{"id":"e062bdc3-6a02-4c59-86e3-101a1474d8b6","arxiv_id":"2603.05950","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"E-AdaPrune sets each image's visual-token budget by the singular-value energy spectrum of its features, improving VLM accuracy under matched average token budgets.","lead":"E-AdaPrune uses the singular-value energy of an image's visual features to decide how many tokens that image needs, giving simple images fewer and complex images more. It works with existing token selectors and claims small average accuracy gains at matched compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy-preservation claim is unverified for the actually selected token subset; the budget k* is computed from full-matrix SVD but attention top-k selection may not preserve that energy.","rationale":"The reader's weakest assumption correctly identifies the gap between the SVD-derived budget and the actual token subset selected by the pruning heuristics. This is the core conceptual link of the paper: the title, abstract, and method section all claim that E-AdaPrune preserves a certain proportion of spectral energy, but there is no verification that the final pruned token set satisfies this property. Without this verification, the mechanism is only an untested proxy. The empirical results (e.g., +5.1% on MMVet) could still be valid and useful, but the paper's central claim is overstated. The proposed concrete test (measuring the retained energy of the selected tokens) would settle whether the mechanism is real or whether the gains arise indirectly from adaptive budgeting. Since this is a conditional concern—the method may still work even if energy preservation fails—the verdict remains CONDITIONAL, matching the reader's assessment. There is no need to change the verdict; the identified gap is already the weakest assumption, and a targeted experiment would either support or refute it.","tokens_in":13638,"tokens_out":4369,"duration_ms":44145,"concrete_test":"Run FastV+E on a random sample of 100 images from MMVet with tau=99.8%. After pruning, record the indices of the retained visual tokens. Compute the spectral energy ratio of the retained token subset to the full feature matrix, using the same feature representation as the budget estimation (Z_V) and also the layer-2 features where FastV prunes. If the average ratio is significantly below 0.998 (e.g., <0.95), the energy-preservation claim is not supported. Additionally, compare this ratio to that of static FastV with the mean token budget; if the adaptive method does not preserve more energy than static, the observed performance gains cannot be attributed to energy preservation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of E-AdaPrune (Section 3.2, Eq. 6) is that computing k* as the minimal rank capturing a tau fraction of the total spectral energy of Z_V yields a budget that 'preserves spectral energy' after pruning. However, the downstream token selectors (FastV, PyramidDrop, VisionZip) choose tokens by importance heuristics (e.g., attention scores), not by the principal components from the SVD. There is no analysis or experiment showing that the token subset actually retained by these selectors contains the promised fraction of the original spectral energy. In fact, because ignored singular components correspond to low-variance directions across the whole feature matrix, while attention top-k picks individual tokens, the retained set can easily have much lower energy. Even if the budget k* is a reasonable heuristic for content complexity, the paper's stated justification—preserving a fixed energy fraction—remains unsubstantiated. This gap matters because it is the load-bearing conceptual link between the SVD and the pruning outcome; without it, the method is just an adaptive budget selector with an unvalidated proxy, and the 'energy-driven' framing is inaccurate. The empirical gains, while real, may stem from a different mechanism (e.g., allocating more tokens to images with flatter spectra, which happen to be harder), not from actual energy preservation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"E-AdaPrune proposes a training-free, plug-and-play module that replaces a static visual token budget in VLM token-pruning pipelines with an image-specific budget k*, computed as the minimum rank at which the cumulative squared singular values of the visual feature matrix reach a threshold τ. This budget is then passed to existing selectors (FastV, PyramidDrop, VisionZip). The paper claims that, under matched average token budgets, this adaptive allocation improves average accuracy by up to 0.6%, with a +5.1% relative gain on MMVet, while adding only 8 ms per image via randomized SVD. Experiments are reported on three LLaVA backbones and nine benchmarks, although the abstract additionally claims Qwen2.5-VL-3B results that do not