{"id":"909c2ea9-1a9c-4911-9028-a558f4a3af1c","arxiv_id":"2608.04610","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"HiSC compresses visual tokens for 3D VLMs by graph-based merging before inference and hierarchical spatial clustering pruning inside the LLM, achieving about 90% token reduction while retaining roughly 92% of original performance.","lead":"HiSC is a training-free token compression method for 3D vision-language models that merges and prunes redundant scene tokens using spatial connectivity and object-like clusters. On five indoor scene benchmarks it removes about 90% of visual tokens while keeping about 92% of original answer quality, with larger gains over prior pruning baselines under aggressive compression.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cluster-budget allocation in §3.3.2 may be infeasible at 90% compression because it requires N(l) ≥ K, yet the paper never reports K or a fallback when the constraint is violated.","rationale":"The paper presents a plausible training-free compression method with strong empirical results, and I would not reject it solely on the reader's averaging concern: SGraM merges only spatially adjacent and semantically similar tokens, and the ablations show it helps overall. However, I identified a more crisp issue in the specification of SCluP. The stated budget allocation constraints (n_k ≥ 1, Σ n_k = N(l)) are jointly satisfiable only if N(l) ≥ K, and with non-uniform importance weights the condition is even stronger. The paper never reports K nor discusses this constraint at 90% compression, where N(l) may be as low as ~10% of the original token count. If the implementation silently relaxes n_k ≥ 1, the claimed object-level coverage guarantee is void; if it inflates N(l), the reported token reduction is not accurate. This does not automatically falsify the empirical results, but it makes the method description incomplete and the central mechanism unverified. The concrete test using the released code would settle it. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":15113,"tokens_out":10301,"duration_ms":115812,"concrete_test":"Run the released HiSC code on a representative set of ScanNet scenes (e.g., the five benchmarks) under the Extreme Compression setting. At each selected pruning layer (1, 4, 7, 16, 23), record the number of object-centric clusters K from Eq. (4) and the per-cluster importance weights. Verify whether the stated n_k ≥ 1 and Σ n_k = N(l) constraints can be satisfied simultaneously. Report the number of scenes/layers where the allocation is infeasible and the fallback actually used. If it is feasible everywhere, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3.2 specifies cluster-level budget allocation with n_k ≥ 1 for every cluster and Σ n_k = N(l). This is jointly satisfiable only if N(l) ≥ K, where K is the number of object-centric clusters from Eq. (4). Under the reported 90% token reduction (Avg. Tokens 9.8%, Table 1), N(l) at the deepest pruned layers is a small fraction of the original token count. K is the number of connected components across K_sem=16 semantic groups in a full ScanNet scene; no K values are reported. If K exceeds N(l) at any selected layer, the allocation is infeasible. The implementation must have either relaxed n_k ≥ 1 (breaking the claimed object-level coverage) or increased N(l) (breaking the claimed compression ratio). The paper provides no fallback or discussion, so the central claim that HiSC preserves object-level completeness under extreme compression is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HiSC, a training-free token-compression framework for multi-view 3D vision-language models. It operates in two stages: SGraM merges redundant tokens by connected components of a joint geometric-semantic adjacency graph before LLM inference, and SCluP performs hierarchical cluster-level budget allocation and intra-cluster importance-diversity sampling during LLM inference. Experiments on five ScanNet-based benchmarks (ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, SQA3D) using a LLaVA-Video 7B backbone compare HiSC against several 2D/3D pruning baselines under roughly 50%, 25%, and 10% token retention, reporting higher retained average scores and lower TFLOPS in all three regimes.","tokens_in":15363,"tokens_out":4775,"duration_ms":51420,"significance":"If the claims hold, HiSC is a useful plug-and-play compressor for multi-view 3D VLMs: it generalizes token selection to object-centric clusters, explicitly models cross-view geometric redundancy, and requires no training. The paper has clear strengths: a unified evaluation across five benchmarks, multiple baselines, component ablations, qualitative visualizations, and a public code link. The main caveats are empirical: no variance estimates are reported, several key thresholds are not analyzed, and the cluster-budget constraint is not verified against the reported compression ratios.","major_comments":[{"comment":"The budget allocation constraint n_k ≥ 1 with Σ_k n_k = N(l) is only satisfiable when N(l) ≥ K for every pruned layer, where K is the number of clusters produced by Eq. (4). The paper never reports K or the per-layer N(l) values, nor does it describe a fallback