ViPC with only 3.92M parameters, competitive ScanObjectNN classification, and additional ModelNet40 and ShapeNetPart results in the supplement. They also provide ablations of the block-transformer, shared encoder, cross-modal encoder, losses, and APR.","tokens_in":14635,"tokens_out":6079,"duration_ms":51228,"significance":"If the reported numbers are reproducible under controlled conditions, the paper would make a useful empirical contribution: it is among the first to use a Mamba-based analogue of cross-attention in 3D vision, and the efficiency story (near-linear complexity, small parameter count) is attractive. The S4/Mamba mathematical background in Eqs. (1)-(3) is presented accurately, the APR idea is clearly motivated, and the supplementary ablations (APR components, loss variants) are a positive feature. However, the central performance and efficiency claims are currently not backed by a controlled comparison: the main tables lack variance, the pretraining protocol differs from the baselines, and the parameter counts are not reconcilable. The contribution is therefore significant conditional on additional verification.","major_comments":[{"comment":"The claimed advantage over EGIInet is confounded by an unequal pretraining protocol. The text states that MambaTron is first trained on complete ShapeNet55 point clouds with image projections at the unimodal stage and then fine-tuned on ShapeNet-ViPC, while the baseline numbers are taken from their published papers; there is no evidence that XMFnet or EGIInet received an equivalent pretraining stage. The average CD difference of 0.012 in the units of Table 1 and the F-score difference of 0.003 could plausibly be an effect of this extra pretraining rather than of the MambaTron architecture. The ablation in Table 4 does not include a no-pretraining condition, so this confound is untested. Please add at least one controlled condition: either train MambaTron from scratch on ShapeNet-ViPC and compare to the published baselines, or pretrain the baselines under the same ShapeNet55 protocol and report the comparison.","section":"§5.1, Tables 1–2"},{"comment":"The headline parameter count is neither derivable nor internally consistent. Section 5.1 reports 3.92M parameters for MambaTron on ShapeNet-ViPC, but the same 'Ours' model on ScanObjectNN in Table 3 reports 13.6M parameters, with no explanation of what differs. The paper gives no layer count, channel widths, tokenizer dimensions, number of MambaTron cells, or decoder configuration, so a reader cannot verify either number. Please provide a full parameter breakdown for all models, state the architecture configuration for each task, and reconcile the 3.92M/13.6M discrepancy.","section":"§5.1, Table 3"},{"comment":"All ShapeNet-ViPC numbers appear to be single runs. No standard deviations, seeds, or significance tests are reported for any metric. Given that the average improvement over EGIInet is 0.012 in the CD units of Table 1 and 0.003 in F-score, run-to-run variance could change the conclusion. Please report mean and standard deviation over at least three seeds for the proposed model and, ideally, for the baselines if rerun.","section":"§5.1, Tables 1–2"},{"comment":"The GPU-memory plot is not interpretable without a methodology. The figure compares MambaTron with Point-MAE and PointMamba as a function of input sequence length, but the manuscript does not state how memory was measured (peak vs. average), what hardware and batch size were used, which model configuration was tested, or how the baseline curves were obtained. Since the abstract and Section 5.2 advertise a 'fraction of the computation resources', this figure needs a complete measurement protocol or should be removed.","section":"Figure 4, §5.2"},{"comment":"Training details needed for reproducibility are missing. The two-stage procedure in Section 4.2 and the experimental protocols in Section 5 do not report optimizer, learning rate schedule, batch size, number of epochs, token masking ratio, loss weights for the terms in Eqs. (8)-(9), or the stack depth of MambaTron cells. Without these, the reported results and ablations cannot be reproduced or checked. Please include a full hyperparameter table and a statement about code availability.","section":"§4.2, §5.1"}],"minor_comments":[{"comment":"There are several typos: 'techninques' in the first paragraph of Section 1, 'theoritical' in Section 2.2, 'paramaters' in Section 2.3.2, and 'a seperate' in Section 5.3.","section":"§1, §2.2, §2.3.2, §5.3"},{"comment":"The phrase 'We observe comparable results comparable to the SOTA methods' repeats 'comparable'; please rewrite the sentence.","section":"§5.2"},{"comment":"The caption reads 'GPU usage comparision'; this should be 'GPU usage comparison'.","section":"Figure 4 caption"},{"comment":"The header 'w/o CD (avg)' is confusing. The first row should be labeled as the full model, and the column header should read 'Average CD (known / novel)' or similar.","section":"Table 4"},{"comment":"Please clarify the two '✓' columns in the APR ablation table: the text describes APR and the affine transform as separate components, but the column headers do not distinguish them.","section":"Supplementary Table 1"},{"comment":"The bullet 'slowing down the training as the loss function takes longer to converge' is a sentence fragment and should be merged into the preceding sentence.