{"id":"b0dcde65-485e-4553-b226-d041da7411fe","arxiv_id":"2504.16003","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MVQA combines a Mamba encoder with a masked fusion of high-resolution fragments and low-resolution semantic content, reaching near-state-of-the-art VQA accuracy at lower compute than FAST-VQA.","lead":"This paper introduces MVQA, a video quality assessment model built on Mamba state-space layers, together with a sampling method that mixes high-resolution detail patches with low-resolution semantic context. Tests on standard benchmarks show it matches or slightly beats existing methods while running about twice as fast as FAST-VQA and using less GPU memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars or multiple-seed results: the reported SROCC gains over baselines are within typical noise for KoNViD/LIVE-VQC/YouTube-UGC, so the central accuracy advantage is not statistically established.","rationale":"I read the paper as a competent empirical VQA study with a sensible architecture and a genuinely interesting sampling idea. The USDS ablation in Table 6 and the cross-architecture validation in Table 5 provide real evidence that the fusion mechanism helps over Fragments in both Mamba and Video Swin backbones, and the efficiency numbers in Table 4 are internally consistent because the fused tensor has the same size as the fragment-only tensor. I therefore do not see a fatal flaw that would justify rejection. The concern I find most load-bearing is not the hand-picked 3:1 USDS ratio (a real limitation, but one that does not invalidate the accuracy claim since the proposed configuration is empirically better than alternatives it was compared against). The deeper issue is that the central 'accuracy at or above existing methods' claim rests on SROCC differences that are small relative to the noise level of the datasets used. For KoNViD-1k, LIVE-VQC, and YouTube-UGC, the approximate standard error of Spearman correlation is 0.03-0.04, and most of the reported advantages over the best baselines are smaller than that. Without confidence intervals, multiple seeds, or significance tests, the reader cannot tell whether the headline improvements are real or run-to-run variation. This issue is fixable and does not by itself overturn the conditional verdict, but it is the condition that should be satisfied before the accuracy claim is accepted. The reader identified the USDS ratio as the weakest assumption and also listed the missing error bars as a secondary weakness, so my emphasis on statistical reliability is a partial agreement: I agree the design choice is under-ablated, but I think the accuracy evidence is the more critical load-bearing point. The concrete test I propose would settle whether the accuracy advantage survives uncertainty quantification, which is the single most important check for the paper's central claim.","tokens_in":17074,"tokens_out":11741,"duration_ms":109349,"concrete_test":"Run the fine-tuning evaluation with 5 random seeds and report paired bootstrap 95% confidence intervals for the SROCC difference between MVQA and the best baseline (e.g., CLiF-VQA) on KoNViD-1k, LIVE-VQC, and YouTube-UGC, using the same fixed train/test splits as the published baselines. If the confidence interval for the difference includes zero on at least two of the three datasets, the central accuracy claim is not supported by the data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: accuracy and efficiency. The efficiency part is structurally sound: the USDS fused tensor is the same size as the fragment-only tensor, and Table 4 shows MVQA-tiny at 34 GFLOPs and 0.028s versus FasterVQA at 70 GFLOPs and 0.032s. The load-bearing part is the accuracy claim, and it is not statistically supported. In Table 3, MVQA-middle improves over the best multi-branch baseline by +0.022 SROCC on KoNViD-1k (N=1200, approximate Spearman standard error ~0.029), +0.012 on LIVE-VQC (N=585, SE~0.041), and +0.007 on YouTube-UGC (N=1067, SE~0.031); none of these individually exceeds one standard error. MVQA-tiny's average advantage over FasterVQA is 0.007, also within noise. The paper reports no confidence intervals, no multiple training seeds, and no significance tests, and it does not specify the train/test split used for fine-tuning in Section 4.1. If the true SROCC of MVQA is within noise of the baselines, the 'accuracy at or above existing methods' claim fails even though the efficiency numbers are correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MVQA, a no-reference video quality assessment (VQA) network built entirely from Mamba state-space layers, together with a new sampling method called Unified Semantic and Distortion Sampling (USDS). USDS fuses grid-sampled high-resolution fragments with a downsampled full-frame view via a fixed binary mask, so the fused tensor has the same size as the fragment-only tensor and adds no computational overhead. The authors pre-train MVQA on LSVQ and evaluate on LSVQ, KoNViD-1k, LIVE-VQC, and YouTube-UGC, reporting that MVQA-middle achieves SROCC at or above existing single-branch and multi-branch methods, while MVQA-tiny runs roughly 2x faster than FAST-VQA and uses much less GPU memory. The paper also reports ablations showing USDS improves over fragment sampling in both the proposed MVQA backbone and a Video Swin