{"id":"02d8fb7c-7296-4f68-a646-cefa91ee2c29","arxiv_id":"2504.17996","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"LVTP prunes transformer tokens during semantic segmentation using multi-scale Tsallis entropy with Sobel edge guidance, reporting 20-46% FLOP reduction with a few mIoU points lost and no retraining.","lead":"This paper proposes a token pruning module for vision transformers that removes uninformative image patches during semantic segmentation, using a two-parameter entropy score plus edge detection to decide which tokens to keep. It reports 20 to 46 percent fewer computations with a small accuracy drop, without retraining the model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline GFLOPS and mIoU for SAM ViT-H RIO differ between Table 2 (820 GFLOPS, 0.53 mIoU) and Table 5 (528 GFLOPS, 46.5%), so the headline 46.2% GFLOPS reduction / 5.0 mIoU drop is not anchored to a stable reference.","rationale":"The reader's weakest assumption concerns the validity of using SAM, a class-agnostic interactive segmenter, to produce semantic-segmentation mIoU on COCO-Tasks and RIO. That is a serious evaluation concern, but the more decisive issue is the paper's own inconsistent reference configurations. The gamma metric in Eq. 11 is defined as a ratio of deltas, so without a fixed baseline the headline 'best gamma' claim is undefined. Table 2 and Table 5 provide two different baselines for the same model and dataset, and the paper offers no explanation for the discrepancy. The proposed test settles the ambiguity by measuring the truly unpruned model and the 0%-second-cluster model under identical settings. If Table 5's 0% row is not the unpruned model, the text mislabels it and the true baseline for Table 1 remains unspecified; if it is the unpruned model, the two tables contradict each other. In either case the central empirical claim cannot be checked without code or a clarified protocol. I therefore keep the reader's REJECT verdict; a future version with released code, explicit baseline definitions, and reported hyperparameters (q1, q2, cluster count, pruning layer) could merit a conditional acceptance.","tokens_in":14856,"tokens_out":10992,"duration_ms":111712,"concrete_test":"Run the exact SAM ViT-H image encoder on RIO with pruning fully disabled and record GFLOPS and mIoU; then run LVTP with a 0% second-cluster pruning rate at layer 16 under the same input resolution and record the same. If the disabled model gives 820 GFLOPS / 0.53 mIoU but the 0%-rate run gives 528 GFLOPS / 46.5%, then Table 5's 0% row is already an LVTP-pruned configuration and Table 1's deltas are computed from a different reference than the ablation table suggests. If the two runs match, Tables 2 and 5 are internally contradictory. Either outcome resolves whether the reported -46.2% GFLOPS and -5.0 mIoU are well-defined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the unspecified and inconsistent reference configuration behind the central claim. Table 1 reports deltas of -46.2% GFLOPS, -5.0 mIoU, and gamma = 9.24 on SAM ViT-H RIO. Table 2 gives the baseline as 820 GFLOPS / 0.53 mIoU and the pruned model as 441 GFLOPS / 0.48 mIoU. Table 5, for the same model and dataset, lists a '0% pruning rate' row with 528 GFLOPS / 46.5% mIoU and a 50% row with 441 GFLOPS / 48.2%. The 50% rows agree, but the reference points do not: 820 vs 528 GFLOPS and 0.53 vs 46.5% mIoU. The paper never states whether Table 5's 0% row is the untouched model or a model that has already undergone first-stage entropy clustering. If it is the latter, calling it '0% pruning rate' is misleading and the baseline used for gamma is unstated; if it is the former, the same configuration is measured twice with a 35% GFLOPS difference. Either way, the deltas and gamma values in Table 1 are not anchored to a well-specified reference, and the central claim of a SOTA cost-accuracy trade-off is not reproducible from the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LVTP, a training-free token pruning framework for ViT-based semantic segmentation. The method scores tokens with a multi-scale Tsallis entropy weighting, performs a first k-means clustering over entropy-weighted features, and then applies a second, Sobel-edge-guided clustering to preserve boundary-critical tokens before pruning at a selected transformer layer. The authors report GFLOPS reductions of roughly 20%-46% with small mIoU drops on RIO, COCO-Tasks, and Massachusetts Roads, and claim a superior trade-off (gamma) compared with CTS, DToP, SViT, and VLTP under a no-fine-tuning protocol. The paper also includes ablations over entropy choices, second-clustering variants, pruning