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is not currently supported by the evidence as presented, but the issues appear addressable with additional controlled experiments (from-scratch versus pretrained comparisons, multiple seeds, parameter and hyperparameter disclosure). The 3.92M versus 13.6M parameter discrepancy should be resolved before any acceptance decision. If the authors can provide these, the contribution may be publishable; in its current form it reads more like an extended technical report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the MambaTron cell is a reasonable and genuinely new combination for cross-modal point cloud completion, but the paper's headline claim—outperforming EGIInet with 3.92M parameters—is not supported by the evidence as reported. I agree with the stress-test note: the confound is real.\n\nWhat's good: the S4/Mamba equations are correct, and the idea of pairing a bidirectional Mamba layer with a block-transformer to compensate for Mamba's weak local reasoning is sensible. APR (a learnable affine reorder before XYZ sorting) is a small but legitimate contribution, and the paper tests the architecture on completion, classification, and segmentation, which is broader than most of these papers. The ablations in the supplementary at least show APR helps.\n\nSoft spots, in order of severity. First, the benchmark comparison is unfair as reported. Section 5.1 says the model is first pretrained on complete ShapeNet55 point clouds with projected views, then fine-tuned on ShapeNet-ViPC; the baselines' numbers are taken from their papers, with no indication they received equivalent pretraining. That alone could explain the 0.012 CD edge. Second, the parameter counts don't reconcile: 3.92M for the ViPC model vs 13.6M for the ScanObjectNN model. That may be due to different decoders, but the paper never says so. Third, there are no error bars or seeds anywhere; the F-score differences are 0.003–0.005, which nobody should trust from a single run. Fourth, the GPU memory plot (Figure 4) has no methodology and no comparison with a transformer baseline. Fifth, the ablation table (Table 4) is under-labeled; rows like \"block-tr\" and \"shared intra\" aren't self-explanatory. None of this kills the architectural idea, but it kills the current claim.\n\nWho this is for: researchers working on Mamba-based 3D models or cross-modal point cloud completion who want ideas for fusing modalities without quadratic attention. They should read it for the cell design, not for the benchmark numbers.\n\nI'd send it to peer review, but only with a clear expectation of major revision: add variance, reproduce with matched baselines (or re-train baselines with same pretraining), report code and hyperparameters, reconcile the parameter counts, and give the GPU plot a real methodology. The idea deserves a referee; the current evidence doesn't yet deserve acceptance.","headline":"MambaTron is a plausible architectural hybrid, but the SOTA claim is confounded by a two-stage pretraining mismatch and missing evaluation details.","tokens_in":15124,"tokens_out":3399,"would_cite":false,"duration_ms":29532,"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":"MambaTron, a hybrid Mamba–Transformer cell, reports state-of-the-art-level view-guided point cloud completion on ShapeNet-ViPC with 3.92M parameters, and its encoder transfers to point cloud classification and segmentation.","keywords":["point cloud completion","view-guided completion","state space models","Mamba","cross-modal fusion","block transformer","point cloud classification","ShapeNet-ViPC"],"falsifier":"Run the released MambaTron model and EGIInet on the same ShapeNet-ViPC test split with identical batch sizes and preprocessing, across at least five random seeds, and compare mean and standard deviation of Chamfer distance and F-score; if EGIInet's mean is within one standard deviation of MambaTron's, or lower, the claim of outperformance is falsified.","tokens_in":14111,"feed_emoji":"🧩","tokens_out":7703,"duration_ms":62477,"temperature":0.7,"pith_summary":"This paper claims that a hybrid cell mixing a