when K > N(l). Under the Extreme setting in Table 1, the average retained token count is 9.8% of the original sequence, so a scene with a large number of connected components could easily violate the constraint; the implementation must have either relaxed n_k ≥ 1 (undermining the object-coverage claim) or retained more tokens than the reported average. Please report K, N(l), and the exact handling of infeasible budgets.","section":"§3.3.2"},{"comment":"All reported numbers come from a single evaluation run, with no error bars, confidence intervals, or significance tests. Several headline comparisons are small: at Light compression, Ours achieves Relative Score 98.87% versus VFlowOpt's 98.15%, and per-metric margins are often below one point (e.g., ScanRefer Acc@0.25: 57.90 vs 57.47). Without variance estimates or significance testing, the claim that HiSC consistently outperforms all baselines is not statistically established.","section":"Table 1 / §4.2.1"},{"comment":"The method has several free hyperparameters—τ_geo, τ_sem, λ in Eq. (2)/(5), K_sem, τ_dist, the pruning layers, and β—but no sensitivity analysis or tuning procedure is reported. For a training-free method that is intended as a drop-in compressor, the absence of any analysis showing how performance varies with these thresholds is a load-bearing gap, especially for the extreme-compression regime where cluster structure determines whether object-level coverage is preserved. Please add a sensitivity study or a principled default-selection rule.","section":"§4.1.3"},{"comment":"Merging every token in a connected component into one super-token via feature and position averaging presumes that all tokens within the component are near-duplicates. This assumption is not tested: the paper does not report component-size statistics, measure merging-induced error, or analyze failure cases at object boundaries (e.g., a poster on a wall). Since SGraM removes tokens before the LLM ever sees them, any information lost at this stage cannot be recovered by SCluP; an analysis of merging error or a comparison against a no-merge oracle would strengthen the central claim.","section":"§3.2.2"}],"minor_comments":[{"comment":"In the abstract, \"show validate\" should be \"validate\" or \"show the effectiveness of\".","section":"Abstract"},{"comment":"The phrase \"HiSC, a acceleration framework\" should be corrected to \"an acceleration framework\".","section":"Introduction, §1"},{"comment":"The sentence \"This decoupled design provides special constraints enforce spatial validity\" is ungrammatical and should be revised.","section":"§3.2.1"},{"comment":"The footnote markers for FastVGGT and BTP are not referenced in the main text; please explain the restrictions for these baselines in the caption or in §4.2.1.","section":"Table 1"},{"comment":"Reference [14] lists the first author as \"kaiyuan Li\"; the capitalization should be corrected to \"Kaiyuan Li\".","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict is fair. The method is not circular, and the empirical design is reasonable, but the infeasible-budget issue in §3.3.2 and the lack of variance and threshold analysis justify a major revision. I would not reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: HiSC is a genuinely useful token-compression method for multi-view 3D VLMs, and the empirical case is decent. The one thing I'd want answered before relying on it is whether the cluster-budget allocation actually holds at the reported extreme compression, because the paper never reports the number of clusters or a fallback when K > N(l). That's a fixable clarification, not a fatal flaw.\n\nWhat's new: most prior token pruning is 2D and token-level. HiSC lifts this to 3D by organizing tokens into spatially grounded clusters using geometric adjacency plus semantic similarity, then merges connected components before the LLM and does hierarchical cluster-level allocation inside the LLM. That combination is new for 3D VLMs, and it shows real gains: at ~10% average tokens it keeps a 92.46% relative score across five benchmarks, comfortably above the best baseline (VisPruner at 89.43%). The ablations make sense and show both components contribute. Training-free, code is public.\n\nWeak spots, in order: (1) The stress-test concern is real. Section 3.3.2 imposes n_k >= 1 for every cluster and sum n_k = N(l), which is only satisfiable when N(l) >= K. At 9.8% average tokens, some layers may have N(l) smaller than the number of connected components K over 16 semantic groups. No K is reported, no fallback is described. The paper's object-coverage claim under extreme compression depends on this, so a referee should ask for the actual K distribution and the exact allocation rule when the constraint binds. (2) No error bars or significance tests; several key thresholds (tau_geo, tau_sem, lambda, beta) are fixed without sensitivity analysis. (3) The 'minimal performance degradation' phrasing overstates what 92.46% relative score means; it's good, but it's a 7.5% average drop, so 'minimal' is a stretch. (4) Minor: abstract has a typo, and the FastVGGT baseline is restricted in coverage, which weakens the comparison slightly.