rates, and pruning layers.","tokens_in":15181,"tokens_out":4345,"duration_ms":42979,"significance":"If the central claims were reliable, the paper would offer a practically attractive plug-in: a no-training, architecture-agnostic token pruning method that uses low-level image structure to retain segmentation boundaries. The motivation is reasonable, and the component-level ablations (Tables 3, 4, and 6) are a useful step toward understanding what drives the behavior. However, the current validation is not trustworthy because the baseline configuration is inconsistent across tables, the use of SAM for semantic segmentation mIoU is not specified, and the headline gamma is selected on the same benchmarks used for evaluation. These issues affect the paper's main quantitative claims rather than only its presentation, so the manuscript is not yet suitable for publication.","major_comments":[{"comment":"The reference configuration for the central claim is inconsistent. For SAM ViT-H on RIO, Table 2 lists a baseline of 820 GFLOPS and 0.53 mIoU, while Table 5 lists a '0% pruning rate' row with 528 GFLOPS and 46.5% mIoU. The 50% rows agree (441 GFLOPS; 0.48/48.2 mIoU), but the reference points do not. The paper does not state whether the 0% row is the untouched model or the model after first-stage entropy clustering. If it is the latter, calling it '0% pruning rate' is misleading; if it is the former, the same configuration is measured with a 35% GFLOPS difference. Either way, the reported -46.2% GFLOPS and gamma = 9.24 in Table 1 are not anchored to a well-specified baseline.","section":"Table 2 and Table 5"},{"comment":"SAM is a class-agnostic, promptable segmentation model, yet the paper reports semantic segmentation mIoU on COCO-Tasks and RIO without describing how class-level semantic masks are derived. There is no discussion of prompts, class mapping, or any task adaptation. Unless this protocol is specified, the mIoU numbers do not measure semantic segmentation as understood in the comparison methods (CTS, DToP, SViT), and the comparison in Table 1 is not valid.","section":"Section 4.1 and Table 2"},{"comment":"The metric gamma is defined as the ratio of the reduction ratio in GFLOPS to the percentage decrease in mIoU, but the units of mIoU are used inconsistently. Table 2 reports mIoU as proportions (0.53, 0.48), while Table 5 reports percentages (46.5%, 48.2%). The claimed -5% mIoU change from 0.53 to 0.48 is an absolute 5-percentage-point drop, not a relative 9.4% decrease. The gamma values therefore depend on an unstated and non-uniform interpretation of Delta mIoU, making them incomparable across tables.","section":"Eq. (11) and Tables 1, 2, 5"},{"comment":"The headline gamma is computed after selecting the pruning layer (16th layer) and the second-clustering pruning rate (50%) from a scan over the same benchmarks in Tables 5 and 6. This is a selection-on-the-test-set procedure, so the reported superiority of LVTP in Table 1 partly reflects the choice of the best configuration rather than an independent evaluation. A held-out validation split or a fixed default configuration should be used to report the main result.","section":"Tables 5 and 6 vs. Table 1"}],"minor_comments":[{"comment":"The values of q1 and q2, the number of clusters k, and the initialization of cluster centers are not reported, even though these are central free parameters of the method; a sensitivity study or at least the chosen values is needed for reproducibility.","section":"Section 3.2-3.3"},{"comment":"Figure 5 is captioned 'Framework for Edge-Enhanced Token Clustering in Transformers', which is the same title as Figure 1; the caption should describe the gamma-versus-pruning-rate visualization instead.","section":"Figure 5 caption"},{"comment":"In the text before Table 4, the RIO Sobel-guided result is described as a GFLOPS reduction of '-48.2%', but the table lists '-46.2%'; the numbers should be reconciled.","section":"Section 4.4.2, Table 4"},{"comment":"The paper should state explicitly whether the '0%' row in Table 5 already includes the first-stage entropy-based clustering; without this clarification the table cannot be interpreted as a pruning-rate ablation.","section":"Section 4.4.3, Table 5"},{"comment":"There are numerous typographical and language errors ('focued', 'becames', 'di fferent', inconsistent article usage) that should be corrected in a revision.