bidirectional Mamba selective state-space layer with a small block-transformer can serve as the cross-modal bridge between an image and a partial point cloud, and that a network built from these cells performs view-guided point cloud completion at state-of-the-art accuracy while using fewer than half the parameters of the leading attention-based models. On the ShapeNet-ViPC benchmark, the proposed model reports an average Chamfer distance of 1.199×10⁻³ on known categories and 2.333×10⁻³ on novel categories, slightly better than EGIInet's 1.211 and 2.354, with F-scores also higher. The authors also show the same encoder transfers to unimodal point cloud classification and segmentation, reporting 90.17% accuracy on ScanObjectNN's PB_T50_RS split with 13.6M parameters. If these numbers hold, the paper's contribution is a concrete demonstration that state-space cross-attention can replace quadratic self/cross-attention in a multimodal 3D task without sacrificing accuracy.","feed_headline":"Point cloud completion bested with 3.92M parameters","feed_subtitle":"MambaTron pairs Mamba scanning with block attention to beat EGIInet on ShapeNet-ViPC using under half the parameters.","key_machinery":"The load-bearing object is the MambaTron cell: a bidirectional Mamba layer (a selective state-space model that scans the whole token sequence in both directions and returns per-token context states) followed by a Block-Transformer layer that splits tokens into blocks of size W=4 and applies local attention in parallel, with geometric center position (GCP) tokens added to the context. This cell is used three times: as a shared intra-modal encoder for image patches and point-cloud groups, and as a cross-modal encoder that concatenates image tokens, point-cloud tokens, and special tokens so each modality's embeddings absorb the other's context. Supporting the point-cloud path is Adjacency-Preserving Reordering (APR), which applies a learnable affine transform (TNet) followed by XYZ sorting before the sequence enters Mamba, giving an infinite family of adjacency-preserving orders. The training objective combines Chamfer distance, a style loss on gram matrices, a projection loss, and a 2D reconstruction loss, staged as unimodal pretraining then cross-modal fine-tuning.","core_discovery":"The central claim is that a MambaTron cell—a fully bidirectional Mamba SSM layer followed by a block-transformer layer operating on blocks of size 4—is enough to build both intra-modal encoders and a cross-modal encoder for view-guided point cloud completion. The model, trained in two stages (unimodal masked autoencoding with projection loss, then cross-modal completion with style loss), takes an incomplete point cloud plus a reference image, fuses them, and decodes a completed point cloud. On the ShapeNet-ViPC benchmark the authors report averages of 1.199 Chamfer distance (×10⁻³) on known categories and 2.333 on novel categories, against EGIInet's 1.211 and 2.354, with corresponding F-scores of 0.839/0.761 versus 0.836/0.750. The whole completion network uses 3.92M parameters, compared with 9.03M for EGIInet and 9.57M for XMFnet. The paper further claims the same pretraining recipe transfers to classification and segmentation, outperforming PointMamba on ScanObjectNN (90.17% vs 89.31% on PB_T50_RS) and reaching 94.3% OA on ModelNet40 and 87.4% class mIoU on ShapeNetPart.","pith_inferences":["A natural extension the paper leaves implicit is applying the MambaTron cell to other paired-modality tasks, such as text-to-point generation or audio-guided scene completion, since the cell itself is modality-agnostic beyond tokenization.","The paper reports single-run comparisons against published baselines; multi-seed runs with confidence intervals on the same test split would be the natural next check to separate the 0.012–0.021 Chamfer-distance margin from run-to-run noise.","The APR ordering depends on a learnable TNet; one could test whether the learned transformations transfer across categories or datasets, or whether a fixed larger set of random affine orders would give the same gain without the extra parameters."],"forward_implications":["If the reported numbers hold, view-guided point cloud completion no longer requires quadratic attention: the MambaTron cell gives subquadratic complexity, so completion can scale to longer point sequences and denser point clouds.","The parameter count drops from 9.03M (EGIInet) and 9.57M (XMFnet) to 3.92M, which would make state-of-the-art-level completion feasible on devices with tight memory budgets.","The same pretrained encoder reaches 90.17% on ScanObjectNN PB_T50_RS, 94.3% OA on ModelNet40, and 87.4% class mIoU on ShapeNetPart, suggesting the cross-modal objective yields transferable 3D representations