{"id":"a5adc00b-f651-47af-bdec-da5f1d666743","arxiv_id":"2608.13226","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CoverPrune and its Lite variant prune 3D VLM visual tokens by optimizing feature-spatial-temporal optimal transport coverage, outperforming diversity-based pruning at 5-20% token retention.","lead":"This paper introduces CoverPrune, a training-free method that prunes 3D vision-language model tokens by selecting a small set that best covers the original tokens under an optimal-transport cost. A reader might care because the method keeps spatial reasoning accuracy at very low token budgets, cutting inference cost for 3D scene understanding.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SGS residual is identically zero under the stated normalization, so the claimed OT approximation is degenerate; the submodularity guarantee is not established for the actual algorithm.","rationale":"The reader's headline weakest_assumption was that the FST OT objective is a faithful surrogate for reasoning quality. I focus instead on an upstream, more concrete flaw: even if the surrogate were perfect, the algorithm described to optimize it does not do so under the stated definitions. The residual degeneracy follows directly from the normalization conventions stated in §3.1 and used in §4.2; it is not a matter of tuning or benchmark detail. This makes the paper's theoretical claim—that SGS approximates the semi-relaxed OT objective with provable guarantees—unsupported. The empirical tables may survive an implementation fix, so I do not recommend rejection; the paper should be revised to specify the partial-capacity normalization, prove or remove the submodularity claim, and release code to verify the selection rule. This aligns with the reader's CONDITIONAL verdict and the issues listed in its rationale (normalization, residual definition, submodularity claim), though the reader's headline weakest_assumption pointed at surrogate validity rather than algorithm degeneracy.","tokens_in":15017,"tokens_out":8413,"duration_ms":90651,"concrete_test":"Implement SGS exactly as stated in §3.5 with u uniform over S_l normalized to sum 1 and v from Eq. (10) normalized to sum 1, and log r_l after each greedy step; if r_l = 0 for all l < K, then Eq. (15) is degenerate and the claimed OT approximation is vacuous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algorithm that instantiates the OT objective is internally inconsistent as written. In §3.1 capacity vectors are normalized to sum 1, and Eq. (10) normalizes v to a distribution; CoverPrune-Lite (§4.2) confirms sum(v)=1. In SGS (§3.5), at each greedy step the plan P_l satisfies P_l 1 = u (Eq. 11). If u is uniform over the l selected tokens and normalized to sum 1, then total transported mass is 1. Since P_l^T 1 <= v componentwise and both sides sum to 1, the inequality is tight: P_l^T 1 = v, hence the residual r_l = [v - P_l^T 1]_+ in Eq. (14) is identically zero. Eq. (15) then scores every candidate as zero and selection is arbitrary. If instead u is intended as partial mass (1/K per selected token), the paper never says so, and the greedy step is no longer the standard submodular maximization greedy rule; the claimed constant-factor guarantee from [17,21] does not apply, and the local residual criterion is an unproven heuristic. Moreover, submodularity of the induced set function f(S) = -LOT(S;T) is asserted without proof, despite the FST cost (Eq. 8) being asymmetric (ReLU temporal term) and sample-normalized, which is exactly the setting where the cited covering results need verification. The paper's central mechanism—SGS approximating the OT coverage objective—therefore rests on an unstated normalization and an unproved approximation guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoverPrune, a training-free inference-time token pruning method for 3D vision-language models (3D VLMs). The method formulates pruning as an optimal transport (OT) problem that selects a compact subset of visual tokens maximizing coverage of the full token set under a feature-spatial-temporal (FST) cost and an informativeness-aware target capacity. To make the combinatorial selection tractable, the paper introduces a Spatial-Guided Greedy Selection (SGS) algorithm that greedily builds the subset using a residual-weighted local cost, and a faster variant, CoverPrune-Lite, based on Morton-order grouping. Experiments on ScanQA, SQA3D, Scan2Cap, and VSI-Bench with two