{"id":"15e4b182-62e2-4cf5-acae-f50e1e12d34a","arxiv_id":"2507.18997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"UPP unifies denoising and completion as point-level prompts in a frozen pre-trained point cloud transformer, improving classification accuracy on noisy and incomplete point clouds without updating backbone weights.","lead":"This paper introduces UPP, a parameter-efficient fine-tuning method that treats point cloud denoising and completion as point-level prompts, letting a frozen pre-trained model classify noisy or incomplete 3D scans more accurately. The method reports gains over full fine-tuning on several corrupted point cloud benchmarks while training only 1.4 million parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"UPP's headline gain may reflect an unmatched training budget: up to 50 extra epochs plus denoising/completion supervision that baselines do not receive, so the superiority claim is not yet isolated.","rationale":"The reader's weakest assumption correctly identifies the training-budget mismatch; my read converges on the same point and adds that the auxiliary losses are a second, equally unmatched resource. UPP's Eq. 14 uses clean point-cloud ground truth for Lrect and Lcomp; the comparison methods only use classification labels. On a benchmark where the test distribution is exactly the corruption used to generate those auxiliary targets, this is a substantial experimental advantage. The staged optimization text directly conflicts with the 'identical training strategies' statement in the supplementary. The parameter-efficiency numbers (1.4M, roughly 6% of backbone) are credible and well documented in Table 8, and the code release is a positive check. But the central claim is not just 'fewer parameters'; it is 'higher accuracy on noisy/incomplete data with fewer parameters'. If that accuracy gap requires 50 extra epochs and clean supervision, the claim needs to be restated as a trade-off, not a dominance. Hence I keep the CONDITIONAL verdict: the paper should be accepted only if the authors supply the matched-budget ablation or explicitly reframe the contribution.","tokens_in":15742,"tokens_out":5918,"duration_ms":61638,"concrete_test":"Ablate UPP on Noisy ModelNet40 to a single-stage, task-only variant: train for exactly 300 epochs with L = Ltask (drop Lrect and Lcomp and the 50-epoch staged warm-start), keeping all other settings from Table 4. Compare the accuracy with Table 1's 92.95 and with the 89.42 full fine-tuning row; repeat for 3 seeds. If the variant falls to roughly 89–90%, the reported superiority depends on the unmatched extra epochs and geometric supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires a fair comparison: UPP must beat full fine-tuning and PEFT because of its point-level prompting, not because it is given more training resources. The supplementary's 'Staged Optimization Strategy' says 'We add 50 epochs to optimize the point-level promoters' (20 for both prompters, then 30 for the Completion Prompter), while Table 4 fixes training epochs at 300 for all methods and the text claims 'identical hyper-parameters and training strategies are applied across fine-tuning and proposed methods'. That is internally inconsistent. In addition, Eq. 14 sums Lrect + Lcomp + Ltask: the rectification loss (Eq. 4) and completion loss (Eq. 9) use clean ground-truth point clouds and missing-part supervision, which full fine-tuning, DAPT, and Point-PEFT do not receive. On Noisy ModelNet40 the reported margin over full fine-tuning is 3.53 points (92.95 vs 89.42); 50 extra epochs and dense geometric supervision on exactly the corruption type being tested are both plausible enough explanations that the prompting mechanism is not the only candidate. Consequently the paper's efficiency story ('more accuracy with 95% fewer trainable parameters') conflates parameter count with total training budget; the later resource is not matched.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UPP, a parameter-efficient fine-tuning (PEFT) method that reformulates point cloud denoising and completion as point-level prompting for downstream classification and segmentation. Three modules are introduced: a Rectification Prompter (Eqs. 1-4) that moves noisy points along predicted vectors, a Completion Prompter (Eqs. 5-9) that reconstructs missing regions using the repurposed MAE decoder, and a Shape-Aware Unit (Eqs. 10-12) that adds spatial-distance-based attention and adapters into each frozen transformer block. Training combines the task loss with rectification and completion losses (Eq. 14). Experiments on synthetic Noisy ModelNet40, Noisy ShapeNet55, real-world ScanObjectNN, and noisy ShapeNetPart report that UPP outperforms full fine-tuning and prior PEFT methods with only 1.4M trainable parameters. The supplementary details the training schedule, additional ablations, and a clean-data PEFT comparison.","tokens_in":16058,"tokens_out":3712,"duration_ms":37606,"significance":"If the reported comparisons are valid, the contribution is a practical one: a single frozen-backbone framework that performs enhancement and