{"id":"4f874762-55be-4895-afc5-82c0360e8c4a","arxiv_id":"2608.04515","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"CARVE compresses 3D medical volume tokens by adaptively windowing the depth axis, selecting anchors and retrieved evidence, and folding redundant tokens, preserving near-full VQA quality at about one fifth of the tokens.","lead":"CARVE is a training-free method that removes about 80% of the visual tokens a 3D medical image model must process, while keeping most of its answer quality. It decides which CT slices and locations carry non-redundant evidence, keeps those as anchors, and folds the rest away before the language model runs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise that encoder cosine/attention scores act as task-relevant token importance is untested; a direct dropped-token prediction test would settle it.","rationale":"The reader and I converge on the same load-bearing premise. The reader's verdict is CONDITIONAL, which I support. The paper's empirical results are credible and well supported by comparative tables; CARVE is first on all four AMOS-MM report metrics, first in aggregate retention on three backbones, and its ablation shows each component contributes. The main correctness risk is that the token-importance proxy (encoder cosine similarity and attention) may not align with what the projector and LLM need for a given query. The paper does not validate this proxy by measuring whether dropped tokens change model predictions. My concrete test would settle this by directly comparing Full and CARVE predictions on a sample of questions, and by checking whether disagreements resolve when more tokens are retained. This is addressable and does not undermine the novelty or the empirical frontier; it just requires an additional validation experiment. I also note the 'Rel.' retention metric could mask failures on rare or localized findings, which is a further reason to run the dropped-token prediction test. On balance, the central concern is real but addressable, so the verdict should remain CONDITIONAL.","tokens_in":13681,"tokens_out":1623,"duration_ms":11603,"concrete_test":"Run a dropped-token prediction test on the AMOS-MM or 3D-RAD validation set: for a random sample of 100-200 questions, compute the Full model's prediction and the CARVE-compressed model's prediction at r=0.2. Flag cases where the predictions disagree. Then, for each disagreement, re-run CARVE with a budget increase (e.g., r=0.25) and check whether the disagreement resolves; if it does, the discarded tokens were task-relevant and the proxy failed. As a second, cheaper check, add CARVE's dropped tokens back at random positions and measure whether the Full model's score changes; if it does, the dropped tokens carry information the LLM uses.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"CARVE's central claim is that the 2.5D allocation scheme, not the token budget, drives the quality-efficiency frontier. That claim depends on the assumption that token importance can be scored entirely in the frozen 2D encoder's feature space: slice redundancy by mean-pooled cosine similarity (Eq. 1), localized deviation by cosine residual to same-coordinate neighbors on adjacent slices (Eq. 2), and anchor usefulness by last-layer self-attention (Eq. 6). The paper never validates this proxy, e.g., by measuring whether dropped tokens change model predictions. If the encoder's cosine-similarity and attention maps do not reflect what the projector and LLM need to answer a query, CARVE can discard task-critical evidence while still respecting the token budget. The qualitative case (Fig. 7) is suggestive but is a single cherry-picked example, not a systematic check. The concern is also aggravated by the 'Rel.' retention metric: it is computed relative to the Full model's score, so a method that drops evidence uniformly can score well on aggregate while failing on rare or localized findings. None of the reported evaluations isolate this failure mode. This is a real soft spot, but it is addressable; the extensive comparative evaluation makes it a conditional acceptance rather than a rejection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CARVE, a training-free token compression framework for slice-based 3D medical vision-language models. It profiles cross-slice feature change, partitions the volume into adaptive windows, splits a target token budget between intra-slice anchors and global inter-slice retrieval, and folds remaining tokens into nearby anchors before the frozen projector and LLM. The authors present scaling analyses suggesting diminishing returns for larger token budgets, and evaluate CARVE on AMOS-MM report