{"id":"7be2f6bc-e3a0-4295-9a91-e874f7f40149","arxiv_id":"2505.04119","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A geometry-aware prompting method reaches competitive 3D point cloud classification accuracy with roughly 2% of full fine-tuning parameters.","lead":"GAPrompt introduces a parameter-efficient way to adapt pre-trained 3D point cloud models using learnable point prompts and instance-specific shape features. It matches or beats full fine-tuning on classification benchmarks with about 2% of the trainable parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix A admits ReCon/Point-FEMAE rows use a stripped Point-MAE backbone, so 'surpasses full fine-tuning' may compare across architectures; this is the weakest link in the central claim.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing issue: the evaluation compares GAPrompt against full fine-tuning of ReCon and Point-FEMAE while using a stripped Point-MAE backbone. This is the correct point of focus because the strongest claim in the paper is that GAPrompt outperforms full fine-tuning, and the evidence for that claim relies on Table 1 rows where the full-fine-tuning baseline architecture differs from the GAPrompt architecture. The concern is explicit in the manuscript itself via Appendix A, so it cannot be dismissed as external speculation. The proposed test—full fine-tuning the same stripped backbone with the same pre-trained weights—would settle whether the 90.67 vs 90.22 margin is real. Secondary issues such as missing error bars and hyperparameter tuning on the evaluation benchmark are real but less decisive; they would weaken confidence in the exact numbers rather than directly test the central comparison. Because the reader already reached CONDITIONAL based on this and closely related concerns, no verdict adjustment is needed; the concern reinforces the conditional status rather than moving it to accept or reject.","tokens_in":15513,"tokens_out":8372,"duration_ms":85983,"concrete_test":"Run full fine-tuning on the exact stripped backbone GAPrompt uses: a Point-MAE model initialized with Point-FEMAE or ReCon pre-trained weights, residual components excluded, on ScanObjectNN PB T50 RS using the Table 5 hyperparameters and at least three seeds. If this stripped-backbone full fine-tuning reaches or exceeds 90.67% for Point-FEMAE weights or 89.76% for ReCon weights, the headline claim of surpassing full fine-tuning is not apples-to-apples; if it falls clearly below, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is comparison fairness for the headline claim that GAPrompt surpasses full fine-tuning. Appendix A states: \"we follow the approach of DAPT by only loading pre-trained weights into a Point-MAE model for efficient fine-tuning, while excluding the residual components of ReCon and Point-FEMAE.\" Therefore the Point-FEMAE and ReCon rows in Table 1 do not compare GAPrompt against full fine-tuning of those exact architectures; GAPrompt runs on a stripped Point-MAE backbone initialized with stronger pre-trained weights. The reported margins are small: 90.67 vs 90.22 on PB T50 RS for Point-FEMAE and 89.76 vs 89.73 for ReCon. These gaps are within typical seed-to-seed variation and may reflect the stripped architecture plus better pre-trained weights rather than the GAPrompt adaptation itself. The comparison against other PEFT methods (IDPT, DAPT, Point-PEFT) is fair because they all follow the same stripped-backbone protocol, so the core PEFT contribution could survive; however, the specific claim of surpassing full fine-tuning for ReCon and Point-FEMAE is not yet supported. The main text labels these rows as 'ReCon' and 'Point-FEMAE' without disclosing the stripped protocol in the caption, making the comparison easy to misread.