\n\nBottom line: this paper deserves a serious referee. The core method is sound, the experiments are broad, and the open questions are answerable with additional reporting rather than rework. I'd send it to review and ask for the allocation details and error bars.","headline":"A solid, useful token-compression method for 3D VLMs with a real but answerable question about cluster-budget feasibility under extreme compression.","tokens_in":15856,"tokens_out":2672,"would_cite":true,"duration_ms":27234,"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":"This paper claims that a training-free pipeline can remove about 90% of the visual tokens fed to a multi-view 3D vision-language model and still keep more than 90% of its benchmark performance, by merging and pruning tokens as spatially…","keywords":["3D vision-language models","visual token pruning","token compression","spatial clustering","graph-based merging","multi-view scene understanding","training-free acceleration","3D scene understanding"],"falsifier":"Take one ScanNet scene with a poster, sign, or other small textured detail on a large planar surface; run HiSC at 90% compression and ask a question whose answer is written on or determined by that detail. If the averaged super-token blurs the detail, the model should answer incorrectly while the uncompressed model and a version that keeps one representative token per component answer correctly.","tokens_in":14946,"feed_emoji":"✂️","tokens_out":7937,"duration_ms":79079,"temperature":0.7,"pith_summary":"HiSC is a training-free compressor for multi-view 3D vision-language models. It argues that most visual tokens in a 3D scene are structured redundancy, coming from repeated views of the same physical surface and from large blank regions, and that this redundancy can be removed before and during LLM inference without retraining. The paper reports that cutting about 90% of visual tokens keeps average benchmark performance above 90% of the uncompressed model, and that moderate compression is nearly lossless. A reader should care because 3D VLMs are expensive to run, and HiSC is an inference-time fix rather than a new training run.","feed_headline":"Cutting 90% of 3D scene tokens keeps reasoning above 90%","feed_subtitle":"A training-free compressor for 3D vision-language models drops most visual tokens while keeping nearly all benchmark accuracy.","key_machinery":"The load-bearing objects are the connectivity graph and the object-centric cluster. The graph edge condition uses Euclidean distance in 3D below a threshold AND cosine similarity above a threshold, so edges exist only between tokens that are both geometrically adjacent and appearance-consistent; connected components of this graph capture physically consistent regions transitively, and each component is collapsed into one averaged super-token. Inside the LLM, SCluP first groups tokens by semantic similarity and then splits each group into spatially connected components, yielding clusters that line up with physical instances. Token budgets are allocated to clusters in proportion to squared importance scores, with a floor of one token per cluster to keep every object represented, and tokens within a cluster are selected by importance with a diversity bias toward tokens far in combined spatial-semantic distance. These mechanisms carry the argument because they turn compression into a coverage problem over physical regions instead of a per-token importance contest.","core_discovery":"The central claim is that token-level pruning, which works for 2D VLMs, fails in multi-view 3D because redundancy is structured by physical space: tokens from the same surface are near-duplicates, and overlapping cameras repeat them. HiSC replaces independent token selection with cluster-level processing. First, a spatial graph-based merging stage builds connections only between tokens that are both spatially close and visually similar, extracts connected components, and averages each component into one super-token before the LLM. Second, a spatial clustering-based pruning stage partitions tokens into object-centric clusters, allocates token budgets across clusters by squared importance with at least one token per cluster, and picks tokens inside each cluster by importance plus a joint spatial-semantic diversity measure. Across grounding, dense captioning, and question answering benchmarks, the paper finds that this pipeline keeps object coverage and fine details even at 10% token retention, where independent pruning fragments objects and over-concentrates tokens.","pith_inferences":["The graph-merge idea should transfer to other multi-view token streams, such as driving cameras or video, where overlapping observations create the same kind of spatially structured redundancy.","A stress test for the 90% claim would be scenes with small text or thin decorations on large planar surfaces, where averaging a connected component may blur task-critical details that independent token selection would keep.","If the connectivity assumption is right, HiSC should improve as camera overlap increases; a controlled experiment varying viewpoint density could separate the benefit of merging from the benefit of pruning.","The paper reports aggregate benchmark scores, so a natural