generation and VQA, 3D-RAD, and M3D-VQA with Hulu-Med-7B, plus transfer to Hulu-Med-4B and Qwen3-VL. At roughly 20% retained tokens, CARVE reports the highest relative retention among compressed baselines on AMOS-MM report generation (Rel. 87.07 vs. the best baseline 80.89) and preserves 96.97-99.08% of full-token performance on the three VQA benchmarks, supported by a budget sweep and ablations that isolate window allocation, selection roles, and folding.","tokens_in":14085,"tokens_out":8868,"duration_ms":69057,"significance":"If the reported results hold, CARVE is a practically useful drop-in compressor for slice-based 3D medical MLLMs: it is training-free, operates at the post-encoder pre-projector interface, and achieves substantial latency/memory reductions with modest aggregate quality loss. The paper's strengths are its breadth of evaluation (three benchmarks, multiple backbones, a budget sweep, a transfer study, and an ablation design that isolates individual factors) and the clarity of its comparative claims. The main risk is that the method's core scoring proxy—encoder cosine similarity and self-attention—is never directly validated against task relevance, and the aggregate relative-retention metric may mask failures on rare localized findings. These issues are addressable, but they are load-bearing for the central quality-efficiency claim.","major_comments":[{"comment":"The entire allocation is driven by scoring rules computed in the frozen 2D encoder's feature space: slice change by mean-pooled cosine similarity (Eq. 1), localized deviation by cosine residual to aligned neighbors (Eq. 2), and anchor usefulness by last-layer self-attention (Eq. 6). The paper never validates that these scores reflect what the projector and LLM need for the target tasks. I ask for a direct diagnostic, for example measuring how often dropped tokens are actually required for a correct answer (by comparing predictions when high-scoring versus low-scoring tokens are removed), or stratifying the evaluation by finding size or localization. The single qualitative example in Fig. 7 is suggestive but cannot carry this load. Without such validation, the reported aggregate 'Rel.' scores cannot rule out systematic loss of rare, task-critical evidence.","section":"Method, Eqs. (1)-(2), (6)"},{"comment":"All results appear to come from a single run per condition; no error bars, confidence intervals, or significance tests are reported. Some of the decisive margins are small (e.g., M3D-VQA ACC: CARVE 76.97 vs. MMTok 77.01; M3D-VQA BERT: CARVE 56.60 vs. FastVID 56.61), so the claim that CARVE 'leads all compression baselines' is not statistically established. In addition, the method introduces several hyperparameters (rho, K_slice, tau, r_nms, kappa, tau_m) that are fixed to single values, with only the budget r swept. A sensitivity analysis over at least rho and K_slice is needed to confirm that the reported quality-efficiency frontier is not an artifact of a particular hyperparameter setting.","section":"§4.1, Tables 1-2, Fig. 5"},{"comment":"The evaluation never isolates the failure mode the method is designed to prevent: localized evidence that appears on only a few slices. The task-type breakdown in Fig. 4 reports aggregate accuracy per category, and the ablation table reports aggregate ACC and ROUGE-L, so a method that preserves coarse anatomy while dropping small findings would not be detected. I request a per-case or per-finding recall measure (e.g., accuracy on cases with small lesions, or retrieval recall against annotated regions) in addition to the aggregate metrics. This is especially important because the paper's stated motivation is maintaining a recall path for localized cross-slice evidence.","section":"§4.2, Fig. 4, Table 3"}],"minor_comments":[{"comment":"The definition of N_z(j) as 'available tokens at the same in-plane coordinate on slices z_j +/- 1' should specify how ties and missing coordinates are handled, especially if slice grids are not perfectly aligned.","section":"Eq. (2)"},{"comment":"The figure repeats '#Tok: 2073' for several methods and uses a 'Better' arrow without a legend, making it hard to identify which symbol corresponds to which method; please add a legend and unique labels.","section":"Fig. 1"},{"comment":"The caption uses 'full method' to refer to the complete CARVE model, while Tables 1-2 use 'Full' for the uncompressed reference; please rename one of these to avoid confusion.","section":"Table 3 caption"},{"comment":"The scaling analysis would benefit from a precise statement of the slice counts and in-plane resolutions used, since the claim that in-plane resolution beats extra slices is central to the