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GAPrompt, a parameter-efficient fine-tuning (PEFT) method for pre-trained 3D point cloud transformers. Three components are proposed: a learnable Point Prompt appended to the raw input coordinates, a Point Shift Prompter that extracts instance-specific shape features and shifts input points, and a Prompt Propagation mechanism that injects the shape-enhanced prompt tokens into the token-wise FPS/KNN hierarchy of each transformer block. Experiments report classification accuracy on ScanObjectNN variants and ModelNet40 using four pre-trained backbones (Point-MAE, ReCon, PointGPT, Point-FEMAE). GAPrompt is reported to outperform existing PEFT methods (IDPT, DAPT, Point-PEFT) and in some cases to surpass full fine-tuning while using about 2% of trainable parameters. The paper includes component ablations, hyperparameter studies, and visualizations.","tokens_in":15728,"tokens_out":3561,"duration_ms":33828,"significance":"If the reported results are taken at face value, GAPrompt is a useful contribution to 3D PEFT: it explicitly exploits geometry at the point level, unlike token-only prompting, and it is shown to be parameter-efficient and computationally light. The paper provides careful component ablations (Tables 3 and 4), hyperparameter sensitivity plots (Figures 4, 7, 8), and a public code repository, which are strengths. However, the headline claim of surpassing full fine-tuning rests on a comparison whose fairness is compromised for two of the four backbones, as stated in Appendix A. The core PEFT comparison against IDPT/DAPT/Point-PEFT on the same protocol is meaningful, but the claim of beating full fine-tuning for ReCon and Point-FEMAE is not supported by the current evidence. The absence of variance reporting is an additional concern for the small reported margins.","major_comments":[{"comment":"The comparison against full fine-tuning of ReCon and Point-FEMAE is not apples-to-apples. Appendix A states that for these two models the authors 'only load pre-trained weights into a Point-MAE model for efficient fine-tuning, while excluding the residual components of ReCon and Point-FEMAE.' Thus the GAPrompt rows in Table 1 labeled 'ReCon' and 'Point-FEMAE' use a stripped Point-MAE backbone initialized with the pre-trained weights, whereas the 'Full Fine-Tuning' rows use the complete original architectures. The reported margins over full fine-tuning (0.03 points for ReCon on PB T50 RS, 0.45 points for Point-FEMAE on PB T50 RS) are therefore not attributable to the GAPrompt adaptation alone; they may reflect the backbone choice and the pre-training weights. The paper should either run GAPrompt on the full ReCon and Point-FEMAE architectures, or limit the claim to 'competitive with full fine-tuning' for these rows and make the protocol disclosure in the main text and table caption explicit.","section":"Appendix A / Table 1"},{"comment":"Table 1 reports a single accuracy per configuration without error bars or standard deviations. The central empirical claim of surpassing full fine-tuning relies on differences as small as 0.03 and 0.45 percentage points, which are typically within seed-to-seed variance for point cloud classification. Since the abstract and Section 4.2 explicitly state that GAPrompt 'surpasses' full fine-tuning for ReCon and Point-FEMAE, the authors should provide results over multiple seeds (at least for the key comparisons) or confidence intervals. Without this, the claimed advantage over full fine-tuning for these two backbones is not statistically established.","section":"Table 1 / Section 4.2"}],"minor_comments":[{"comment":"Section 4 states that the hyperparameters are set as βa = 0.5, βp = 0.5, and P = 20, but Table 5 shows different Point Prompt numbers for each dataset (20, 10, 20, 5). This is confusing; either the main text should acknowledge that P is tuned per dataset, or Table 5 should be reconciled with the stated default.","section":"Section 4 / Table 5"},{"comment":"The phrase 'sight computational saving' in Appendix A should be 'slight computational saving.'","section":"Appendix A"},{"comment":"The full fine-tuning row is labeled 'PointGPT-L' in Table 1 but the text in Section 4.2 and elsewhere uses 'Point-GPT'; please unify the naming.","section":"Table 1"},{"comment":"The table caption does not disclose that the ReCon and Point-FEMAE GAPrompt rows use a Point-MAE backbone with only the pre-trained weights loaded. Adding this caveat in the caption would prevent misreading of the results.","section":"Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's core PEFT contribution is plausible and the comparisons against IDPT/DAPT/Point-PEFT appear fair. The main risk is the overstatement of the full fine-tuning comparison for ReCon and Point-FEMAE due to the stripped-backbone protocol admitted in Appendix A. The authors should either run the actual architectures or soften the claim. Also, given the small margins, variance reporting is needed. The paper is within scope for CV and the experiments are extensive, so I see no reason to reject at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real step forward for 3D PEFT, and the main PEFT-vs-PEFT comparison is fair and favorable. Don't trust the 'surpasses full fine-tuning' line for the ReCon and Point-FEMAE rows without reading Appendix A.