base models (GS-Reasoner and VLM-3R) report consistent improvements over attention- and diversity-based baselines, especially under aggressive pruning (10% and 5% retention).","tokens_in":15398,"tokens_out":7789,"duration_ms":76349,"significance":"If the approach holds, it offers a new coverage-based paradigm for token pruning in 3D VLMs, moving beyond diversity- and attention-based heuristics, and it is training-free and plug-and-play. The empirical evaluation is a strength: results are reported across two base models, four benchmarks, and multiple retention ratios, with external benchmarks and no apparent fitting to test labels. The ablation study isolates the contributions of the FST cost terms and the capacity weighting. The Lite variant provides a practical low-overhead approximation. However, the theoretical justification of the SGS algorithm is a major weakness: the semi-relaxed OT formulation as written degenerates under the stated capacity normalization, and the submodularity/approximation claims are not established for the asymmetric, sample-normalized FST cost.","major_comments":[{"comment":"Under the normalization stated in §3.2 and confirmed in §4.2, both capacity vectors u and v sum to 1. The semi-relaxed problem in Eq. (11) imposes P1 = u and P^T 1 ≤ v. Since the total transported mass is 1 and v sums to 1, the inequality must be tight: P^T 1 = v exactly. Consequently the residual r_ℓ = [v − P_ℓ^T 1]_+ in Eq. (14) is identically zero at every greedy step, and the selection rule in Eq. (15) assigns a score of zero to every candidate token, making SGS arbitrary. If the authors intended u to carry partial mass (for example u_i = 1/K for each selected token, so that the total mass is ℓ/K at step ℓ), this must be stated explicitly, because it contradicts the normalization in §3.2 and §4.2, and the cited approximation guarantees from [17,21] would need to be rederived for that setting. As written, the central algorithm cannot be expected to produce the reported results without additional unspecified implementation details.","section":"§3.5, Eq. (14)"},{"comment":"The claim that the induced set function for subset selection under the relaxed Wasserstein objective is monotone submodular and therefore admits greedy constant-factor approximation is not established for the FST cost in Eq. (8). The temporal term ReLU(τ_s − τ_t) is asymmetric, and each discrepancy is min-max normalized within the sample, so the cost matrix is not a symmetric metric; the cited results [17,21] do not directly apply. Moreover, the SGS rule in Eq. (15) is not the standard greedy marginal-gain update for submodular maximization: it replaces the global marginal gain with a local residual-weighted sum over a 3D neighborhood. Even if the full OT objective were submodular, the approximation guarantee would not transfer to this local heuristic. The authors should provide a rigorous proof of submodularity for their specific cost or explicitly characterize SGS as a heuristic without theoretical guarantees.","section":"§3.5, third paragraph"}],"minor_comments":[{"comment":"The header 'Acc.%↑B-4↑Rouge↑CIDEr↑Meteor↑' contains a typo: 'B-4' should be 'BLEU-4', and the repeated 'Acc.%' labels are confusing. Please clean up the table formatting.","section":"Table 1 caption"},{"comment":"The rows for DTC, CoverPrune, and CoverPrune-Lite are run together with missing spaces (e.g., '62840.73.4723.879.85'), making the efficiency numbers difficult to read. Please fix the LaTeX/table formatting.","section":"Table 5"},{"comment":"The sentence 'At 20% token retention, CoverPrune preserves 92.4% of full-token performance' is supported by Table 2 (59.76/64.70 ≈ 0.9236), but the abstract claims 'near-zero performance loss with 10% visual tokens on general 3D tasks'; please clarify which benchmark and retention ratio each claim refers to.","section":"§5.2"},{"comment":"The phrase 'normalize it to match the pruning budget' is ambiguous: Eq. (10) normalizes v to a probability distribution summing to 1, not to the budget K. Please rephrase to avoid confusion, especially since §4.2 later uses group masses of 1/K.","section":"§3.4, Eq. (10)"},{"comment":"The FST cost weights λ_f, λ_x, λ_τ and capacity weights α_f, α_x, α_τ are all set to 1 in experiments, but no sensitivity analysis is provided. A short robustness study would strengthen the claim that the multi-domain weighting is well-behaved.","section":"§5.1"},{"comment":"Some references are