appear in the body.","tokens_in":13982,"tokens_out":17070,"duration_ms":143586,"significance":"If the claims hold, the contribution is a simple and attractive addition to the visual-token-pruning toolbox: it is training-free, introduces no learnable parameters, and can be dropped into existing pruning pipelines by replacing a static budget with a per-image budget. The matched-average-budget evaluation protocol is a strength, and the rSVD latency analysis is a useful engineering result. The central weakness is that the paper's stated justification—that the selected token set preserves a fixed fraction of spectral energy—is never verified; the selectors do not retain the singular components on which k* is based. In addition, the abstract overstates the experimental scope by promising Qwen2.5-VL-3B results that are absent. With the requested revisions, the core idea could be publishable as an adaptive budget heuristic rather than as a principled energy-preserving method. I agree with the stress-test assessment: the circularity concern does not land, because k* is defined from the feature matrix alone, but the energy-preservation mechanism is unsupported.","major_comments":[{"comment":"The central claim that k* 'preserves spectral energy' is not established. k* is the minimal rank of the full visual feature matrix Z_V whose cumulative squared singular values reach τ, but the selectors used here do not retain the top singular components: FastV chooses attention-top-k tokens after layer 2, PDrop drops layer-wise by importance scores, and VisionZip merges tokens by similarity. The retained token subset can therefore have an energy fraction substantially different from τ, and the paper never measures this quantity. A simple check is to report ||Z_retained||_F^2 / ||Z_V||_F^2 for the actually selected tokens. Table 4 also provides internal evidence against the energy link: with q=0 the average budget changes from 130 to 156 tokens but MMB is essentially unchanged (62.8 vs 62.5). Please either verify retained-energy preservation or reframe k* as a heuristic budget, ideally w","section":"§3.2, Eq. (6); §4.1"},{"comment":"The abstract states that E-AdaPrune is evaluated 'across four VLM backbones' and gives concrete Qwen2.5-VL-3B numbers for SQA^I (35.8% fewer tokens, 0.52% relative decrease, 1.94% improvement). The experiments in the body are limited to LLaVA-1.5-7B, LLaVA-1.5-13B, and LLaVA-NeXT-8B; no Qwen experiment appears. This is a factual overstatement of the reported scope and should be corrected by adding the Qwen experiments or removing the claims.","section":"Abstract; §4.1; §4.3"},{"comment":"Many of the core improvements are within the range of a few tenths of a percentage point, and no significance testing, confidence intervals, or repeated runs are reported. At τ=99.0% the paper itself reports average improvements of only 0.3%, 0.1%, and 0.1% for the three baselines. The phrase 'consistently yields' is stronger than the evidence supports. The tables also mix raw scores and relative-to-base percentages without a clear mapping, making it difficult to verify the stated averages. Please add statistical support and clarify the reporting scheme.","section":"§4.2; Tables 1-2"},{"comment":"The method depends on τ, kmin/kmax, and the PDrop ρ clamp intervals ([0.4,0.6] at τ=99.8%, [0.6,0.8] at τ=99.0%). None of these are justified by sensitivity analysis, and kmin/kmax are not reported. Because the static baseline budget is the mean of the adaptive budgets, these choices directly determine the comparison operating point. A sensitivity table over τ and the clamp ranges is needed to show that the reported gains are not artifacts of calibration.","section":"§4.1; Algorithm 1"}],"minor_comments":[{"comment":"The text says 'nine benchmarks' but enumerates only eight; Table 1 actually contains nine columns (MMB and MMBCN are counted separately). Please align the text and the table.","section":"§4.1"},{"comment":"Figures 2 and 3 appear to contain the same diagram with different captions. If this is not a layout artifact, one of the figures should be replaced.","section":"Figures 2 and 3"},{"comment":"Algorithm 2 introduces p, k_sub, and a target dimension without clearly connecting them to the t and q used in §4.4. The notation should be made consistent for reproducibility.","section":"Appendix, Algorithm 2"},{"comment":"The paper positions E-AdaPrune against learned adaptive methods (ATP-LLaVA, VCM, VisionThink) but provides no experimental comparison with any of them. Even a single reference comparison would help clarify the trade-offs.","section":"§2.3; §4"},{"comment":"The abstract's 'up to 0.6%' and '+5.1% relative gain' mix absolute and relative performance changes. The paper should state clearly which quantities are relative to Base and which are relative to the static