\n\nThe new idea is the point-level prompt instead of only token-level: a small set of learnable 3D points concatenated to the input, plus a Point Shift Prompter that predicts per-point displacements and extracts a global shape vector used to modulate prompt tokens and adapters. The Prompt Propagation module, which injects those shape-aware tokens into the FPS/KNN neighborhoods before attention, is genuinely different from what IDPT, DAPT, and Point-PEFT do. The paper shows consistently better accuracy than those three under the same protocol, with the fewest trainable parameters (0.6M) and only a tiny FLOP increase. Ablations cover each component, the scale factors, prompt count, initialization, and downstream head inputs. Code is public. That is a solid, useful contribution.\n\nThe weak spot is real but bounded. For ReCon and Point-FEMAE, the 'full fine-tuning' row is the complete model, while GAPrompt and the other PEFT rows run on a stripped Point-MAE backbone with only the pre-trained weights loaded—the residual modules in ReCon/Point-FEMAE are dropped. That is disclosed in Appendix A, but the main text and Table 1 give no hint in the caption, so the reader naturally reads 'surpasses full fine-tuning' as architecture-matched. The margins on PB T50 RS are 0.45 for Point-FEMAE and 0.03 for ReCon; those are within seed noise, and the comparison is not apples-to-apples. The claim should be reworded as 'competitive with full fine-tuning' for those two, or the comparison should be redone properly. This does not sink the paper, because the PEFT-vs-PEFT comparison is fair and GAPrompt wins consistently.\n\nOther, smaller issues: no error bars or multiple seeds anywhere, which matters for small gaps; and beta_a, beta_p, and P are tuned on PB T50 RS and then reported on it as the headline result. Both are standard oversights in the field but should be acknowledged.\n\nBottom line: the core method is sound, the efficiency numbers are real, and the novelty is clear. If this crosses your desk, send it to a serious referee—it deserves the ICML slot it got. The authors need to fix the framing of the full-fine-tuning comparison and add variance estimates.","headline":"Genuinely 3D-aware PEFT with a solid PEFT-vs-PEFT story; the 'surpasses full fine-tuning' claim for ReCon and Point-FEMAE is weakened by the stripped-backbone protocol disclosed only in Appendix A.","tokens_in":16327,"tokens_out":3456,"would_cite":true,"duration_ms":29470,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GAPrompt claims that geometry-aware prompting—trainable point clouds, instance-specific point shifts, and shape-feature propagation—lets a frozen 3D vision model beat full fine-tuning with 2.19% of the parameters.","keywords":["point cloud classification","parameter-efficient fine-tuning","prompt tuning","geometry-aware prompting","3D vision transformers","pre-trained 3D models","Point Shift Prompter","Prompt Propagation"],"falsifier":"Fully fine-tune the complete, unmodified ReCon and Point-FEMAE models (residual modules included) and run GAPrompt on those same complete backbones; if the 0.6M-parameter prompt no longer matches or exceeds full fine-tuning on ScanObjectNN PB T50 RS, the headline claim is refuted. A simpler check is to fully fine-tune the stripped Point-MAE-style backbones and see whether the 90.22% Point-FEMAE baseline itself shifts once its residual components are removed.","tokens_in":15267,"feed_emoji":"📐","tokens_out":7919,"duration_ms":65027,"temperature":0.7,"pith_summary":"GAPrompt is a parameter-efficient fine-tuning method for pre-trained 3D point cloud vision models. It freezes the backbone and trains only about 0.6M parameters, yet reports 90.67% accuracy on the hardest ScanObjectNN variant starting from Point-FEMAE, above the 90.22% of full fine-tuning. The central idea is that prompts for point clouds should carry geometry—a small learnable point cloud, instance-specific shifts of the input points, and a global shape feature