{"id":"45ea1b6b-2740-4f4b-9751-cf4e64200230","arxiv_id":"2505.20941","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"PMA adapts frozen point cloud models by ordering and fusing all intermediate layer features with Mamba, achieving parameter-efficient gains on ScanObjectNN, ModelNet40, and ShapeNetPart.","lead":"Point Mamba Adapter (PMA) is a parameter-efficient fine-tuning method that uses a Mamba state-space model to fuse intermediate features from all layers of a frozen point cloud backbone, guided by a geometry-constrained gate prompt. On classification, segmentation, and few-shot benchmarks it reports gains over full fine-tuning and prior PEFT methods with far fewer trainable parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed end-to-end optimization of the G2PG token ordering is unsupported: the one-hot + argmax reorder in §4.4.2 has zero gradient through the order path, so the task loss cannot train the permutation.","rationale":"The reader's weakest assumption is the non-differentiable G2PG reorder, and my reading confirms that this is the most load-bearing gap in the paper's argument. The headline results depend on a method whose core contribution is 'dynamically optimizing the spatial order'; if that path has zero gradient, the claimed end-to-end ordering mechanism is not actually trained, and the Table 7 comparison against static Hilbert/Z-order does not demonstrate what the text says it demonstrates. I considered the Table 1 vs Table 2 inconsistencies and the missing error bars; these are real quality issues, but they concern reporting of baselines rather than the central mechanism. The argmax issue directly attacks a stated novelty and the only ablation isolating it, so it is the single most consequential concern. The concern is addressable: code inspection and a gradient/ablation experiment can settle whether order optimization is real. Because the core empirical results may still hold, a CONDITIONAL verdict remains appropriate; no change to the reader's verdict is needed.","tokens_in":15502,"tokens_out":7636,"duration_ms":85654,"concrete_test":"Inspect the released code (github.com/zyh16143998882/PMA) to check whether the Argmax/One-hot reorder is detached or uses a differentiable relaxation (straight-through, soft sort, Gumbel-Softmax). Then run one targeted ablation on ScanObjectNN PB-T50-RS: (a) stop gradients through the order indices while keeping all other components identical; (b) replace the hard permutation with a differentiable soft-sort/Gumbel-Softmax with the same parameter budget. If (a) matches the published 86.43 and (b) does not improve it, the Table 7 gain is not due to end-to-end learned ordering. If (b) improves accuracy, the paper's stated mechanism is unsupported as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty includes a 'geometry-constrained dynamic sorting' strategy that is said to be optimized 'in an end-to-end manner' (Sec. 4.1, Sec. 5.5.2). However, Figure 3 and §4.4.2 specify that the order is obtained by applying Softmax, then One-hot encoding, then Argmax. Argmax is non-differentiable almost everywhere; using its output to permute tokens and then backpropagating through Mamba gives zero gradient to the permutation indices. Since the tokens T_i come from a frozen backbone, the G2PG parameters receive no gradient from the task loss through the reorder path; they are updated only through the geometric prompt path P_i. Thus the order can change during training indirectly as G2PG changes, but it is not optimized by gradient descent for the downstream task. The sole evidence for dynamic ordering's benefit is Table 7, where 'Ours' beats Z-order by 0.48 accuracy on PB-T50-RS. If the reorder is not actually trainable, this gain cannot be attributed to 'task-specific optimization of the ordering process' as claimed in Sec. 5.5.1. A straight-through estimator, Gumbel-Softmax, or differentiable sorting could make the claim true, but none is described. This concern does not by itself invalidate the reported accuracies, but it undermines a core mechanism and the ablation that supports it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Point Mamba Adapter (PMA), a parameter-efficient fine-tuning method for point cloud Transformers. PMA freezes the pre-trained backbone, collects token features from every layer, concatenates them into a long token sequence, and processes the sequence with a Mamba state-space model. A shared Geometry-constrained Gate Prompt Generator (G2PG) produces per-layer geometric prompts injected into Mamba's output matrix C and generates a token