incomplete: [3] lacks a year/venue, and [15] is missing the full proceedings information. Please complete all bibliographic entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the SGS residual is confirmed by the manuscript text: with u and v both normalized to sum to 1, Eq. (14) is identically zero. This is a load-bearing technical flaw in the central algorithm as written. The empirical results are extensive and consistent, so the paper may be salvageable if the authors clarify the capacity normalization (e.g., partial-mass source capacities), re-derive or remove the theoretical guarantees, and describe the actual algorithm accurately. I recommend major revision with a request for a corrected derivation and, if needed, rerun experiments under the clarified protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2608.13226. First, the empirical package is solid: across ScanQA, SQA3D, Scan2Cap, and VSI-Bench, with two base models (GS-Reasoner and VLM-3R), CoverPrune and CoverPrune-Lite consistently beat attention- and diversity-based baselines at 10–20% token retention, and the ablations show each FST term contributes. Second, the theoretical core does not survive contact with its own equations. As written, SGS's residual in Eq. (14) is identically zero: u and v are both normalized to sum to 1, so P_l 1 = u forces P_l^T 1 ≤ v with equal total mass, hence equality, and every candidate scores zero in Eq. (15). The paper never states the alternative (partial mass, 1/K per selected token), and if that were intended, the greedy rule is no longer standard submodular maximization, so the cited approximation guarantees from [17,21] do not apply. Submodularity of the FST-induced set function is also asserted, not proved, for a cost with an asymmetric ReLU temporal term.\n\nWhat is genuinely new: the coverage framing for 3D VLM pruning, the FST transport cost, the informativeness-aware capacity, and the Lite variant's Morton-order grouping with capacity-balanced blocks. These are real contrasts to diversity-based methods. The OT prototype machinery is prior work (SPOT, Partial Wasserstein Covering), and the paper cites it honestly. The reported gains are a few points on average, not a breakthrough, but consistent.\n\nMinor soft spots: hyperparameters like κ, n, and g in the FST cost and capacity are never given, and no code is attached, so the numbers are not independently reproducible. The ablations report only point estimates without variance across runs. Some relative accuracy numbers exceed 100%, suggesting pruning occasionally regularizes; the paper does not comment on that.\n\nThe central argument — minimizing OT coverage loss preserves the evidence the VLM needs — is plausible but demonstrated only indirectly through downstream accuracy. That is acceptable for a systems paper, but it means the theory should be either fixed or softened. The Lite variant is conceptually cleaner than full SGS and likely carries the practical value.\n\nReading group? Yes — it is a nice example of a paper whose empirical results can outrun its own derivations. I would not cite the theoretical claims, and I would wait for a revised draft with the SGS normalization fixed and the missing hyperparameters. Recommendation: send it to peer review. The empirical direction is useful enough that a serious referee should engage, but the authors need to fix SGS, prove or drop the submodularity claim, and release the code and hyperparameters before acceptance.","headline":"Strong empirical token pruning results for 3D VLMs, but the SGS algorithm is degenerate as stated and the theoretical guarantees collapse; the Lite variant is the cleaner contribution.","tokens_in":15942,"tokens_out":5560,"would_cite":false,"duration_ms":51280,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Pruning visual tokens by coverage, formalized as optimal transport, preserves spatial reasoning under extreme compression: CoverPrune keeps 92.4% of full-token VSI-Bench performance at 20% retention and beats attention- and…","keywords":["visual token pruning","optimal transport","3D vision-language models","spatial reasoning","evidence coverage","token efficiency","inference acceleration"],"falsifier":"Take the kept sets produced by CoverPrune and by each baseline at the same budget on VSI-Bench, compute the FST transport cost of Eqs. (8)–(10) for each kept set, and compare those costs with per-question accuracy. If the kept set with the lowest transport cost is