Transformer.","tokens_in":17243,"tokens_out":8228,"duration_ms":70833,"significance":"If the empirical results are correct, the paper makes a useful contribution by demonstrating that a pure state-space backbone is viable for VQA and by providing a simple, structurally efficient way to inject global semantic context into fragment-based sampling. The efficiency advantage of the tiny variant is structurally guaranteed because USDS does not increase the input tensor size, and the cross-architecture validation in Table 5 is a valuable consistency check. The paper does not ship code, but the method is described in enough detail to reimplement once the notation and hyperparameter omissions are fixed. The main weakness is statistical: the reported accuracy gains over strong baselines are not accompanied by error bars, multiple seeds, or significance tests, and some gains are within the typical Spearman standard error for the dataset sizes involved. The lack of an ablation of the hand-chosen 3:1 fusion ratio also leaves a gap between the proposed design and the evidence for it.","major_comments":[{"comment":"The central accuracy claim is not statistically supported. The paper reports single-run SROCC and PLCC values without error bars, multiple training seeds, or significance tests. For example, on KoNViD-1k (N=1200) the approximate Spearman standard error is about 0.029, so the reported +0.022 SROCC gain of MVQA-middle over the best multi-branch baseline (0.925 vs 0.903) is within one standard error; the gains on LIVE-VQC (N=585, SE approx 0.041) and YouTube-UGC (N=1067, SE approx 0.031) are likewise within one standard error. Please provide repeated-run statistics (at least mean and standard deviation over several seeds) or a statistical comparison against the strongest baselines before claiming state-of-the-art accuracy.","section":"Section 4.2, Tables 1 and 3"},{"comment":"The fine-tuning protocol is underspecified. The text states that the model is fine-tuned on KoNViD-1k, LIVE-VQC, and YouTube-UGC but does not give the train/test split ratio, the number of epochs, or the learning-rate schedule used for fine-tuning, and it does not state whether the same splits as those used by the cited baselines were employed. Since Table 3 computes 'improvements' against published numbers, differences in the evaluation split can change the ranking; please specify the exact protocol and, if possible, reuse the standard publicly available splits used in prior work.","section":"Section 4.1 and Table 3"},{"comment":"The USDS fusion ratio is a hand-chosen design parameter that is never ablated. The mask in Eq. (11) places high-resolution fragments in three quadrants of each 32x32 block and a low-resolution full-frame patch in the bottom-right quadrant (a 3:1 blend), and this specific ratio and mask placement are not varied in any experiment. The ablation in Table 6 compares USDS as a whole against resize, crop, MRET, and fragments, but does not isolate the effect of the fusion ratio or mask location (e.g., other quadrants, 2:2, or 1:3 blends). Without such an ablation, it is unclear whether the reported gains come from the specific fusion design or merely from adding any global low-resolution context; please add experiments varying the ratio and mask placement.","section":"Section 3.3, Eq. (10)-(11) and Algorithm 1"}],"minor_comments":[{"comment":"Eq. (13) appears to contain a typo: 'T_l = Vim(T_{l-1} + T_{l-1})' would double the input to the vision Mamba block. Presumably the intended residual form is T_l = Vim(T_{l-1}) + T_{l-1}; please correct this.","section":"Eq. (13)"},{"comment":"The dimension broadcasting in Eq. (12) is unclear: a spatial embedding of length (H/16 * W/16 + 1) and a temporal embedding of length t are added to a token sequence of length T * (H/16 * W/16) + 1. Please explain how these embeddings are broadcast across the T frames.","section":"Eq. (12)"},{"comment":"Please state whether the reported FLOPs and runtimes include the USDS sampling and preprocessing steps or only the backbone forward pass. If the preprocessing is excluded, the end-to-end speed advantage may be smaller than the '2x faster' claim suggests.","section":"Section 4.3 and Table 4"},{"comment":"The '1/5 GPU memory' claim should be supported by a measurement table or numeric annotations giving the measurement conditions (batch size, input resolution, and whether the number is peak activation memory or total training memory). Fig. 1 currently shows panels without numeric memory values.","section":"Abstract and Fig. 1"},{"comment":"The claim 'this is the first work to successfully apply state-space models in the VQA domain' is very strong. Given the rapid growth of SSM-based vision methods, please verify against concurrent literature and either provide a thorough literature search or soften the claim to 'to the best of our knowledge'.","section":"Abstract and Section 1"},{"comment":"The CLIP-based semantic analysis is purely qualitative. Adding a quantitative measure (e.g., average CLIP cosine similarity between the sampled frames and the original frames) would make the semantic-preservation claim more concrete.","section":"Section 4.4, Figs. 2 and 6"},{"comment":"The resolution ablation is presented only as plots without numerical values. Please include the actual SROCC/PLCC numbers in the text or in a table so that the