ordering via Softmax, one-hot encoding, and argmax, with the goal of dynamic, task-specific spatial ordering. The method is evaluated on classification (ScanObjectNN, ModelNet40), part segmentation (ShapeNetPart), and few-shot classification, using Point-BERT, Point-MAE, Recon, and PointGPT-L as frozen backbones, and reports state-of-the-art PEFT accuracy with 1 to 5 million trainable parameters. Code is publicly released.","tokens_in":15775,"tokens_out":7225,"duration_ms":70059,"significance":"The core idea of fusing intermediate features of a frozen backbone with a linear-complexity sequence model is sensible and orthogonal to existing point-cloud PEFT methods. The paper's strengths are its breadth of experiments (three tasks, four backbones, comparisons to full fine-tuning and several PEFT baselines), the public code release, and the motivating observation experiment in Sec. 4.1. If the dynamic ordering were genuinely optimized end-to-end, the G2PG mechanism would be a useful contribution. However, the current description does not support that claim, and the reported baseline numbers are internally inconsistent, so the central empirical claims need substantiation before the paper can be accepted.","major_comments":[{"comment":"The token reordering is described as Softmax followed by One-hot encoding and Argmax. Argmax is non-differentiable almost everywhere, and because the backbone is frozen, gradients from the task loss cannot reach the G2PG through the reorder path in Eqs. (7)-(8); the only gradient path is through the geometric prompt P_i. Therefore the ordering is not actually optimized by gradient descent for the downstream task, contrary to the claims of \"end-to-end\" optimization in Sec. 5.5.1 and Sec. 5.5.2. As a result, the 0.48-point advantage of \"Ours\" over Z-Order in Table 7 cannot be attributed to task-specific ordering optimization; it may come entirely from the prompt generation branch. Please specify a gradient estimator (e.g., straight-through estimator, Gumbel-Softmax, or differentiable sorting) or provide a controlled experiment that isolates the ordering effect from the prompt effect.","section":"Sec. 4.4.2, Fig. 3, Sec. 5.5.2"},{"comment":"The baseline numbers are internally inconsistent. Table 1 lists \"Point-MAE w/ IDPT\" with 2.7M trainable parameters and accuracies 95.18/93.29/90.63 on the three ScanObjectNN variants, values identical to the Recon FFT row, whereas Table 2 lists Point-MAE w/ IDPT with 1.7M parameters and accuracies 91.22/90.02/84.94. Similarly, PointGPT-L w/ PointGST is given as 0.6M parameters in Table 1 but 2.4M in Table 2, with the same accuracy values. Because IDPT, DAPT, and PointGST are the main PEFT baselines, these discrepancies directly affect the claimed rankings and the parameter-efficiency comparison. Please correct and reconcile all baseline rows, including #TP and accuracy, across Tables 1, 2, 3, and 5.","section":"Table 1 vs. Table 2"},{"comment":"The component ablation in Table 6 and the ordering-strategy comparison in Table 7 report single-run accuracies, with gaps as small as 0.48 points (86.43 vs. 85.95). The main classification and segmentation tables also report no variance or number of seeds. Given that few-shot results in Table 4 show standard deviations of 1-5 points, the small margins in Tables 1, 2, 3, and 7 may be within run-to-run noise. Please report mean and standard deviation over at least three seeds for the central comparisons, or explicitly state the experimental protocol used to obtain the reported numbers.","section":"Tables 6 and 7 (and main accuracy tables)"}],"minor_comments":[{"comment":"The dimensions of A, B, C, and D are stated inconsistently: Eqs. (1)-(4) define A in R^{S x S}, B in R^{1 x N}, C in R^{1 x N}, and D in R^{1 x N}, but Sec. 4.4.2 says the dimension of C is set to the number of patches, with S=128 as an example. Please clarify whether S denotes the Mamba state dimension or the number of tokens, and align the notation throughout.","section":"Preliminaries and Sec. 4.4.2"},{"comment":"The observation experiment in Sec. 4.1 cites drops of 2.6% and 0.14% when using the first 3 layers, but Fig. 1 does not show absolute accuracy values or error bars, making it difficult to assess the significance of these drops. Please include numerical values and, if possible, variance in the figure or its caption.","section":"Sec. 4.1, Fig. 1"},{"comment":"Equation (9) refers to features F_pre from the first N-1 layers, but the architecture description in Sec. 4.3 uses L layers and never defines N. Please replace N with L or define N