motivation.","section":"Fig. 2"},{"comment":"There is a typo: 'The scores j remains fixed' should read 'The score s_j remains fixed.'","section":"Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the contribution is potentially significant. The main concern is the unvalidated scoring proxy; if the authors add the requested diagnostic and sensitivity analysis, I would expect the revised version to be acceptable. I also note that several baselines are recent preprints, so the authors should verify that the citations are accurate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers what it promises: a training-free, pre-LLM token compression method for slice-based 3D medical MLLMs that removes about 80% of tokens while keeping near-full VQA performance and beating all compression baselines on AMOS-MM report generation. The central idea is genuinely new in this space: instead of pruning whole slices (as MedPruner does) or treating the volume as a flat bag of tokens, CARVE allocates one budget anisotropically between in-plane anchors and cross-slice retrieval, using adaptive depth windows and local folding. The scaling analysis is also a useful contribution on its own — the finding that in-plane resolution matters more than adding slices at matched budgets is clearly demonstrated and motivates the method.\n\nThe evidence base is substantial and mostly convincing. Three benchmarks, four backbones including a transfer study, a budget sweep, and a careful ablation that isolates each component all point the same way. The comparative tables show CARVE leads on every AMOS-MM report metric and achieves the best aggregate retention on every backbone. The qualitative case in Fig. 7 is suggestive, though of course it is a single example.\n\nThe soft spots are real but addressable. First, the scoring proxy — mean-pooled cosine similarity for slice redundancy, token-level cosine residuals for localized deviations, and last-layer attention for anchor saliency — is never directly validated against what the projector and LLM actually need. The paper does not test whether dropped tokens would change predictions if the scoring were wrong. I agree with the stress-test note that this is the load-bearing assumption; the aggregate retention metrics could mask failures on rare or localized findings. But the paper's evaluation is broad enough that I don't think this is fatal — it is a missing experiment, not a known failure. Second, the hyperparameters (tau, r_nms, kappa, tau_m, and the heuristic multipliers in Eqs. 5 and 7) are not reported in the main text, and no error bars or significance tests are given. Third, code is not released, which makes reproducibility harder for a method whose value depends on exact implementation details.\n\nNone of these undermine the core result. The central claim — that anisotropic 2.5D allocation, not a larger budget, drives the quality-efficiency frontier — holds up under the evidence presented. The paper deserves a serious referee: the method is practical, the experiments are well designed, and the limitations are fixable. I would ask the authors to report all hyperparameters, release code, add error bars, and run a direct dropped-token sensitivity test that checks whether the scoring proxy agrees with downstream task importance. With those additions this would be a solid contribution to the medical MLLM efficiency literature.","headline":"CARVE is a legitimately useful, well-evaluated training-free token compression method for 3D medical MLLMs; the empirical case is strong, though the feature-space proxy for token importance is under-validated and hyperparameters are unreported.","tokens_in":14522,"tokens_out":1441,"would_cite":true,"duration_ms":14115,"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":"CARVE compresses 3D medical volumes to about one fifth of their visual tokens before the LLM, retaining 87% of full-token report-generation quality and near-full VQA accuracy.","keywords":["3D medical image understanding","visual token compression","multimodal large language models","slice-based MLLMs","budget-constrained allocation","training-free inference acceleration","2.5D representation","report generation"],"falsifier":"Take a random sample of AMOS-MM volumes at the r=0.2 budget and compare per-query answers from the full model against the model with CARVE's dropped tokens manually restored one at a time; if restoring tokens that CARVE assigns low residual scores changes answers on cases where all compared baselines are correct, the proxy scores are missing task-critical evidence. Alternatively, annotate the ground-truth finding regions on AMOS-MM and measure whether CARVE's retrieved tokens