baseline.","section":"Abstract; Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a usable core idea and a fair matched-average-budget protocol, but it is not publishable as is. The two blocking issues are (1) the unsupported energy-preservation mechanism and (2) the abstract's claim of Qwen2.5-VL-3B results with no corresponding experiments. Both are fixable with additional experiments or reframing, so I recommend a major revision rather than rejection. The small effect sizes also need to be addressed with statistical support or more cautious language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a simple, training-free trick that works about as well as advertised on the LLaVA models actually tested. The abstract, however, promises more than the paper delivers, and the 'energy preservation' framing is not backed by an analysis of the tokens actually retained.\n\nThe new bit is swapping the fixed top-k budget in FastV, PyramidDrop, and VisionZip for a per-image budget k* defined by the cumulative squared-singular-value energy of Z_V. That's a clean, parameter-light idea, and the matched-average-budget comparison is the right way to isolate the benefit. The headline results — up to +0.6% average, and +5.1% relative on MMVet for PDrop — are plausible and worth a serious look. The rSVD overhead table is honest: 8ms per image and ~5% total runtime is a fair price.\n\nWhere it gets soft. First, the abstract is not the paper. It mentions four backbones including Qwen2.5-VL-3B and reports a specific SQA^I gain for it; the body only has LLaVA-1.5-7B, LLaVA-1.5-13B, and LLaVA-NeXT-8B. That is a serious reporting inconsistency that has to be fixed before this is publishable. Second, the 'preserving spectral energy' story is not verified at the level of the actually selected token set. k* is computed from the full-matrix SVD, but the selectors keep tokens by attention or similarity, not by principal components. The paper never checks that the kept tokens contain the promised energy fraction. It may still work as a content-complexity proxy, but the paper should either add that analysis or soften the claim. Third, there are several hand-tuned knobs (τ, kmin/kmax, PDrop ρ-clamps, rSVD t and q) and no sensitivity analysis or error bars. Many of the deltas are 0.1–0.3%, which is within the noise for these benchmarks; the MMVet gain is the one standout.\n\nWho benefits: anyone working on VLM inference acceleration who wants a plug-and-play adaptive budget for an existing selector. This is a modest but useful contribution, not a branch-level rethinking. It deserves a serious referee, not a desk reject, but the authors need to reconcile the abstract with the body and either verify or reframe the energy-preservation claim. I'd send it to review with a clear list of required revisions.","headline":"Simple adaptive-budget trick with real but modest gains; the abstract overclaims (Qwen2.5 missing) and the energy-preservation story is unverified.","tokens_in":14434,"tokens_out":6399,"would_cite":true,"duration_ms":53795,"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":"Spectral energy of visual features can set per-image token budgets for VLMs, improving accuracy at matched compute.","keywords":["vision-language models","token pruning","adaptive token budget","singular value decomposition","spectral energy","inference acceleration","training-free"],"falsifier":"Compute, on a benchmark set, the squared spectral energy contained in the actually selected top-k tokens (according to the integrated selector) for each image and compare it to the τ threshold that determined k*. If the retained energy falls far below τ for many images, or if an image with a flat spectrum (high k*) answers correctly with far fewer tokens, the proxy is not doing the causal work the paper claims.","tokens_in":13533,"feed_emoji":"⚡","tokens_out":3769,"duration_ms":33069,"temperature":0.7,"pith_summary":"E-AdaPrune proposes that the right number of visual tokens for a vision-language model is an intrinsic property of the image, measurable from the singular value spectrum of the visual feature matrix. It computes an image-specific token budget by keeping only enough singular components to retain a set fraction (τ) of the total squared spectral energy, then feeds that budget to any existing top-k token selector. Under matched average token budgets across nine benchmarks and three model families, this adaptive budgeting improves average accuracy by up to 0.6%, with the largest relative gain (+5.1%) on MMVet reasoning. The method is training-free and adds roughly 8 ms per image when budget estimation uses randomized SVD.","feed_headline":"Per-image token budgets from SVD spectrum boost VLM accuracy","feed_subtitle":"Training-free adaptive pruning beats static top-k at matched compute across nine benchmarks","key_machinery":"The central object is the