explicitly.","section":"Eq. (9)"},{"comment":"Some references are incomplete or inconsistently formatted, e.g., PointGPT [2] gives only \"2024\" without a venue, and several arXiv preprints do not include arXiv identifiers. Please normalize the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The non-differentiable argmax ordering issue is the key technical concern: if the authors cannot provide a gradient estimator or reframe the ordering as input-dependent but not task-optimized, the central novelty is substantially weakened and Table 7's interpretation must change. The Table 1 vs. Table 2 inconsistency is serious and should be resolved before any conditional acceptance; please ask for a detailed rebuttal and, ideally, the exact experimental setup and code for the reordering ablation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical PEFT paper with a real new combination — fusing all intermediate layer features from a frozen point cloud transformer using a Mamba adapter, plus a shared geometry-constrained gate prompt generator. The results are consistently positive across classification, part segmentation, and few-shot on multiple backbones. That part deserves attention.\n\nWhat is genuinely new: previous point cloud PEFT methods (IDPT, DAPT, PointGST) either use prompts/adapters per layer or operate on final features. PMA explicitly concatenates features from every layer into one long sequence and processes it with Mamba, which is a sensible way to keep complexity linear in the number of tokens. The G2PG idea — using point cloud geometry to generate both output-matrix prompts and a token order — is also new. The experiments are broad: three backbones, three ScanObjectNN variants, ModelNet40, ShapeNetPart, few-shot. With 4.9M trainable parameters on PointGPT-L they beat full fine-tuning on PB-T50-RS by 1.78 points. That is a practically useful result.\n\nNow the soft spots, in order of severity.\n\nFirst, the end-to-end training claim for the dynamic ordering does not hold as stated. The order is produced by softmax -> one-hot -> argmax. Argmax has zero gradient, and the paper does not describe a straight-through estimator or any relaxation. So the permutation indices receive no gradient from the task loss. The G2PG parameters do get gradients through the geometric prompt path P_i, so the order changes during training, but it is not optimized by gradient descent for the downstream task. The ablation in Table 7 (86.43 vs 85.95 for Z-order) therefore cannot be attributed to 'task-specific optimization' of the ordering. This needs to be fixed — either by using Gumbel-Softmax, differentiable sorting, or by significantly tempering the claim.\n\nSecond, the baseline numbers are inconsistent between Table 1 and Table 2. The most glaring example: Point-MAE w/ IDPT is 95.18 / 93.29 / 90.63 in Table 1 but 91.22 / 90.02 / 84.94 in Table 2. The Table 1 values look like they were copied from the Recon full fine-tuning row. This is a serious reporting error; it makes the comparison against existing PEFT methods unreliable until corrected.\n\nThird, the main classification tables have no error bars or multiple runs. Few-shot has standard deviations, which is good, but the central claims on ScanObjectNN and ModelNet40 would be stronger with variance reported.\n\nOverall: the core architecture and results are plausible and the direction is worth pursuing. The flaws are addressable and do not obviously invalidate the main story, but the current version overclaims a mechanism that is not actually trained end-to-end, and the baseline inconsistency must be fixed. I would send this to a serious referee — with a note to check Tables 1/2 and to ask the authors to clarify the gradient path for the ordering. I would not cite it in its current form.","headline":"Useful parameter-efficient adapter for point clouds, but the headline ordering mechanism is not actually optimized end-to-end the way the paper claims.","tokens_in":16303,"tokens_out":3246,"would_cite":false,"duration_ms":31971,"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":"Fusing all layers of a frozen pre-trained point-cloud model with a Mamba adapter lets parameter-efficient fine-tuning beat full fine-tuning while using 99% fewer trainable parameters.","keywords":["point cloud understanding","parameter-efficient fine-tuning","Mamba","state space models","intermediate features","3D perception","feature fusion"],"falsifier":"Inspect the released code for the reorder branch: if the argmax-selected indices are detached or no straight-through estimator is used, then the ordering