not the one with the highest accuracy across the benchmark, then the defined coverage objective is not the mechanism producing the pruning gains.","tokens_in":14836,"feed_emoji":"🎯","tokens_out":12267,"duration_ms":105159,"temperature":0.7,"pith_summary":"Three-dimensional vision-language models gain spatial reasoning by injecting geometry into the visual token stream, but the resulting token counts make inference expensive. CoverPrune argues that the right pruning objective is coverage — keeping a compact set of prototype tokens that collectively explain the full evidence — rather than diversity, which tends to discard representative tokens in favor of outliers. The paper formalizes coverage as an optimal transport problem: the retained tokens serve as prototypes that must carry the mass of the full token set under a cost combining feature similarity, 3D distance, and temporal order, with informative tokens receiving extra target capacity. A semi-relaxed transport formulation and a spatial-guided greedy selection make the intractable subset choice practical, and a block-structured variant runs in near-linear time. If the argument holds, aggressive token pruning can accelerate 3D VLMs without sacrificing spatial reasoning; the reported numbers show 92.4% of full-token VSI-Bench performance retained at 20% of the tokens.","feed_headline":"Prune 3D-VLM tokens by coverage to keep 92% of reasoning","feed_subtitle":"At 20% token retention, optimal-transport coverage beats attention- and diversity-based pruning on VSI-Bench.","key_machinery":"The central machinery is the coverage-as-OT objective: a semi-relaxed optimal transport problem in which the retained tokens are the source, the original tokens are the target, and the objective is to minimize transport cost while allowing unused target capacity. The cost matrix is the Feature-Spatial-Temporal (FST) cost $C_{ij} = \\lambda_f \\hat{d}_f + \\lambda_x \\phi_\\kappa(\\hat{d}_x) + \\lambda_\\tau \\hat{d}_\\tau$, blending normalized feature cosine distance, a log-compressed 3D spatial distance, and a one-sided temporal penalty $\\mathrm{ReLU}(\\tau_s - \\tau_t)$ so a later token cannot cheaply cover an earlier one. Per-token target capacities $v_j$ are reweighted by local FST distinctiveness, concentrating transport mass on tokens that are hard to approximate from their 3D neighbors. The semi-relaxed inequality $P^\\top \\mathbf{1} \\le v$ is the load-bearing relaxation: it turns the combinatorial subset selection into a monotone submodular objective with greedy constant-factor guarantees, and Spatial-Guided Greedy Selection implements the greedy step by evaluating marginal costs only within each candidate's $g$ nearest 3D neighbors. CoverPrune-Lite replaces this iteration with a Morton-code ordering, capacity-balanced grouping into $K$ groups, and one within-group prototype per group, reducing the pruning pass to $O(N \\log N)$.","core_discovery":"The paper's central claim is that coverage, not diversity, is the correct objective for pruning visual tokens in 3D VLMs, and that this objective can be cast as an optimal transport problem with a Feature-Spatial-Temporal cost. In this view, the selected subset $S$ is the source support and the full token set $T$ is the target support; pruning minimizes the cost of transporting the target distribution to $S$, subject to uniform source capacities and an informativeness-weighted target capacity $v$ computed from local FST distinctiveness. The semi-relaxed constraint $P^\\top \\mathbf{1} \\le v$ leaves residual target capacity unused, which is what makes subset selection a monotone submodular maximization admitting greedy approximation. The paper reports that this coverage-based selection outperforms attention-based and diversity-based pruning on Scan2Cap, ScanQA, SQA3D, and VSI-Bench across two base models, and that at 20% token retention it preserves 92.4% of full-token VSI-Bench performance, with the advantage growing as the retention ratio drops to 10% and 5%.","pith_inferences":["A direct test the paper does not run: compute the FST transport cost of the kept sets produced by each pruning method and correlate it with accuracy across VSI-Bench questions. If the lowest-cost kept set is not the most accurate, the coverage proxy itself is not the operative mechanism.","The paper's transfer claim is untested beyond two 3D VLMs on indoor-scan benchmarks; a natural extension is outdoor scenes or video-based spatial