analysis jointly, with clearly specified losses and a released codebase. The explicit mathematical formulation of the losses and the careful description of the prompting mechanism are strengths, as is the evaluation across multiple backbone choices and noise regimes. However, the central efficiency-and-accuracy claim is not yet isolated from the training budget and auxiliary supervision the method receives; the margin over full fine-tuning could plausibly be driven by extra epochs and dense geometric losses rather than by the prompting design. The paper is therefore promising but requires a substantially fairer experimental protocol before its headline claims are supported.","major_comments":[{"comment":"The training budget for UPP is not matched to the baselines. The supplementary states: 'We add 50 epochs to optimize the point-level promoters' (20 epochs for both prompters, then 30 for the Completion Prompter), while Table 4 fixes Training epochs at 300 for all methods and the text claims 'identical hyper-parameters and training strategies are applied across fine-tuning and proposed methods.' These statements are internally inconsistent: UPP receives up to 350 epochs in total. Since Table 1 reports margins of +3.53 over full fine-tuning on Noisy ModelNet40 and +2.27 on Noisy ShapeNet55, an additional 50 epochs of training is a plausible alternative explanation for the improvement. This is load-bearing because the paper's central selling point is higher accuracy with 95% fewer trainable parameters; trainable parameter count is not the same as total training budget. Please rerun all baselines under a matched total-epoch budget (e.g., 350 epochs for every method), or restrict UPP to 300 total epochs including the prompter warm-up, and report the resulting numbers.","section":"Supplementary 'Staged Optimization Strategy' and Table 4"},{"comment":"The total loss L = Lrect + Lcomp + Ltask gives UPP dense geometric supervision that the baselines do not receive. Lrect uses the clean point cloud as target and a separate noisy-point set, while Lcomp uses the ground-truth complete point cloud Pgt and the missing part Pm; full fine-tuning, DAPT, and Point-PEFT are trained with cross-entropy alone. Because UPP's synthetic corruption during training (24 outliers, 64 surface noise, 25% missing points, and supplementary random cropping of 25%-50%) matches the corruption at test time, the gain could come from the auxiliary losses on the test distribution rather than from the point-level prompting mechanism. Please add an ablation in which the baselines receive analogous auxiliary denoising/completion supervision (or in which UPP is trained with only Ltask) to isolate the contribution of the prompting design.","section":"Eq. (14) and the auxiliary losses in Eqs. (4) and (9)"},{"comment":"No error bars, seeds, or statistical significance tests are reported anywhere in the paper or supplement, and the ablation design in Table 3 and the prompting-order ablation (Table 7) are selected on Noisy ModelNet40, the same benchmark that produces the headline numbers. With component increments of 1.50%, 1.50%, and 1.54% in Table 3 and margins as small as +0.67 in Table 1, single-run differences may be within run-to-run variance. Please report mean and standard deviation over at least three seeds for the main tables and the key ablations, and specify whether the model selection is performed on a validation split or on the test set.","section":"Section 5.4 and all experimental tables"},{"comment":"On Noisy ShapeNetPart, UPP (82.2/84.4 with Point-MAE and 82.5/84.8 with Point-FEMAE) does not surpass full fine-tuning (83.3/85.6 and 83.5/85.9), and the text acknowledges that PEFT methods including UPP 'exhibit greater susceptibility to noise and incompleteness compared to full fine-tuning.' This is stated in the supplement but not mentioned in the main paper's conclusion, which claims general 'superiority' over existing methods. The main text should either include the segmentation results or explicitly qualify the claim to the classification benchmarks.","section":"Supplementary Table 5"}],"minor_comments":[{"comment":"The Point-FEMAE row is duplicated, and PCP-MAE is labeled with reference [41] although it is reference [43] in the bibliography; please correct the table and the citation indices.","section":"Table 2"},{"comment":"The grouping is confusing: the Point-MAE row listed under 'Full Fine-Tuning (FFT)' and the 'Point-MAE (baseline)' row under 'Parameter-Efficient Fine-Tuning' are identical in the table yet represent different training regimes; please restructure the table so that the baseline configuration is unambiguous.","section":"Table 1"},{"comment":"The phrase 'we optionally enable the training of the two point-level promoters with the Shape-Aware Unit when the learning rate narrows to 0.0001' is vague; please specify exactly which epochs or which LR milestones trigger this phase and whether the 50 extra epochs are included in the 300 reported in Table 4.","section":"Supplementary 'Staged Optimization Strategy'"},{"comment":"The y-axis range and the numeric labels under