singular value decomposition (SVD) of the visual feature matrix Z_V ∈ R^{n_v × d_v}. The method defines 'image energy' as the sum of squared singular values and selects the minimum number of components k* that capture a fraction τ of that energy (Eq. 6), clamped to a practical range [k_min, k_max]. This energy-based rank acts as a content-aware token budget that is orthogonal to—and can be combined with—any existing token-scoring heuristic (FastV, PyramidDrop, VisionZip). Randomized SVD (rSVD) is used to estimate the spectrum cheaply, adding about 8 ms per image.","core_discovery":"The paper's central claim is that visual token budgets should be adaptive, and that the correct budget per image is encoded in the spectral energy distribution of the visual feature matrix. By computing the smallest rank k* whose cumulative squared singular values reach a threshold τ of total energy, E-AdaPrune allocates more tokens to information-dense scenes (flat spectra) and fewer to redundant scenes (steep spectral decay). This k* is then used as the cut-off for existing selectors such as attention-based top-k or merging methods, decoupling budget determination from token selection. The paper reports consistent average gains up to 0.6% over static baselines at matched mean token budgets","pith_inferences":["The abstract and full text disagree on scope: the abstract reports a fourth backbone (Qwen2.5-VL-3B) and a specific SQA^I result (35.8% fewer tokens, 52.5% more tokens on hard cases) that do not appear in the body; readers should treat the body as authoritative for what was actually run.","The validity of the energy proxy itself is not directly verified: the paper assumes the top-k tokens selected by the downstream selector retain the spectral energy that justified k*, but never measures the energy contained in the retained token set.","A natural extension is to turn τ into a tunable per-query or per-task knob rather than a fixed global hyperparameter, since the paper's results show gains shrink at lower τ (99.0% vs 99.8%).","The spectral-energy idea may transfer to other sequence-compression settings (e.g., video frames or long documents), but the claim of being text-agnostic has not been stress-tested beyond the VLM benchmarks reported."],"forward_implications":["If the spectral-energy budget is a valid proxy for what a VLM needs, static top-k pruning is suboptimal: simple images can be pruned more aggressively and the saved compute can be redirected to dense images.","Because E-AdaPrune is training-free and model-agnostic, it can be dropped into existing acceleration pipelines without altering their scoring mechanisms, potentially combining with many future selectors.","The energy-threshold τ acts as a global information-preservation guarantee across heterogeneous datasets, which may make adaptive pruning more predictable across benchmarks.","The measured +5.1% relative MMVet gain suggests that reasoning-heavy, information-dense scenes are precisely where static budgets discard critical tokens, so adaptive budgeting may help high-level VQA tasks specifically."],"fun_headline_variants":["SVD spectrum sets per-image token budgets for VLMs","Adaptive token pruning from singular value energy","Spectral energy guides VLM token allocation","Per-image token budgets via SVD energy","Energy-adaptive pruning lifts VLM accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The number of singular-value components needed to reach a τ fraction of the total squared spectral energy of the whole visual feature matrix is a valid proxy for the number of visual tokens the VLM actually needs, and this energy is actually preserved after the downstream token selector picks its top tokens—two assumptions the paper does not directly verify.","fun_headline_variants_meta":{"raw":{"variants":["SVD spectrum sets per-image token budgets for VLMs","Adaptive token pruning from singular value energy","Spectral energy guides VLM token allocation","Per-image token budgets via SVD energy","Energy-adaptive pruning lifts VLM accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1381,"prompt_tokens":768,"completion_tokens":613,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":544}},"tokens_in":512,"tokens_out":613,"duration_ms":5118,"temperature":1.0,"reasoning_tokens":544,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:40:20.013611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on a benchmark set, the squared spectral energy contained in the actually selected top-k tokens (according to the integrated selector) for each image and compare it to the τ threshold that determined k*. If the retained energy falls far below τ for many images, or if an image with a flat spectrum (high k*) answers correctly with far fewer tokens, the proxy is not doing the causal work the paper claims.","supporting_citations":[],"review_version":1}