grounding, where the temporal penalty direction and 3D coordinate estimation would have to be re-examined.","The one-sided temporal cost encodes an assumption about evidence order, namely that earlier tokens should cover later ones. Reversing that direction on the same benchmarks would isolate how much of the gain comes from temporal asymmetry rather than from spatial and feature coverage.","Since the method operates on raw tokens before the prefill stage, it should compose with later-stage accelerations such as KV-cache eviction or attention sparsification; the paper does not measure such compounded speedups."],"forward_implications":["At 20% retention, CoverPrune keeps 92.4% of full-token VSI-Bench accuracy with the stronger base model, while at 10% and 5% it holds a clear lead over all four baselines, indicating the coverage objective degrades more gracefully under compression.","On general 3D reasoning benchmarks, 10%-retention CoverPrune roughly matches or exceeds the accuracy that diversity-based baselines reach at 20% retention on multiple metrics, so aggressive pruning no longer forces a proportional accuracy loss.","Ablations show that removing the feature term from the FST cost costs about 3.6 overall VSI-Bench points, while removing the geometry and temporal terms costs roughly 0.7–0.8 points each, so semantic affinity is the dominant term but spatial and temporal cues are not noise.","CoverPrune-Lite cuts pruning time from 2.53 seconds to 0.41 seconds on VSI-Bench while slightly improving relative accuracy (88.01% vs 87.84%), making coverage-based pruning usable at deployment latencies.","Because CoverPrune is training-free and plugs in right before the language-model prefill stage, its benefits are available without fine-tuning, which is what the two-base-model experiments demonstrate."],"supporting_citations":[{"why":"defines the optimal transport problem that the coverage objective is built on.","marker":"[33]"},{"why":"establishes NP-hardness and greedy approximation guarantees for OT-based prototype selection.","marker":"[17]"},{"why":"supplies the partial-Wasserstein covering view behind the semi-relaxed target inequality.","marker":"[21]"},{"why":"provides the semi-relaxed OT formulation and entropic regularization used by the solver.","marker":"[30]"},{"why":"the attention-based pruning baseline that CoverPrune must surpass under matched budgets.","marker":"[39]"},{"why":"the diversity-based video pruning baseline representing the competing paradigm.","marker":"[32]"},{"why":"the 3D-specific diversity-driven token compression baseline.","marker":"[19]"},{"why":"the attention-diversity fused baseline for egocentric video reasoning.","marker":"[23]"},{"why":"the VSI-Bench spatial-reasoning benchmark where the 92.4% retention result is measured.","marker":"[38]"},{"why":"the stronger of the two base 3D VLMs used for the main experiments.","marker":"[9]"}],"fun_headline_variants":[],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that minimizing the transport cost between the kept tokens and all original tokens, under the paper's feature-spatial-temporal weighting and capacity vector, faithfully tracks whether the retained visual evidence is enough for correct spatial reasoning; the paper checks this only through final benchmark accuracy.","fun_headline_variants_meta":{"error":"Client error '402 Payment Required' for url 'https://api.deepseek.com/chat/completions'\nFor more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/402"},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:52:45.749948+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the kept sets produced by CoverPrune and by each baseline at the same budget on VSI-Bench, compute the FST transport cost of Eqs. (8)–(10) for each kept set, and compare those costs with per-question accuracy. If the kept set with the lowest transport cost is not the one with the highest accuracy across the benchmark, then the defined coverage objective is not the mechanism producing the pruning gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the partial-Wasserstein covering view behind the semi-relaxed target inequality."},{"cited_title":"In: The Thirty-ninth Annual Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"the diversity-based video pruning baseline representing the competing paradigm."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"the 3D-specific diversity-driven token compression baseline."}],"review_version":1}