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The central quantitative claim is not reproducible from the manuscript because the baseline is inconsistent and the evaluation protocol for SAM is unspecified; these are load-bearing issues, not presentation problems. The paper may have a salvageable idea, but as submitted the experiments do not support the stated conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead arXiv:2504.17996. Short version: the method is a plausible and genuinely new combination — two-parameter Tsallis entropy weighting with a dynamic alpha/beta, then a Sobel-edge-guided second clustering round — and the ablation layout is more transparent than most. But the evaluation as written cannot support the headline claims. The reference configuration shifts between tables, and the SAM mIoU protocol is never specified. The stress-test numbers check out: Table 2 gives SAM ViT-H on RIO as 820 GFLOPS / 0.53 mIoU, while Table 5's 0-percent row gives 528 GFLOPS / 46.5 percent for the same claimed setting.\n\nWhat's new is the specific assembly: I don't see the q1/q2 Tsallis sum with norm-based weights plus Sobel-guided reclustering in the cited literature. The ablations cover the right dimensions — entropy guidance (none/Shannon/M-Tsallis), second clustering (none/simple/Sobel), a pruning-rate sweep, and a layer sweep — and they report the full scan, not just the winning point. The pruned-state numbers reproduce across tables (441 GFLOPS and roughly 48 mIoU appear in Tables 2, 3, and 5), so the inference configuration is internally consistent. The idea of protecting edge tokens is sensible for segmentation, and the Discussion does acknowledge robustness limits under extreme lighting, occlusion, and noise — though it never flags the baseline or protocol problems, which are the real trouble.\n\nSoft spots, in proportion. Load-bearing: Table 1's deltas and gamma (46.2 / 5.0 / 9.24) are not anchored to a stable baseline. If Table 5's 0-percent row is the model after first-stage clustering, calling it zero-percent pruning is misleading, and the first clustering alone appears to cost about 6.5 mIoU points that the paper never discusses. If it is the untouched model, the same configuration is measured twice with a 35-percent GFLOPS difference. Either way, the headline number is not reproducible from the manuscript. Second, the SAM mIoU protocol is missing: SAM is class-agnostic, and the paper gives no account of how masks are labeled or matched to compute semantic mIoU on RIO and COCO-Tasks; without that, the comparison against CTS/DToP/VLTP may not measure the same task. Third, Table 1 uses the best layer (16) and rate (50 percent) selected from Tables 5-6, with no evidence the baselines got the same tuning; the selection is visible, so this is moderate, not fatal. Fourth, one middle layer's pruning is said to cut 46 percent of global GFLOPS, but the paper never says whether tokens are physically dropped or how the decoder handles pruned positions. Minor: the Tsallis citation [55] points to a k-means paper; RIO is misdescribed as 3D spatial when the cited RIO is a 2D object-reasoning benchmark; 'multi-scale' oversells a two-parameter sum; figure captions are duplicated; no code.\n\nThis is for people working on efficient ViT inference and token pruning for dense prediction. The method is a reasonable incremental idea, and the ablation design is a useful template for mapping a pruning scheme's sensitivity. The numbers, as published, are not trustworthy.\n\nMy recommendation: it deserves peer review, not desk rejection — the idea is new enough, and the failures sit in the evaluation and reporting rather than in the concept. A serious referee could put it on a major-revision path: stabilize and specify the baseline, document the SAM mIoU protocol, and release code. In the current form I would not accept it.","headline":"A plausible new token-pruning recipe with transparent ablations, but the reference baseline shifts between tables and the SAM mIoU protocol is never specified, so the headline trade-off numbers are not yet reproducible.","tokens_in":15766,"tokens_out":20444,"would_cite":false,"duration_ms":173652,"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":"A two-stage entropy-and-edge score prunes ViT tokens with no training, cutting compute by up to 46% while holding segmentation accuracy.","keywords":["Vision Transformers","token pruning","Tsallis entropy","semantic segmentation","Sobel edge detection","plug-and-play compression","model efficiency","edge-aware pruning"],"falsifier":"Re-run the same no-training protocol on a class-labeled semantic segmentation benchmark such as ADE20K or Cityscapes with a ViT segmentor whose task head is fixed; if a 45% GFLOPS reduction comes with an mIoU drop larger than the claimed 3 points on COCO-style data, or if removing the Sobel-guided second clustering does not increase mIoU relative to a simple