concentrate on those regions; if they do not on cross-slice relation questions, the residual objective is not tracking the evidence the benchmark rewards.","tokens_in":13462,"feed_emoji":"🩻","tokens_out":8368,"duration_ms":63458,"temperature":0.7,"pith_summary":"The paper asks whether a growing visual token budget is the right way to make slice-based 3D medical vision-language models better, and answers no: scaling analyses on AMOS-MM and M3D-VQA show accuracy saturates while latency, memory, and FLOPs keep rising, and in-plane resolution helps more than adding slices at matched budgets. It then presents CARVE, a training-free method that compresses the visual sequence before the frozen projector and LLM by casting token reduction as a budget-constrained 2.5D allocation problem. CARVE partitions the depth axis into coherent windows, spends a slice of the budget on cross-slice evidence retrieval from the full volume and the rest on multi-granular spatial anchors on representative slices, then folds surplus eligible tokens into nearby anchors. At roughly 20% retained tokens, CARVE attains 87.07% of full-token quality on AMOS-MM report generation and preserves 96.97% and 99.08% on 3D-RAD and M3D-VQA, leading all compared compression baselines.","feed_headline":"CARVE keeps 87% of report quality on 80% fewer tokens","feed_subtitle":"Training-free token reallocation tops compression baselines on AMOS-MM and preserves near-full VQA accuracy.","key_machinery":"The load-bearing object is the 2.5D budget-allocation pipeline, built from three scoring functions and two selection loops. Inter-slice change scores $g_t = 1 - \\cos(\\bar{f}_t, \\bar{f}_{t+1})$ define adaptive depth windows; token-level residuals $\\tilde{r}_j$, computed against aligned neighbours on adjacent slices with the slice median subtracted, isolate local deviations; the allocation splits the budget into $K_{\\text{inter}} = \\lfloor \\rho K \\rfloor$ for global retrieval and the rest for anchors, distributed over windows by a softmax over normalised window statistics. Intra-slice anchor selection uses a best-first quadtree whose refinement priority $\\pi(B)$ combines block heterogeneity with the averaged last-layer self-attention saliency $a_j$, placing anchors on representative slices. Global retrieval then ranks all non-anchor tokens by the volume-normalised residual, keeps the strongest quartile as candidates, and applies 3D non-maximum suppression with a spacing-aware Chebyshev radius in physical coordinates to enforce dispersion before the retrieved tokens are kept fixed. Finally, remaining eligible tokens are folded into their closest same-window anchor with a cosine-weighted mean, and the compressed sequence retains its 3D position indices.","core_discovery":"The central claim is that the redundant visual sequence produced by slice-wise encoding of a 3D volume is best compressed not by a flat importance ranking or by dropping whole slices, but by an anisotropic allocation that distinguishes two roles: a spatial scaffold of anchors that summarise recurring anatomy, and a dispersed set of retrieved tokens that capture locally changing findings on any slice. Because the depth axis repeatedly re-images aligned anatomy with sparse embedded changes, token importance is not isotropic in the volume, and a single flat ranking can waste budget on correlated tokens while missing local deviations. CARVE operationalises this by splitting the budget across depth windows and in-plane positions, and by keeping every slice eligible for retrieval rather than committing to a subset of planes. The paper reports that this allocation, rather than a larger budget, drives the quality-efficiency frontier: at a 20% target keep ratio CARVE uses fewer retained tokens than any baseline while scoring highest on every AMOS-MM report metric and near-full retention on VQA tasks.","pith_inferences":["If the paper's proxy scores were validated against the projector and LLM's actual evidence needs, a natural next test is whether the budget split $\\rho$ should adapt per question type; the paper uses a fixed $\\rho=0.25$, but report generation and closed-ended VQA may require different retrieval shares.","The scaling finding that in-plane resolution beats slice count at matched token budgets is stated for a specific 2D-encoder stack; a testable extension is whether the same ordering holds for volumetric encoders or lower-resolution whole-volume inputs, which would change acquisition and compression guidance.","CARVE's saliency signal is the frozen 2D encoder's last-layer attention; replacing it with a task-conditioned signal, such as the language model's cross-attention, is a plausible way