is not optimized by gradient descent; retraining on PB-T50-RS with the learned ordering replaced by a fixed random permutation would then show whether the reported 86.43% accuracy actually depends on the dynamic order.","tokens_in":15313,"feed_emoji":"🧊","tokens_out":6822,"duration_ms":63063,"temperature":0.7,"pith_summary":"This paper tries to show that a frozen pre-trained point-cloud model holds usable information in every layer, not just the last one, and that a lightweight Mamba-based adapter can harvest it. The proposed Point Mamba Adapter (PMA) concatenates tokens from all layers into one ordered sequence, reorders them with a geometry-constrained gate prompt generator (G2PG), and fuses them with a state-space model. Across ScanObjectNN, ModelNet40, ShapeNetPart, and few-shot benchmarks, PMA is reported to beat full fine-tuning and prior parameter-efficient methods while training only 4.9M parameters on the 360M-parameter PointGPT-L backbone. If true, this makes large pre-trained 3D models much cheaper to adapt, especially for fine-grained tasks like part segmentation.","feed_headline":"A 4.9M-parameter adapter beats full fine-tuning on point clouds","feed_subtitle":"Fusing every layer of a frozen backbone lifts few-shot and segmentation accuracy with 99% fewer parameters.","key_machinery":"The central object is the Point Mamba Adapter, built on Mamba, a selective state-space sequence model whose recurrence is $h_t = A h_{t-1} + B x_t$ with output $y_t = (C+P)h_t + D x_t$. The adapter fuses the $L \\times M$ patch tokens collected from all layers of the frozen backbone. The second load-bearing object is the geometry-constrained gate prompt generator (G2PG), which builds a KNN geometric graph over tokens, aggregates neighbor features with down-projection and max pooling, applies softmax, and then produces two things: geometric prompts $P$ added to the output matrix $C$, and a one-hot plus argmax index that reorders tokens before fusion. The paper claims this injects spatial-neighborhood constraints into the sequence model and makes the ordering task-adaptable.","core_discovery":"The central claim is that intermediate features from a frozen pre-trained backbone are nearly as informative as the final features and contain complementary information, so discarding them wastes most of the model's representational power. PMA is the proposed fix: extract patch-token outputs from every Transformer layer, feed them through a shared G2PG that produces both a geometric prompt for Mamba's output matrix and a dynamically learned token order, concatenate all layers' tokens chronologically, and run a Mamba adapter whose output state is modulated by the geometric prompt. The fused output is combined with the last-layer CLS token and final features before the task head. The paper reports that this arrangement beats full fine-tuning and earlier PEFT methods across classification, part segmentation, and few-shot learning, with a 99% reduction in trainable parameters on PointGPT-L, and that the dynamic geometry-constrained ordering outperforms fixed Hilbert, z-order, and axis-orderings.","pith_inferences":["Editorial inference: because one-hot plus argmax has no gradient, the claim of end-to-end learned ordering depends on an unstated gradient approximation such as a straight-through estimator; checking the released code for that estimator is the fastest way to test the paper's main novelty.","Editorial inference: the concatenate-all-layers-and-scan design is not specific to point clouds, so the same adapter recipe could be transferred to other frozen architectures with layer-wise patch tokens, such as 2D or video transformers.","Editorial inference: the sequence length grows as $L \\times M$, so a practical deployment of PMA should be measured for latency and memory on long point clouds; the linear-time advantage of Mamba may shrink when the fused sequence is very long."],"forward_implications":["If PMA's claims hold, parameter-efficient fine-tuning can close most of the gap to full fine-tuning on fine-grained point-level tasks like part segmentation.","The 360M-parameter PointGPT-L backbone can be adapted with only 4.9M trainable parameters while improving classification accuracy on ScanObjectNN and ModelNet40.","Point-cloud PEFT methods that use only final-layer features leave measurable performance unused; multi-layer fusion becomes a standard component of future adapters.","The learned geometry-constrained ordering gives a reported gain over Hilbert, z-order, and axis-orderings (86.43 versus about 85.9 on the PB-T50-RS variant), so