each outlier-noise level are difficult to read, and the relationship between the plotted percentages and the 'Outlier Noise Number / Input Resolution' axis label is unclear; please annotate the figure more precisely.","section":"Figure 3"},{"comment":"The abstract and introduction describe the conventional denoising/completion ensemble as 'failing' and 'inefficient,' but the quantitative comparison to such ensemble methods is not shown; adding one such comparison, even in the supplement, would strengthen the positioning relative to the described baseline.","section":"Section 2.2 and abstract"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test concern: the skipped training-budget match is real and load-bearing. The paper's internal inconsistency between Table 4 and the supplementary staged-optimization text means the headline margins are not yet trustworthy. The fix is straightforward experimentally (matched total epochs, matched auxiliary supervision, multiple seeds), so a major revision rather than rejection seems appropriate. The benchmark-tuning worry is also present: the corruption settings used for training are identical to the test corruption, so the robustness claims would be much stronger if some generalization to unseen noise levels or unseen corruption types were demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: UPP is a genuinely new architecture that reformulates denoising and completion as input-space point prompts for frozen backbones, and it reports consistent accuracy gains over full fine-tuning and PEFT baselines across three backbones on noisy/incomplete classification. The code is released, the equations are coherent, and the main numbers are internally consistent. I think the method itself is worth taking seriously.\n\nWhat's new: combining a rectification prompter (KNN-filtered displacement vectors) with a completion prompter (MAE decoder reuse) and a spatial-distance-based shape-aware attention, all on top of a frozen transformer, is not in the cited Point-PEFT, DAPT, or IDPT papers. The per-component ablation on Noisy ModelNet40 gives each module credit.\n\nWhere it gets soft: the training budget. The supplementary says \"We add 50 epochs to optimize the point-level promoters\" (20 both, then 30 completion only), and also says \"optionally enable\" them later with the SA-Unit when LR narrows. Yet Table 4 says \"identical hyper-parameters and training strategies are applied across fine-tuning and proposed methods.\" That's flatly inconsistent. Baselines are trained for 300 epochs on the same task. UPP gets up to 50 extra epochs plus two auxiliary losses (Eq. 4 and Eq. 9) that use clean ground-truth surfaces and missing-part supervision on exactly the corruption types used at test time. So the 3.53-point margin over full fine-tuning on Noisy ModelNet40 could easily be explained by the extra budget and supervision, not by the prompting mechanism. The paper's efficiency story (“more accuracy with 95% fewer parameters”) conflates parameter count with total training cost, and the latter isn't matched. That's the load-bearing issue and it's addressable: retrain baselines with the same epochs and auxiliary losses, or run UPP without the staged pre-training.\n\nMinor soft spots: no error bars or multiple seeds; the prompting-order ablation (Table 7) is selected on the test set; the segmentation results in Table 5 show UPP does not beat full fine-tuning, yet the abstract and conclusion claim broad superiority without qualification; and there's no comparison with the authors' own GAPrompt (arXiv:2505.04119), which is reviewed in the related work.\n\nBottom line: this paper is for researchers working on parameter-efficient adaptation of 3D transformers. The architecture deserves a careful referee. I'd accept it for peer review with the expectation that the authors equalize training budgets, add variance estimates, and soften the claims to match the segmentation evidence. I'd cite it for the novelty of point-level prompting, not for the absolute numbers.\n\nRecommendation: send to peer review; require the matched-budget experiments before acceptance.","headline":"UPP is a genuinely new point-level prompting architecture that reports consistent gains on noisy/incomplete point cloud classification, but the headline comparison gives it up to 50 extra training epochs plus dense geometric supervision the baselines don't get, so the efficiency claim isn't isolated.","tokens_in":16576,"tokens_out":2935,"would_cite":true,"duration_ms":27911,"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":"The paper claims that reformulating point cloud denoising and completion as point-level prompts lets a frozen pre-trained backbone beat full fine-tuning on noisy and incomplete classification with only 1.4M trainable parameters.","keywords":["point cloud analysis","parameter-efficient fine-tuning","point-level prompting","denoising","completion","3D classification","noisy point clouds","Shape-Aware Attention"],"falsifier":"Run the Table 1 comparison with identical total epochs for every method and the same auxiliary geometric losses; if UPP's