to close the residual gap to full tokens, but would require the training-free property to be re-examined."],"forward_implications":["At a 20% target keep ratio, CARVE leads every compared compression method on all four AMOS-MM report-generation metrics, with an aggregate retention of 87.07% versus 80.89% for the strongest baseline and a range of 66.76-80.89% for the other baselines.","On 3D-RAD and M3D-VQA, CARVE preserves 96.97% and 99.08% of full-token aggregate performance at realized keep ratios of 19.7% and 19.2%, respectively, and is first on seven of the eight open-ended scores across the two benchmarks.","The gains transfer across backbones: CARVE achieves the highest aggregate retention on Hulu-Med-4B, Qwen3-VL-4B, and Qwen3-VL-8B, with the best mean rank (2.42) across three backbones and six evaluation tracks.","The ablation shows both selection branches are necessary: running intra-slice anchors alone drops 1.69 ROUGE-L points on AMOS-MM report generation, inter-slice retrieval alone drops 1.92 points, and dropping folding entirely costs 3.40 points.","Quality improves with budget for all methods, but the spread narrows as the budget grows: at the most binding ratio CARVE leads by 0.48 ACC and 1.76 ROUGE-L, while near the 0.4 ratio all methods approach full-token quality."],"supporting_citations":[{"why":"Supplies the Hulu-Med-7B backbone on which the main compression comparisons are run.","marker":"Jiang et al. 2025a"},{"why":"Supplies the AMOS-MM VQA and report-generation benchmark that drives the central evaluation.","marker":"Ji et al. 2022"},{"why":"Supplies the M3D-VQA benchmark used for one of the two scaling analyses and the VQA transfer evaluation.","marker":"Bai et al. 2024"},{"why":"Supplies the 3D-RAD benchmark for closed-ended and open-ended VQA evaluation.","marker":"Gai et al. 2026"},{"why":"Provides VisionZip, the image-token compression baseline compared against CARVE in the main tables.","marker":"Yang et al. 2025"},{"why":"Provides DivPrune, the diversity-based pruning baseline that is CARVE's strongest aggregate-retention competitor in the report-generation comparison.","marker":"Alvar et al. 2025"},{"why":"Provides MedPruner, the closest medical-volume compression baseline, which filters whole slices before token retention and motivates CARVE's retrieval over all slices.","marker":"Liu et al. 2026"},{"why":"Supplies the best-first quadtree and multi-granular merging mechanism used for CARVE's intra-slice anchor selection.","marker":"Hyun et al. 2025"}],"fun_headline_variants":["CARVE: 80% fewer tokens, 98% VQA retention via anisotropic allocation","CARVE reallocates tokens across slices, keeps 98% of full-token quality","Training-free CARVE cuts 80% tokens, preserves near-full VQA","CARVE's 2.5D token budget outperforms flat compression on AMOS-MM","Not all slices equal: CARVE allocates tokens by cross-slice evidence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that its three proxy scores—mean-pooled slice cosine similarity, per-token cross-slice residual, and last-layer self-attention—identify the evidence the projector and LLM actually need, even though the paper does not validate whether dropped tokens would have changed a prediction.","fun_headline_variants_meta":{"raw":{"variants":["CARVE: 80% fewer tokens, 98% VQA retention via anisotropic allocation","CARVE reallocates tokens across slices, keeps 98% of full-token quality","Training-free CARVE cuts 80% tokens, preserves near-full VQA","CARVE's 2.5D token budget outperforms flat compression on AMOS-MM","Not all slices equal: CARVE allocates tokens by cross-slice evidence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001133,"raw_usage":{"total_tokens":4757,"prompt_tokens":1047,"completion_tokens":3710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":3600}},"tokens_in":663,"tokens_out":3710,"duration_ms":23843,"temperature":1.0,"reasoning_tokens":3600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:37:45.462628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of AMOS-MM volumes at the r=0.2 budget and compare per-query answers from the full model against the model with CARVE's dropped tokens manually restored one at a time; if restoring tokens that CARVE assigns low residual scores changes answers on cases where all compared baselines are correct, the proxy scores are missing task-critical evidence. Alternatively, annotate the ground-truth finding regions on AMOS-MM and measure whether CARVE's retrieved tokens concentrate on those regions; if they do not on cross-slice relation questions, the residual objective is not tracking the evidence the benchmark rewards.","supporting_citations":[],"review_version":2}