ordering strategy matters for sequence-based fusion.","PMA transfers across different pre-trained backbones, including Point-BERT, Point-MAE, Recon, and PointGPT-L, suggesting the approach is not tied to one pre-training method."],"supporting_citations":[{"why":"Supplies the selective state-space model (Mamba) that PMA uses to fuse the ordered multi-layer token sequence.","marker":"[8]"},{"why":"The Point-MAE pre-trained backbone used for the intermediate-feature observation study and for most PEFT comparisons.","marker":"[32]"},{"why":"The PointGPT-L pre-trained model used as the main large backbone; its 360M-parameter full fine-tuning is the baseline PMA outperforms with 4.9M parameters.","marker":"[2]"},{"why":"IDPT, an existing point-cloud prompt-tuning PEFT method that PMA compares against on classification and segmentation.","marker":"[52]"},{"why":"DAPT, an existing point-cloud adapter-plus-prompt PEFT method that PMA compares against and whose evaluation protocol it follows.","marker":"[62]"},{"why":"PointGST, the current point-cloud PEFT baseline that PMA matches or exceeds on ScanObjectNN variants.","marker":"[24]"},{"why":"Recon, a contrastive-generative pre-trained backbone used in part-segmentation and few-shot evaluations.","marker":"[35]"},{"why":"Point-BERT, a masked-point-modeling backbone used across the classification, part-segmentation, and few-shot tables.","marker":"[49]"},{"why":"PointMamba, an earlier state-space model for point clouds that motivates the use of Mamba for point-cloud sequences.","marker":"[25]"}],"fun_headline_variants":["4.9M-param Mamba adapter beats full point tuning","Fusing every layer: 99% fewer params, higher point accuracy","Point Mamba Adapter: learn order, fuse layers, win","All-layer fusion with Mamba outdoes full fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned order is computed by a hard “take the maximum” step over one-hot codes, and the paper does not say how gradients pass through it; if they do not, the ordering is not truly trained end-to-end and the reported benefit of dynamic ordering over fixed Hilbert or z-order curves is not established.","fun_headline_variants_meta":{"raw":{"variants":["4.9M-param Mamba adapter beats full point tuning","Fusing every layer: 99% fewer params, higher point accuracy","Point Mamba Adapter: learn order, fuse layers, win","All-layer fusion with Mamba outdoes full fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1345,"prompt_tokens":955,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":316}},"tokens_in":571,"tokens_out":390,"duration_ms":4924,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:42:33.897483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code for the reorder branch: if the argmax-selected indices are detached or no straight-through estimator is used, then the ordering is not optimized by gradient descent; retraining on PB-T50-RS with the learned ordering replaced by a fixed random permutation would then show whether the reported 86.43% accuracy actually depends on the dynamic order.","supporting_citations":[{"cited_title":"Masked autoencoders for point cloud self-supervised learning","cited_arxiv_id":null,"evidence_quote":"The Point-MAE pre-trained backbone used for the intermediate-feature observation study and for most PEFT comparisons."},{"cited_title":"Pointgpt: Auto-regressively generative pre- training from point clouds","cited_arxiv_id":null,"evidence_quote":"The PointGPT-L pre-trained model used as the main large backbone; its 360M-parameter full fine-tuning is the baseline PMA outperforms with 4.9M parameters."},{"cited_title":"Instance-aware dynamic prompt tuning for pre-trained point cloud models","cited_arxiv_id":null,"evidence_quote":"IDPT, an existing point-cloud prompt-tuning PEFT method that PMA compares against on classification and segmentation."},{"cited_title":"Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis","cited_arxiv_id":null,"evidence_quote":"DAPT, an existing point-cloud adapter-plus-prompt PEFT method that PMA compares against and whose evaluation protocol it follows."},{"cited_title":"Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining","cited_arxiv_id":null,"evidence_quote":"Recon, a contrastive-generative pre-trained backbone used in part-segmentation and few-shot evaluations."},{"cited_title":"Point-bert: Pre-training 3d point cloud transformers with masked point modeling","cited_arxiv_id":null,"evidence_quote":"Point-BERT, a masked-point-modeling backbone used across the classification, part-segmentation, and few-shot tables."}],"review_version":1}