second clustering, the paper's central claim fails.","tokens_in":14582,"feed_emoji":"✂️","tokens_out":11752,"duration_ms":97789,"temperature":0.7,"pith_summary":"The paper introduces LVTP, a training-free token pruning method for Vision Transformers used in semantic segmentation. It claims that scoring tokens with multi-scale Tsallis entropy and then re-checking survivors with Sobel edge maps preserves segmentation accuracy far better than methods that rely on high-level semantics alone. On a frozen SAM (Segment Anything) ViT-H it reports a 46.2% GFLOPS reduction with a 5 point mIoU drop on RIO, and a 45.5% GFLOPS reduction with a 3 point mIoU drop on COCO-Tasks. If these numbers hold, LVTP offers a plug-and-play route to cheaper ViT-based segmentation on resource-constrained devices without retraining or architectural change.","feed_headline":"Edge-guided token pruning halves ViT compute with tiny accuracy loss","feed_subtitle":"No training or fine-tuning needed: multi-scale entropy plus Sobel edges drops redundant tokens while keeping boundaries sharp.","key_machinery":"The load-bearing object is a two-scale Tsallis entropy score, a non-extensive generalization of Shannon entropy, written as $S_{q_1,q_2}(P)=\\alpha\\cdot(1-\\sum_i p_i^{q_1})/(q_1-1)+\\beta\\cdot(1-\\sum_i p_i^{q_2})/(q_2-1)$, with $q_1>1$ and $q_2<1$, and dynamic weights $\\alpha=\\|F\\|_2/\\sqrt{D}$, $\\beta=1-\\alpha$. This entropy weight matrix modulates token features before cosine-distance k-means clustering, and a second clustering is driven by the Sobel gradient magnitude $G(x,y)=G_x(x,y)^2+G_y(x,y)^2$. The machinery classifies tokens as redundant or informative at two complementary granularities: entropy catches semantic richness, while Sobel catches boundary and texture cues that high-level scoring would prune away.","core_discovery":"The paper's central claim is that token redundancy in ViT-based semantic segmentation can be removed in two stages. The first stage clusters tokens with an entropy-weighted feature representation using multi-scale Tsallis entropy, which scores semantic importance while balancing salient high-probability features ($q_1>1$) and detail-oriented low-probability features ($q_2<1$). The second stage refines that decision with Sobel edge maps so boundary tokens survive pruning. Across SAM ViT-H/ViT-L and Swin-Unet backbones on RIO, COCO-Tasks, and Massachusetts Road, this progressive scheme reduces GFLOPS by about 20% to 46% while keeping mIoU degradation between roughly 0 and 6 points, and it outperforms CTS, DToP, SViT, and VLTP on the gamma tradeoff metric, defined as the ratio of GFLOPS reduction to mIoU decrease, under a no-fine-tuning comparison.","pith_inferences":["The Sobel-guided second stage only needs the input image, so the same two-stage scoring should transfer to other dense prediction tasks such as instance or panoptic segmentation; the paper lists this as future work, not a demonstrated result.","The RIO finding that accuracy improves as the second-stage pruning rate rises to 40% suggests entropy-guided pruning also acts as a mild regularizer by suppressing noisy background tokens, an effect the paper does not claim explicitly.","A testable extension is to replace the Sobel operator with learned boundary maps or structured edge detectors to see whether the extreme conditions named in the discussion, such as lighting changes, occlusion, and noise, can be handled without retraining the segmentor.","The gamma ratio combines a percentage GFLOPS change with a percentage-point mIoU change, so cross-method comparisons should be read alongside the absolute mIoU and GFLOPS numbers, not the ratio alone."],"forward_implications":["LVTP is plug-and-play: it applies to frozen ViT, Swin, and SAM backbones with no architectural change and no fine-tuning.","Pruning at a middle transformer layer (layer 16 of SAM ViT-H) gives the best cost-accuracy tradeoff; pruning layer 8 destroys primitive features and pruning layer 24 disrupts global semantics.","The Sobel-guided second clustering is what preserves boundary accuracy: removing it raises the mIoU drop from 5.0 to 11.9 points on RIO.","Moderate pruning can even improve accuracy on RIO, where mIoU rises from 46.5% at 0% second-stage pruning to 48.5% at 40%.","LVTP has the best GFLOPS-to-mIoU tradeoff among the training-free token pruning methods compared in Table 1."],"supporting_citations":[{"why":"Supplies the frozen SAM ViT-H and ViT-L backbones used in the main pruning experiments.","marker":"[62]"},{"why":"Supplies VLTP, the vision-language guided token pruning baseline compared in Table 1 and the motivation for the low-level feature