Noisy ModelNet40 accuracy drops to or below the 89.42% full-fine-tuning baseline, the advertised gain comes from the extra training budget, not from point-level prompting.","tokens_in":1682,"feed_emoji":"🧊","tokens_out":2395,"duration_ms":78715,"temperature":0.7,"pith_summary":"Pre-trained point cloud models lose accuracy when input scans are noisy or missing points, a common situation with real 3D sensors. The paper proposes UPP, a parameter-efficient fine-tuning method that treats denoising and completion as point-level prompts rather than as separate pre-processing models. A Rectification Prompter predicts per-point displacement vectors and masks unreliable points, a Completion Prompter generates missing points using a repurposed pre-trained decoder, and a Shape-Aware Unit adapts the frozen backbone's features through spatial-distance attention. With only 1.4 million trainable parameters, UPP reports higher classification accuracy than full fine-tuning on Noisy ModelNet40, Noisy ShapeNet55, and ScanObjectNN, across three different pre-trained backbones. The paper also shows that applying rectification before completion, and the addition of the Shape-Aware Unit, account for most of the gain.","feed_headline":"Point-level prompts beat full fine-tuning on noisy 3D scans","feed_subtitle":"Frozen-backbone UPP hits 92.95% on Noisy ModelNet40 with 1.4M trainable parameters, beating full fine-tuning.","key_machinery":"The machinery is three lightweight modules inserted into a frozen pre-trained point cloud transformer: the Rectification Prompter, which estimates per-point displacement vectors $v_r \\in \\mathbb{R}^3$ and masks points with $\\|v_r\\| > \\tau$; the Completion Prompter, which reuses the discarded decoder of the masked autoencoder to produce coarse centers and dense point prompts $x_m$; and the Shape-Aware Unit, which adds trainable prompt tokens $k_i$ to each block and combines feature-similarity attention with spatial-distance K-nearest-neighbor interpolation. The completion losses are Chamfer distances between predicted and ground-truth clouds, the rectification loss is an L2 distance to the nearest clean-surface points, and the total loss is the sum of the two plus the downstream cross-entropy. This design lets the model move and add points in the input space while keeping the backbone weights frozen.","core_discovery":"UPP's central claim is that denoising and completion should not be separate pre-processing stages: they can be expressed as point-level prompts inside the analysis model, so the enhancement and the downstream task share one frozen backbone. Given a noisy, incomplete point cloud, the Rectification Prompter predicts a per-point rectification vector, keeps only points whose predicted displacement magnitude is below a threshold, and moves those points toward the estimated clean surface. The Completion Prompter then repurposes the pre-trained masked-autoencoder decoder to predict coarse missing centers and dense auxiliary points, and the fused point set is resampled with farthest point sampling. The Shape-Aware Unit prepends prompt tokens in every transformer block and augments feature-similarity attention with spatial-distance K-nearest-neighbor attention, which the paper argues is naturally insensitive to outliers. On Noisy ModelNet40 with the Point-MAE backbone this reaches 92.95% accuracy with 1.4M trainable parameters, versus 89.42% for full fine-tuning and 86.43% for DAPT.","pith_inferences":["A testable extension is to isolate the source of the gain by giving DAPT and full fine-tuning the same auxiliary Chamfer-distance supervision and extra epochs; if the gap narrows, the contribution is partly the auxiliary geometric loss rather than the prompting mechanism itself.","The masking of low-reliability points by a hard threshold could be replaced with a soft, differentiable weighting, which might preserve more thin structures such as chair legs or table edges.","The same point-level prompting recipe could be applied to other corruption types, such as varying occlusion rates or sensor-specific noise patterns, and to 3D detection, where noisy partial scans are also common."],"forward_implications":["If UPP's results hold, freezing the backbone while injecting point-level prompts is enough to exceed full fine-tuning on noisy and incomplete classification, so full fine-tuning is not necessary for robustness on these benchmarks.","The method's 1.4M trainable parameters, a reduction of more than 95% versus full fine-tuning, imply that the accuracy gain does not require task-specific denoising and completion networks with their own feature extractors.","The ordering result, where rectification first reaches 92.95% versus 91.18% for completion first, implies that cleaning input geometry before completing the shape is a design principle for unified enhancement pipelines.","On clean data, the Shape-Aware Unit alone reaches 94.2% on ModelNet40 with a Point-FEMAE backbone and 0.6M parameters, suggesting the module is also a general parameter-efficient fine-tuning method, not only a noise-handling device.","Part segmentation results