next measurement is per-answer reliability: whether the errors under 90% compression concentrate on questions about small objects, colors, or relationships rather than being spread evenly."],"forward_implications":["At 50% and 75% token budgets, HiSC keeps nearly all of the uncompressed model's accuracy on the five reported benchmarks, indicating that structured redundancy can be removed before it dilutes attention.","At 10% token retention, HiSC reports a relative score of 92.46%, higher than the compared baselines, which cluster-level coverage matters most under aggressive compression.","The two stages are complementary: the pre-LLM merge removes redundant tokens so the in-LLM budget is spent on informative regions, while the in-LLM prune adapts allocation to reasoning-relevant objects.","Because HiSC is training-free and operates on tokens already constructed by a multi-view 3D VLM, the same pipeline can be attached to other models with the same token format."],"supporting_citations":[{"why":"Video-3D LLM is the base model whose pretrained weights and benchmark protocol HiSC compresses; the upper-bound performance comes from this model.","marker":"[55]"},{"why":"FastVGGT identifies token collapse from view overlap in multi-view geometry transformers and motivates structured merging; it is also a baseline in the high-compression comparison.","marker":"[30]"},{"why":"GitMerge3D documents severe token redundancy in 3D transformers and supplies the graph-merging motivation for spatially consistent regions.","marker":"[32]"},{"why":"Balanced token pruning provides the intra-cluster selection strategy adopted by SCluP and is a baseline at medium compression.","marker":"[14]"},{"why":"FastV is an attention-based token pruning baseline compared at all compression levels.","marker":"[6]"},{"why":"ScanRefer is the object grounding benchmark used for Acc@0.25 and Acc@0.5 evaluation.","marker":"[4]"},{"why":"Scan2Cap is the dense captioning benchmark used for BLEU-4 and CIDEr at 0.5 IoU.","marker":"[8]"},{"why":"ScanQA is the 3D question answering benchmark used for CIDEr and exact match evaluation.","marker":"[2]"},{"why":"SQA3D is the situated question answering benchmark used for exact match evaluation.","marker":"[24]"}],"fun_headline_variants":["Spatial clustering compresses 3D scene tokens 90% without training","HiSC: 90% fewer 3D tokens, near-full reasoning","Training-free 3D token compression via spatial clustering","Cut 90% of 3D vision tokens, keep reasoning intact","Spatial graph merges redundant 3D tokens, pruning preserves details"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every token inside a spatially connected, appearance-similar region is a near-duplicate whose averaged super-token preserves what the LLM needs; if such a region contains subtle but task-critical variation, like a poster on a wall or a boundary between similar objects, averaging can erase it.","fun_headline_variants_meta":{"raw":{"variants":["Spatial clustering compresses 3D scene tokens 90% without training","HiSC: 90% fewer 3D tokens, near-full reasoning","Training-free 3D token compression via spatial clustering","Cut 90% of 3D vision tokens, keep reasoning intact","Spatial graph merges redundant 3D tokens, pruning preserves details"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1580,"prompt_tokens":1002,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":482}},"tokens_in":618,"tokens_out":578,"duration_ms":5439,"temperature":1.0,"reasoning_tokens":482,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:46:18.524001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one ScanNet scene with a poster, sign, or other small textured detail on a large planar surface; run HiSC at 90% compression and ask a question whose answer is written on or determined by that detail. If the averaged super-token blurs the detail, the model should answer incorrectly while the uncompressed model and a version that keeps one representative token per component answer correctly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Video-3D LLM is the base model whose pretrained weights and benchmark protocol HiSC compresses; the upper-bound performance comes from this model."},{"cited_title":"Doan, Roger Wattenhofer, Vien Anh Ngo, Mathias Niepert, Daniel Sonntag, and Paul Swoboda","cited_arxiv_id":null,"evidence_quote":"GitMerge3D documents severe token redundancy in 3D transformers and supplies the graph-merging motivation for spatially consistent regions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Balanced token pruning provides the intra-cluster selection strategy adopted by SCluP and is a baseline at medium compression."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FastV is an attention-based token pruning baseline compared at all compression levels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ScanRefer is the object grounding benchmark used for Acc@0.25 and Acc@0.5 evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Scan2Cap is the dense captioning benchmark used for BLEU-4 and CIDEr at 0.5 IoU."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ScanQA is the 3D question answering benchmark used for CIDEr and exact match evaluation."}],"review_version":1}