{"id":"f85236e4-3634-457e-b1f0-82c856c4a7cc","arxiv_id":"2607.19711","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PSFT couples minimally-influential point selection with prompt tuning and stochastic feature filtering, cutting corruption error on ModelNet-C and ModelNet40-C across four pre-trained 3D backbones.","lead":"This paper combines point selection, layer-wise prompts, and a lightweight feature filter to keep 3D point cloud classifiers robust when fine-tuning pre-trained models. It reports lower corruption error on ModelNet-C and ModelNet40-C across all four tested backbones and on two of four backbones on ScanObjectNN-C.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection module's core premise—that highly influential points are corrupted outliers—is untested and may fail for strong backbones, so the mechanism behind PSFT's robustness gains is unverified.","rationale":"The reader's weakest_assumption is precisely the selection premise, and the paper itself flags the ScanObjectNN-C mixed pattern, which strengthens the concern. This is load-bearing because the point selection module is the central novelty distinguishing PSFT from ordinary PEFT: if the selection is not removing outliers, the method's robustness gains might come from the feature filter or prompt tuning alone, or from an implicit regularization effect, rather than from the claimed influence-based suppression. The paper only tests the premise indirectly through end-to-end benchmarks, which cannot isolate whether the selection mechanism behaves as stated. A direct point-level analysis on the corrupted points is the cleanest way to settle this. I do not think this concern, if it lands, would fully overturn the empirical results—PSFT may still be a useful robust fine-tuning recipe—but it would change the paper's explanatory claims and the interpretation of its novelty. Therefore the existing CONDITIONAL verdict is appropriate; no verdict change is needed.","tokens_in":16716,"tokens_out":17934,"duration_ms":199575,"concrete_test":"Using the released code, run the PSFT forward pass on ModelNet-C with ULIP-2 and record which points are dropped by the selection module for each corruption type. For Add-G and Add-L, the identity of the added corrupted points is known; for Jitter, mark points with displacement above a high percentile as corrupted. Compute recall@k and precision@k of the dropped-point set against this ground-truth corrupted set. If recall@k is low (e.g., <0.3 for Add-G/Add-L), the selection module is not preferentially removing outliers, and the premise underlying the method's mechanism fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The point selection module (Sec. 3.1) discards the most influential points under max-pooling, citing Refocusing [11]'s observation that highly influential points are likely corrupted outliers. This premise is the entire basis for the PS component, and Table 5 shows PS alone accounts for a large share of the improvement (ModelNet-C mCE 0.863→0.656 for ULIP-2). However, the paper never directly verifies that the dropped points are actually the corrupted points in ModelNet-C/ModelNet40-C/ScanObjectNN-C, nor that the premise transfers to modern transformer backbones such as ULIP-2 and Uni3D-B. For these multimodal backbones, the most influential points may be semantically salient clean structure rather than noise. The ScanObjectNN-C results are consistent with this worry: PSFT *hurts* Point-BERT and Point-MAE (Table 3), suggesting selection can remove informative points under realistic corruptions. If the selection is not actually targeting outliers, the reported gains may be attributable to the FFM/prompt modules or to a regularization effect, and the paper's central explanatory claim—that influence-based point selection suppresses corruption—would be unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PSFT, a three-module robust fine-tuning framework for 3D point cloud classification: (i) a point selection module that drops points with high max-pooling influence, following the refocusing idea of [11]; (ii) a prompt generation branch that creates layer-wise prompts from the selected subset for a frozen backbone; and (iii) a feature filter module with Beta-gated residual blending to refine patch tokens. Experiments on ModelNet-C, ModelNet40-C, and ScanObjectNN-C across Point-BERT, Point-MAE, ULIP-2, and Uni3D-B report consistent mCE/mER improvements on the two synthetic corruption benchmarks and strong ScanObjectNN-C results for ULIP-2 and Uni3D-B, while using about 10.9% of the trainable parameters of full fine-tuning. The paper also provides ablations, sensitivity analyses, and public code.","tokens_in":17056,"tokens_out":12873,"duration_ms":130224,"significance":"If the empirical claims are reliable, PSFT is a useful contribution: it offers a parameter-efficient alternative to full fine-tuning that improves or at least preserves corruption robustness on common benchmarks. Strengths include the public code, systematic module ablations, evaluation across four substantially different backbones, and honest reporting of the mixed ScanObjectNN-C behavior. The main limitations are methodological: the selection mechanism's premise is not directly verified, test-set information is used to choose hyperparameters, and the reported aggregate gains come from single seeds without variance estimates. These issues currently limit confidence in the central explanatory claim, although they are addressable in revision.","major_comments":[{"comment":"The influence score I_F(j) is computed from X_f, the final pre-pooling feature map, which requires a complete forward pass through the frozen backbone with all N points. The subsequent prompt-injected forward pass in Eq. (6) uses the selected subset, implying a second forward pass. The manuscript never states this explicitly, nor does it report the added inference cost. If the second pass indeed runs, the influence scores are stale with respect to the selected subset; if there is no second pass, then selection only affects the prompt/FFM branches and cannot 'suppress outliers' in the backbone's pooled representation as claimed. Please specify the exact data flow, including whether the patch encoder and transformer layers operate on the selected subset, and give the computational overhead.","section":"Sec. 3.1 (Eq. 1) and Sec. 3.2 (Eq. 6)"},{"comment":"The central premise—that the most influential points are corrupted outliers—is adopted from [11] and is not tested on any corruption type or backbone used here. Table 5 shows that point selection alone accounts for a large part of the ULIP-2 ModelNet-C gain (mCE from 0.863 to 0.656), so this premise is load-bearing. Table 3 shows that on ScanObjectNN-C the same selection degrades clean OA and mCE for Point-BERT and Point-MAE, which is consistent with the possibility that informative structure is being discarded. Please provide a direct test, e.g., per-corruption overlap between the discarded points and known corrupted points, or a comparison of influence-based selection with random and inverse-influence selection, reported per backbone. This would substantiate or correct the explanatory claim.","section":"Sec. 3.1 / Table 5"},{"comment":"Test-time α is selected by reading mCE on ModelNet-C, and λ=0.5 is chosen as the best training-time Beta shape on the same benchmark. Since Fig. 4 shows mCE varies substantially with α (roughly 0.53–0.66), this use of the evaluation benchmark to set hyperparameters can inflate the reported improvements over baselines whose own hyperparameters are not similarly tuned. Please fix α=0.5 a priori (it is the Beta mean and a natural default) or select α and λ on a separate validation split, and report the corresponding test results. Also state which α/λ values were used to produce Tables 1–3.","section":"Sec. 4.3 / Fig. 4"},{"comment":"All results appear to be single runs with no error bars or multi-seed statistics. The claimed 'consistent' improvements are aggregate means over corruption types, and some per-corruption cells move in the opposite direction (e.g., Table 1 Rotate for Point-BERT: 1.316→1.591; Table 2 Rotation/Shear for Point-BERT and Point-MAE). A single seed is insufficient to establish the consistency claim. Please report mean±std over at least three seeds, or provide a statistical test for the headline mCE/mER differences.","section":"Tables 1–3"}],"minor_comments":[{"comment":"Typos: 'Feature Fiter Module' and 'Feature Fiter' should be 'Feature Filter Module' and 'Feature Filter'.","section":"Fig. 2 / Fig. 3 captions"},{"comment":"The baseline row (e.g., 'Point-BERT') is implicitly full fine-tuning, but this is not stated in the captions. Since the comparison to full fine-tuning is central to the parameter-efficiency claim, make the baseline explicit.","section":"Tables 1–3"},{"comment":"Per-corruption tradeoffs are not discussed, e.g., PSFT improves Add-G but hurts Jitter for ULIP-2. A sentence acknowledging that mCE gains are not uniform across corruptions would improve transparency.","section":"Sec. 4.2, ScanObjectNN-C paragraph"},{"comment":"On ScanObjectNN-C, the full model (PS+PG+FFM) has mCE 0.685, which is worse than PS+FFM (0.654) and lower OA (86.1 vs 87.4). The text notes this, but the choice of the full model over PS+FFM as the reported method needs a justification, e.g., average performance across all three benchmarks.","section":"Table 5"},{"comment":"Implementation details are sparse. Please provide batch size, learning rate schedule, number of input points, patch size (for transformer backbones), bottleneck dimension d', and EdgeConv hyperparameters to allow reproduction.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope. The two-pass ambiguity and the unverified selection premise are the main risks to the paper's central mechanism, but both are addressable with a clarifying revision and an additional diagnostic. The test-set tuning of α and λ is a protocol issue that should be corrected. I would be willing to review a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent and useful empirical paper. The genuinely new thing is the coupling: influence-based point selection (borrowed from Refocusing), frozen-backbone prompt tuning, and a bottleneck-MLP feature filter with Beta-sampled residual blending. The paper shows consistent mCE/mER gains over full fine-tuning and IDPT on ModelNet-C and ModelNet40-C across four pre-trained backbones, and it is honest about the mixed ScanObjectNN-C results — PSFT helps ULIP-2 and Uni3D-B there but hurts Point-BERT and Point-MAE. The ablation table showing that point selection alone accounts for a large part of the gain is a real contribution, and the parameter-efficiency numbers are clear. Code is linked, though not pinned to a commit.