on noisy ShapeNetPart show UPP beats other PEFT methods but still trails full fine-tuning in fine-grained tasks, so the claimed advantage is strongest for classification."],"supporting_citations":[{"why":"Provides the Point-MAE pre-trained transformer that UPP freezes and adapts, and the full fine-tuning baseline it outperforms.","marker":"[26]"},{"why":"DAPT is the main PEFT baseline in Tables 1 and 2, and the method UPP surpasses by several accuracy points.","marker":"[44]"},{"why":"Point-PEFT is the other 3D PEFT baseline included in the main comparisons.","marker":"[32]"},{"why":"ReCon is a second pre-trained backbone used to show UPP transfers across backbones.","marker":"[28]"},{"why":"Point-FEMAE is a third pre-trained backbone, also used for the clean-data PEFT comparison in the supplement.","marker":"[41]"},{"why":"ModelNet40 is the source of the Noisy ModelNet40 benchmark with injected noise and 25% missing points.","marker":"[34]"},{"why":"ShapeNet55 is the source of the Noisy ShapeNet55 benchmark used in the main classification experiments.","marker":"[3]"},{"why":"ScanObjectNN provides real-world scanned objects with inherent noise for the incomplete real-data test.","marker":"[33]"},{"why":"PoinTr is the source of the online cropping procedure used to simulate incompleteness.","marker":"[38]"},{"why":"ScoreDenoise is cited for the surface and outlier noise injection procedure used to build the noisy benchmarks.","marker":"[24]"}],"fun_headline_variants":["Unified point prompts make point cloud models robust to noise and gaps","Point-level prompting beats full fine-tuning on noisy 3D scans","UPP: prompting for denoising and completion in one frozen backbone","Prompt-based rectification and completion for robust point clouds","Parameter-efficient prompting tames noisy and incomplete 3D data"],"cache_read_input_tokens":18688,"weakest_assumption_plain":"The comparison assumes a matched training budget, but the staged optimization gives UPP up to 50 extra epochs plus auxiliary Chamfer supervision that the baselines do not receive, so the accuracy margin may partly reflect extra training rather than the prompting mechanism alone.","fun_headline_variants_meta":{"raw":{"variants":["Unified point prompts make point cloud models robust to noise and gaps","Point-level prompting beats full fine-tuning on noisy 3D scans","UPP: prompting for denoising and completion in one frozen backbone","Prompt-based rectification and completion for robust point clouds","Parameter-efficient prompting tames noisy and incomplete 3D data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1597,"prompt_tokens":1059,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":450}},"tokens_in":675,"tokens_out":538,"duration_ms":4729,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:03:50.272292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Table 1 comparison with identical total epochs for every method and the same auxiliary geometric losses; if UPP's Noisy ModelNet40 accuracy drops to or below the 89.42% full-fine-tuning baseline, the advertised gain comes from the extra training budget, not from point-level prompting.","supporting_citations":[{"cited_title":"Masked autoencoders for point cloud self-supervised learning","cited_arxiv_id":null,"evidence_quote":"Provides the Point-MAE pre-trained transformer that UPP freezes and adapts, and the full fine-tuning baseline it outperforms."},{"cited_title":"Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis","cited_arxiv_id":null,"evidence_quote":"DAPT is the main PEFT baseline in Tables 1 and 2, and the method UPP surpasses by several accuracy points."},{"cited_title":"Point- peft: Parameter-efficient fine-tuning for 3d pre-trained mod- els","cited_arxiv_id":null,"evidence_quote":"Point-PEFT is the other 3D PEFT baseline included in the main comparisons."},{"cited_title":"Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining","cited_arxiv_id":null,"evidence_quote":"ReCon is a second pre-trained backbone used to show UPP transfers across backbones."},{"cited_title":"Towards compact 3d representations via point feature enhancement masked au- toencoders","cited_arxiv_id":null,"evidence_quote":"Point-FEMAE is a third pre-trained backbone, also used for the clean-data PEFT comparison in the supplement."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"ModelNet40 is the source of the Noisy ModelNet40 benchmark with injected noise and 25% missing points."},{"cited_title":"Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data","cited_arxiv_id":null,"evidence_quote":"ScanObjectNN provides real-world scanned objects with inherent noise for the incomplete real-data test."},{"cited_title":"Pointr: Diverse point cloud comple- tion with geometry-aware transformers","cited_arxiv_id":null,"evidence_quote":"PoinTr is the source of the online cropping procedure used to simulate incompleteness."},{"cited_title":"Score-based point cloud denoising","cited_arxiv_id":null,"evidence_quote":"ScoreDenoise is cited for the surface and outlier noise injection procedure used to build the noisy benchmarks."}],"review_version":2}