propagated through the encoder—instead of only generic token prompts. If the claim holds, users can swap full fine-tuning for a tiny trainable prompt that preserves the pre-trained model and costs little extra compute.","feed_headline":"Geometry prompts let 0.6M parameters beat full fine-tuning","feed_subtitle":"Frozen 3D encoders learn instance shape from tiny learnable point prompts, reaching 90.67% on ScanObjectNN.","key_machinery":"The central object is the GAPrompt module trio. The Point Prompt is a small set of learnable 3D coordinates initialized uniformly and concatenated with the raw point cloud, giving the model explicit tunable geometry at the input; the Point Shift Prompter uses multi-resolution grouping with farthest point sampling and k-nearest neighbour search, a lightweight PointNet encoder, and a shift head to produce both a global shape feature $f$ and a shifted point cloud; the Prompt Propagation mechanism then runs FPS and KNN over the token set, randomly injects the shape-enhanced prompt tokens into the center and neighbour tokens, and propagates interpolated features back to all tokens. The shape feature $f$ also scales the prompt tokens and adapters through factors $\\beta_p$ and $\\beta_a$. This carries the argument by moving prompting from token space into point-level and feature-propagation space, which the paper argues is where point cloud geometry lives.","core_discovery":"The paper claims that a prompting scheme built around geometry, not just latent tokens, can adapt frozen pre-trained 3D vision models to downstream classification at a fraction of the parameter cost. With 0.6M trainable parameters (2.19% of Point-FEMAE), GAPrompt reaches 90.67% on ScanObjectNN PB T50 RS, surpassing full fine-tuning at 90.22%, and it improves on or matches full fine-tuning on Point-MAE, ReCon, and PointGPT-L as well. The authors trace this to three mechanisms: a Point Prompt that concatenates learnable 3D coordinates directly into the input point cloud, a Point Shift Prompter that extracts a global shape feature and produces per-point coordinate shifts, and a Prompt Propagation step that injects shape-enhanced prompt tokens into local feature interpolation inside each transformer block.","pith_inferences":["If geometry-aware prompting is the active ingredient, the same Point Shift and Prompt Propagation designs should transfer to segmentation and detection, where instance geometry is even more discriminative than in object classification.","A testable extension is to ablate the PointNet-style prompter against a simpler or fully symmetric shape encoder to separate the contribution of global shape extraction from the propagation mechanism itself.","The 'surpasses full fine-tuning' result for ReCon and Point-FEMAE rests on a stripped-backbone comparison; a direct run on the complete unmodified models would show whether the gain is due to the prompt or to dropping residual modules.","One could check whether the learned Point Prompt settles on semantically meaningful surface regions by measuring overlap with human keypoint annotations, which would directly test the claim that it captures fine-grained geometry."],"forward_implications":["With 0.6M trainable parameters, GAPrompt matches or exceeds full fine-tuning on four pre-trained backbones across ScanObjectNN and ModelNet40, so downstream users could freeze a 3D encoder and store only a small prompt per task.","The added compute is modest (5.0G FLOPs versus 4.8–5.3G for the base backbones), keeping inference close to the frozen-model cost, in contrast to IDPT (7.2G) and Point-PEFT (7.0G).","On PointGPT-L, GAPrompt reports 98.97% on ScanObjectNN OBJ BG and 96.2% on ModelNet40 without voting, which would make it the best reported efficient-tuning result on that backbone.","The same global shape feature feeds the downstream head together with the [CLS] token and max-pooled patch tokens, so instance geometry contributes directly to the final decision."],"supporting_citations":[{"why":"Supplies the Point-MAE pre-trained backbone, its token-embedding pipeline, and the full fine-tuning baseline (85.18 on PB T50 RS) that GAPrompt improves to 85.57.","marker":"Pang et al., 2022"},{"why":"Supplies Point-FEMAE, the pre-trained model used for the headline 90.67 vs 90.22 comparison on ScanObjectNN PB T50 RS.","marker":"Zha et al., 2024"},{"why":"Supplies ReCon, one