for classification and segmentation.","APR's learnable affine reordering appears to be worth 0.118 Chamfer distance on known categories and 0.278 on novel ones, indicating that token ordering is a first-order design choice for Mamba-based point cloud encoders."],"supporting_citations":[{"why":"EGIInet is the state-of-the-art baseline the paper claims to outperform on ShapeNet-ViPC; supplies the numbers in Tables 1 and 2.","marker":"[42]"},{"why":"XMFnet is the transformer-based cross-attention baseline for view-guided completion; supplies the comparison for parameter count and accuracy.","marker":"[2]"},{"why":"ViPC introduced the view-guided completion task and the ShapeNet-ViPC dataset used for evaluation.","marker":"[49]"},{"why":"Mamba provides the selective state-space layer that forms half of the MambaTron cell.","marker":"[13]"},{"why":"PointMamba supplies the point-cloud tokenization scheme and the unimodal baseline compared in classification.","marker":"[24]"},{"why":"Joint-MAE supplies the two-stage pretraining recipe, projection loss, and decoder design.","marker":"[17]"},{"why":"Block-Recurrent Transformer is the recurrent block design the MambaTron cell is based on.","marker":"[22]"},{"why":"Block-State Transformers is the block-parallel SSM design the authors draw on for subquadratic complexity.","marker":"[6]"}],"fun_headline_variants":["MambaTron: 3.92M params to beat EGIInet on point cloud completion","Image-guided point cloud completion with a Mamba-Transformer cell","Under half the parameters, MambaTron outperforms EGIInet","MambaTron: Mamba cross-attention for point cloud completion","Efficient cross-modal point cloud enhancement with MambaTron"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire comparison against EGIInet and XMFnet assumes the published baseline numbers in Tables 1 and 2 were measured under comparable conditions to the authors' single run; no variance, seeds, or significance tests are reported, so the small Chamfer-distance margins could disappear under rerunning.","fun_headline_variants_meta":{"raw":{"variants":["MambaTron: 3.92M params to beat EGIInet on point cloud completion","Image-guided point cloud completion with a Mamba-Transformer cell","Under half the parameters, MambaTron outperforms EGIInet","MambaTron: Mamba cross-attention for point cloud completion","Efficient cross-modal point cloud enhancement with MambaTron"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001109,"raw_usage":{"total_tokens":4695,"prompt_tokens":1093,"completion_tokens":3602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":3502}},"tokens_in":709,"tokens_out":3602,"duration_ms":27006,"temperature":1.0,"reasoning_tokens":3502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:37:37.111868+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released MambaTron model and EGIInet on the same ShapeNet-ViPC test split with identical batch sizes and preprocessing, across at least five random seeds, and compare mean and standard deviation of Chamfer distance and F-score; if EGIInet's mean is within one standard deviation of MambaTron's, or lower, the claim of outperformance is falsified.","supporting_citations":[{"cited_title":"Explicitly guided information interaction network for cross-modal point cloud completion, 2024","cited_arxiv_id":null,"evidence_quote":"EGIInet is the state-of-the-art baseline the paper claims to outperform on ShapeNet-ViPC; supplies the numbers in Tables 1 and 2."},{"cited_title":"Cross-modal learning for image-guided point cloud shape completion","cited_arxiv_id":null,"evidence_quote":"XMFnet is the transformer-based cross-attention baseline for view-guided completion; supplies the comparison for parameter count and accuracy."},{"cited_title":"View-guided point cloud completion","cited_arxiv_id":null,"evidence_quote":"ViPC introduced the view-guided completion task and the ShapeNet-ViPC dataset used for evaluation."},{"cited_title":"Pointmamba: A simple state space model for point cloud analysis","cited_arxiv_id":null,"evidence_quote":"PointMamba supplies the point-cloud tokenization scheme and the unimodal baseline compared in classification."},{"cited_title":"Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training","cited_arxiv_id":null,"evidence_quote":"Joint-MAE supplies the two-stage pretraining recipe, projection loss, and decoder design."},{"cited_title":"Block-recurrent transformers, 2022","cited_arxiv_id":null,"evidence_quote":"Block-Recurrent Transformer is the recurrent block design the MambaTron cell is based on."},{"cited_title":"Block-state transformers, 2023","cited_arxiv_id":null,"evidence_quote":"Block-State Transformers is the block-parallel SSM design the authors draw on for subquadratic complexity."}],"review_version":1}