gap.","marker":"[24]"},{"why":"Supplies CTS, the content-aware token sharing baseline compared in Table 1.","marker":"[63]"},{"why":"Supplies DToP, the dynamic token pruning baseline compared in Table 1.","marker":"[64]"},{"why":"Supplies SViT, the token pruning baseline compared in Table 1.","marker":"[65]"},{"why":"Supplies the Tsallis entropy formulation that the paper extends to multi-scale form with q1 and q2.","marker":"[56]"},{"why":"Supplies the Sobel operator used for low-level edge guidance in the second clustering.","marker":"[57]"},{"why":"Supplies the Massachusetts Road dataset used in the Swin-Unet generalization experiments.","marker":"[58]"},{"why":"Supplies the COCO-Tasks dataset used for multi-class dense target evaluation.","marker":"[59]"},{"why":"Supplies the RIO dataset used for indoor scene understanding evaluation.","marker":"[60]"}],"fun_headline_variants":["Two-stage token pruning cuts ViT compute by 45% with no retraining","Edge-aware entropy pruning drops 45% of ViT tokens, keeps accuracy","Multi-scale entropy + Sobel edges slash ViT compute without finetuning","ViT token pruning with edge awareness: 20-46% less GFLOPS, tiny mIoU drop","Entropy+edges: 45% less ViT compute, no retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central measurements assume SAM's outputs can be scored as class-level semantic segmentation mIoU on the RIO and COCO datasets without task adaptation, even though SAM is an interactive, class-agnostic segmenter; if its masks are not class-labeled semantic predictions, the reported mIoU drops do not measure semantic segmentation quality.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage token pruning cuts ViT compute by 45% with no retraining","Edge-aware entropy pruning drops 45% of ViT tokens, keeps accuracy","Multi-scale entropy + Sobel edges slash ViT compute without finetuning","ViT token pruning with edge awareness: 20-46% less GFLOPS, tiny mIoU drop","Entropy+edges: 45% less ViT compute, no retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000984,"raw_usage":{"total_tokens":4143,"prompt_tokens":879,"completion_tokens":3264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":3153}},"tokens_in":495,"tokens_out":3264,"duration_ms":22234,"temperature":1.0,"reasoning_tokens":3153,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:27:34.491634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same no-training protocol on a class-labeled semantic segmentation benchmark such as ADE20K or Cityscapes with a ViT segmentor whose task head is fixed; if a 45% GFLOPS reduction comes with an mIoU drop larger than the claimed 3 points on COCO-style data, or if removing the Sobel-guided second clustering does not increase mIoU relative to a simple second clustering, the paper's central claim fails.","supporting_citations":[{"cited_title":"Revisiting token pruning 19 for object detection and instance segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies SViT, the token pruning baseline compared in Table 1."},{"cited_title":"Segment anything","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen SAM ViT-H and ViT-L backbones used in the main pruning experiments."},{"cited_title":"Content-aware token sharing for e fficient semantic segmen- tation with vision transformers","cited_arxiv_id":null,"evidence_quote":"Supplies CTS, the content-aware token sharing baseline compared in Table 1."},{"cited_title":"Dynamic token pruning in plain vision transform- ers for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies DToP, the dynamic token pruning baseline compared in Table 1."},{"cited_title":"Multilevel image thresholding based on 2d histogram and maximum tsallis entropy—a differential evolution approach","cited_arxiv_id":null,"evidence_quote":"Supplies the Tsallis entropy formulation that the paper extends to multi-scale form with q1 and q2."},{"cited_title":"Edge detection using guided sobel image filtering","cited_arxiv_id":null,"evidence_quote":"Supplies the Sobel operator used for low-level edge guidance in the second clustering."},{"cited_title":"Machine learning for aerial image labeling","cited_arxiv_id":null,"evidence_quote":"Supplies the Massachusetts Road dataset used in the Swin-Unet generalization experiments."},{"cited_title":"What object should i use?-task driven object detection","cited_arxiv_id":null,"evidence_quote":"Supplies the COCO-Tasks dataset used for multi-class dense target evaluation."},{"cited_title":"Rio: A benchmark for reasoning intention-oriented objects in open environments","cited_arxiv_id":null,"evidence_quote":"Supplies the RIO dataset used for indoor scene understanding evaluation."}],"review_version":1}