\n\nThe soft spots are real but not fatal. There are no error bars or multi-seed runs; some of the differences in Tables 1–3 could be run-to-run noise. Alpha and lambda are selected using the same corruption benchmarks on which the method is evaluated (Fig. 4), so the headline numbers partly reflect test-set tuning. And the central mechanism — that minimally influential points are cleaner — is inherited from [11] but never directly validated here. Point selection alone drops ModelNet-C mCE from 0.863 to 0.656 for ULIP-2; that is a large single-module effect, and without showing that the dropped points are actually the corrupted ones, the gain could also come from a generic regularization effect of dropping high-activation points. The ScanObjectNN-C backbone dependence suggests the selection is not universally safe. That said, the stress-test worry that selection fails specifically for strong multimodal backbones is not supported: on ScanObjectNN-C, the strong backbones are exactly where it helps. The concern should be reframed as \"when does the premise hold?\", not \"strong backbones break it.\"\n\nThe math is clean: the selection rule is a definition, there is no circularity, and the Beta-gating formula is simple and clearly explained. The citation pattern looks reasonable; Refocusing, VPT/IDPT, and the benchmarks are properly credited.\n\nBottom line: this deserves a serious referee. I would send it out. For acceptance I would want multi-seed statistics, a direct check of whether selected points correlate with corruption, a Refocusing comparison, and a pinned code commit. For someone working on 3D robustness or PEFT, it is worth reading, and I would cite it.","headline":"A useful, honestly reported robustness recipe for 3D PEFT, but the point-selection mechanism is under-verified and the stats need tightening.","tokens_in":17479,"tokens_out":3017,"would_cite":true,"duration_ms":29453,"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":"PSFT, a point-selection fine-tuning framework, claims that retaining minimally influential points, prompting a frozen backbone, and Beta-gated feature filtering reduces corruption error on ModelNet-C and ModelNet40-C across four pre-trained","keywords":["point cloud classification","corruption robustness","parameter-efficient fine-tuning","prompt tuning","point selection","feature filtering","frozen backbone","3D pre-trained models"],"falsifier":"On ModelNet-C with additive-Gaussian noise, measure the retained ratio S/N under increasing noise severity; if S/N stays near 1 (i.e., the entropy rule does not respond to corruption) while PSFT still lowers mCE relative to full fine-tuning, then the robustness gain comes from the prompt/filter modules rather than point selection, and the selection premise is not the active mechanism.","tokens_in":16645,"feed_emoji":"🧊","tokens_out":7445,"duration_ms":75162,"temperature":0.7,"pith_summary":"The paper argues that full fine-tuning of 3D pre-trained models can harm corruption robustness by overwriting robust priors, while naive parameter-efficient tuning remains sensitive to corrupted points. It proposes PSFT, which ranks points by their influence on the pooled representation, discards the most influential ones (assumed to be outliers), adapts a frozen transformer via layer-wise prompts generated from the selected subset, and refines patch tokens with a bottleneck MLP blended through a Beta-gated residual connection. The central claim is that this coupling yields lower corruption error than full fine-tuning on ModelNet-C and ModelNet40-C for all four tested backbones (Point-BERT, Point-MAE, ULIP-2, Uni3D-B), and the best ScanObjectNN-C result among compared tuning strategies for the two strongest backbones, while using only 10.9% of full fine-tuning's trainable parameters. A sympathetic reader would care because robust, parameter-efficient adaptation could make large 3D foundation models practical in noisy perception settings, and because the paper's own ScanObjectNN-C numbers show the benefit is not universal: for Point-BERT and Point-MAE, PSFT worsens mCE relative to full fine-tuning.","feed_headline":"Point-selection tuning beats full fine-tuning on corrupt 3D data","feed_subtitle":"PSFT keeps the backbone frozen, prunes influential outliers, and filters residual noise, cutting corruption error on ModelNet-C and ModelNet","key_machinery":"The load-bearing object is the point-wise influence score I_F(j) = Σ_k 1{ j = argmax_n X_f(n,k) }, counted over the pre-pooling feature map's channels, together with the entropy-guided retention rule S = round((1 - H_n(p))·N) that keeps the minimally influential points. The prompt generation branch, an EdgeConv-based network, turns the selected subset into per-layer prompt tokens injected as (T_cls, P_i, T_i) into each frozen transformer layer. The feature filter module is a bottleneck MLP (down-projection, GELU, dropout, up-projection) whose output is blended with the original token by x_final = (1-α)x_f + αx_o, with α ~ Beta(λ,λ) during training — a bounded, stochastic residual mixing that","core_discovery":"PSFT's central discovery is that robustness during adaptation can be improved by not fine-tuning the backbone at all, instead feeding a frozen pre-trained model a deliberately pruned point cloud. The pruning is driven by a per-point influence score: for each feature channel before global pooling, the point with the maximum activation gets a vote, and the votes are aggregated; the entropy of the normalized vote distribution sets the retention ratio S/N = 1 - H_n, and the least influential S points are kept. On top of this, an EdgeConv-based branch predicts one learnable prompt token per transformer layer from the selected points, and a funnel MLP with Beta-sampled residual blending (x_final =","pith_inferences":["If the refocusing premise is right, PSFT-style influence-based pruning could be used as a plug-in input filter for other PEFT methods (adapters, LoRA), since it is computed from the frozen backbone's own features and requires no extra supervision.","The entropy rule S = (1-H_n)N means that under corruptions that spread influence evenly (e.g., rotation, shear), selection retains nearly all points; the paper's per-corruption tables show PSFT improves even those categories, so either the prompt branch or the feature filter — not point pruning — is doing the work there. A per-corruption ablation would separate these contributions.","The paper's own ScanObjectNN-C results (Point-BERT mCE 0.931→0.970, Point-MAE 0.895→1.014) and the supplementary's finding that WOLFMix raises ScanObjectNN-C mCE by 0.204 on average suggest a dataset–augmentation mismatch: on cluttered real scans, dropping minimally influential points can remove useful geometry. A testable fix would be density-aware or class-conditional selection.","The Beta-gated blending with optimal test-time α≈0.5 suggests that the filter's value is its bounded correction of the frozen feature, not the stochasticity per se; fixing α=0.5 during training and testing would isolate the interpolation from the regularization effect."],"forward_implications":["PSFT lowers mean corruption error on ModelNet-C for all four evaluated backbones and on ModelNet40-C lowers mean error rate for all four, with the largest gains on strong backbones (ULIP-2 mCE 0.530, Uni3D-B mER 14.2).","Robustness gains are compatible with data augmentation: adding WOLFMix further improves synthetic-benchmark results (best ModelNet-C mCE 0.465 with Uni3D-B), establishing complementarity.","Parameter efficiency is preserved: 2.4M trainable parameters (10.9% of full fine-tuning) suffice to beat full fine-tuning on synthetic corruptions.","Ablations attribute gains to all three modules, with selection alone reducing mCE substantially and selection plus feature filtering being the strongest pair on ScanObjectNN-C.","The transfer to real-world corruptions is backbone-dependent: on ScanObjectNN-C, PSFT beats full fine-tuning only for ULIP-2 and Uni3D-B, and degrades mCE for Point-BERT and Point-MAE."],"fun_headline_variants":["Frozen backbone with pruned points cuts 3D corruption error","Influence-based point pruning beats full fine-tuning","Keep backbone frozen, prune outliers for robust 3D","Point selection beats fine-tuning on corrupt point clouds","PSFT: prune points, prompt frozen net for 3D robustness"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise, taken from prior refocusing work and adopted in Sec. 3.1, is that the points winning the most max-pooling channels are usually corrupted outliers, so that discarding the most influential points removes noise instead of informative structure; the paper tests this only indirectly through end-to-end benchmarks, and its ScanObjectNN-C results for Point-BERT and Point-MAE show the premise can fail on real-world scans.","fun_headline_variants_meta":{"raw":{"variants":["Frozen backbone with pruned points cuts 3D corruption error","Influence-based point pruning beats full fine-tuning","Keep backbone frozen, prune outliers for robust 3D","Point selection beats fine-tuning on corrupt point clouds","PSFT: prune points, prompt frozen net for 3D robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1488,"prompt_tokens":752,"completion_tokens":736,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":654}},"tokens_in":496,"tokens_out":736,"duration_ms":6981,"temperature":1.0,"reasoning_tokens":654,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:54:14.788445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On ModelNet-C with additive-Gaussian noise, measure the retained ratio S/N under increasing noise severity; if S/N stays near 1 (i.e., the entropy rule does not respond to corruption) while PSFT still lowers mCE relative to full fine-tuning, then the robustness gain comes from the prompt/filter modules rather than point selection, and the selection premise is not the active mechanism.","supporting_citations":[],"review_version":1}