of the pre-trained models GAPrompt adapts, including the 89.73 full fine-tuning baseline.","marker":"Qi et al., 2023"},{"why":"Supplies PointGPT-L, the largest backbone tested, where GAPrompt reports 98.97 on OBJ BG and 96.2 on ModelNet40.","marker":"Chen et al., 2024"},{"why":"Supplies IDPT, the prior instance-aware dynamic prompt tuning method whose 88.38 on PB T50 RS is the PEFT baseline GAPrompt surpasses.","marker":"Zha et al., 2023"},{"why":"Supplies DAPT, the dynamic adapter baseline (88.51) it surpasses, and the protocol of loading only pre-trained weights into a Point-MAE model that the authors adopt.","marker":"Zhou et al., 2024"},{"why":"Supplies Point-PEFT, the prompt+adapter+bias baseline (89.35) it surpasses, showing the prior state of the PEFT art.","marker":"Tang et al., 2024"},{"why":"Provides the FPS/KNN multi-resolution grouping and feature propagation operations reused by the Point Shift Prompter and Prompt Propagation.","marker":"Qi et al., 2017b"},{"why":"Provides the attention equivalence (linear interpolation plus offset subspace) used to justify the prompt-token design in Eq. 19.","marker":"He et al., 2021"}],"fun_headline_variants":["Geometry-aware prompts beat full fine-tuning with 2% parameters","0.6M parameter prompts outperform full fine-tuning on 3D","Point prompts squeeze 3D adaptation into 2% of parameters","GAPrompt: geometric cues unlock frozen 3D encoders","Frozen 3D models adapt via geometry prompts, not token tweaks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against full fine-tuning of ReCon and Point-FEMAE assumes that loading only their pre-trained weights into a Point-MAE-style backbone, without their extra residual modules, preserves the representational capacity those models would have if fully fine-tuned in their original architecture.","fun_headline_variants_meta":{"raw":{"variants":["Geometry-aware prompts beat full fine-tuning with 2% parameters","0.6M parameter prompts outperform full fine-tuning on 3D","Point prompts squeeze 3D adaptation into 2% of parameters","GAPrompt: geometric cues unlock frozen 3D encoders","Frozen 3D models adapt via geometry prompts, not token tweaks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000592,"raw_usage":{"total_tokens":2786,"prompt_tokens":966,"completion_tokens":1820,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":1727}},"tokens_in":582,"tokens_out":1820,"duration_ms":11801,"temperature":1.0,"reasoning_tokens":1727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:24.908338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fully fine-tune the complete, unmodified ReCon and Point-FEMAE models (residual modules included) and run GAPrompt on those same complete backbones; if the 0.6M-parameter prompt no longer matches or exceeds full fine-tuning on ScanObjectNN PB T50 RS, the headline claim is refuted. A simpler check is to fully fine-tune the stripped Point-MAE-style backbones and see whether the 90.22% Point-FEMAE baseline itself shifts once its residual components are removed.","supporting_citations":[{"cited_title":"E., Liu, W., Tian, Y., and Yuan, L","cited_arxiv_id":null,"evidence_quote":"Supplies the Point-MAE pre-trained backbone, its token-embedding pipeline, and the full fine-tuning baseline (85.18 on PB T50 RS) that GAPrompt improves to 85.57."},{"cited_title":"Pointgpt: Auto-regressively generative pre-training from point clouds","cited_arxiv_id":null,"evidence_quote":"Supplies PointGPT-L, the largest backbone tested, where GAPrompt reports 98.97 on OBJ BG and 96.2 on ModelNet40."},{"cited_title":"Instance-aware dynamic prompt tuning for pre-trained point cloud models","cited_arxiv_id":null,"evidence_quote":"Supplies IDPT, the prior instance-aware dynamic prompt tuning method whose 88.38 on PB T50 RS is the PEFT baseline GAPrompt surpasses."},{"cited_title":"Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis","cited_arxiv_id":null,"evidence_quote":"Supplies DAPT, the dynamic adapter baseline (88.51) it surpasses, and the protocol of loading only pre-trained weights into a Point-MAE model that the authors adopt."},{"cited_title":"Point-peft: Parameter-efficient fine-tuning for 3d pre-trained models","cited_arxiv_id":null,"evidence_quote":"Supplies Point-PEFT, the prompt+adapter+bias baseline (89.35) it surpasses, showing the prior state of the PEFT art."}],"review_version":1}