magnitude of the resolution-dependent improvement can be assessed.","section":"Section 4.5, Fig. 7"},{"comment":"The notation in Eq. (10) and Algorithm 1 is inconsistent: the block indices are introduced as (k', l') in Eq. (10) but as (i, j) in Algorithm 1, and R_{k,l} uses undefined indices. Please unify the notation.","section":"Algorithm 1 and Eq. (10)"},{"comment":"Please report the numerical values of the loss weights alpha and beta in Eq. (17) and the number of frames T after sampling (Table 2 suggests T=32), as these are needed for reproducibility.","section":"Section 4.1 and Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important and timely topic, and the efficiency mechanism is elegant. However, the accuracy claim, which is the load-bearing part, needs statistical support (seeds, confidence intervals, or significance tests), and the fine-tuning protocol must be specified for the results to be reproducible and comparable. The 'first work' novelty claim is risky because there is active concurrent work on SSM-based VQA; the authors should temper it. If the authors can provide the missing statistical evidence and ablations, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the USDS sampling idea is the real contribution—masked fusion of high-res fragments with a low-res semantic view is neat, and the cross-architecture test in Video Swin shows it transfers beyond the Mamba backbone. The efficiency side is solid too: Table 4 shows MVQA-tiny at 34 GFLOPs and 0.028s, roughly half of FasterVQA's time with a fraction of the FLOPs. That's genuinely useful.\n\nThe soft spot is the accuracy claim. In Table 3, MVQA-middle beats the best multi-branch baseline by +0.022 SROCC on KoNViD, +0.012 on LIVE-VQC, and +0.007 on YouTube-UGC. With dataset sizes of 1200, 585, and 1067, those gaps are within one standard error. No confidence intervals, no multiple seeds, no significance tests. The abstract hedges with 'comparable performance,' which is closer to what the data support. The central 'accuracy at or above existing methods' claim is not statistically established.\n\nAlso, the 'first to apply SSM to VQA' claim is not substantiated—no prior Mamba VQA work is cited, and I suspect some exists. The fusion ratio (3:1) is never ablated, and the loss weights in Eq. (17) are unreported. Eq. (13) contains a typo in the residual connection—it reads Tl = Vim(Tl−1 + Tl−1), likely should be Tl = Vim(Tl−1) + Tl−1. Minor, but sloppy.\n\nNone of this suggests a hollow paper. The sampling method is worth considering, the efficiency win is structural, and the cross-architecture validation is a good practice. It just needs a more honest framing and basic statistical care before the accuracy advantage is believable.\n\nWho is this for? Anyone working on efficient VQA or sampling-based video models. It deserves a serious referee, but the referee should push for error bars, an ablation of the fusion ratio, and a citation check.","headline":"Useful sampling idea with real efficiency gains, but the accuracy advantage is within noise and the 'first Mamba VQA' claim is overreaching.","tokens_in":17910,"tokens_out":2480,"would_cite":false,"duration_ms":23083,"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":"MVQA claims that a state-space (Mamba) encoder with mask-fused sampling matches or beats existing VQA methods while running about twice as fast and using a fraction of the GPU memory.","keywords":["video quality assessment","no-reference video quality assessment","state space models","Mamba","unified sampling","semantic information","fragment sampling","linear complexity"],"falsifier":"Retrain MVQA-tiny on LSVQ with the USDS mask inverted so the low-resolution semantic view occupies the top-left quadrant of each block instead of the bottom-right, keeping everything else identical, and compare SROCC on LSVQ test; if the gain over Fragments disappears or reverses, the reported accuracy advantage is tied to the specific mask placement rather than to unified semantic-and-distortion sampling.","tokens_in":16764,"feed_emoji":"🎬","tokens_out":9696,"duration_ms":82787,"temperature":0.7,"pith_summary":"MVQA sets out to show that a video quality assessment network can be built entirely from Mamba/state-space layers, without CNN or Transformer backbones, and still match or beat existing methods. Its sampling module, USDS, fuses two views of each frame into one tensor: high-resolution grid fragments that preserve distortion, and a downsampled full-frame version that preserves semantic content, placed in the bottom-right quadrant of every 32x32 block. Because the fused tensor is the same size as the fragment-only input, the semantic information costs no extra compute. Reported results put MVQA-middle's average SROCC at 0.901 after fine-tuning, above the best single-branch (0.867) and multi-branch (0.886) comparators, with MVQA-tiny at 34 GFLOPs and 0.028 s. If correct, this means efficient VQA does not need a semantic side branch or expensive backbones.","feed_headline":"Mamba-based video quality scorer beats multi-branch rivals at 2x speed","feed_subtitle":"Mask-fused sampling gives one Mamba branch both distortion and semantics, beating multi-branch accuracy at lower cost.","key_machinery":"The central mechanism is USDS's fixed mask fusion. A binary mask M marks the bottom-right 16x16 quadrant of every 32x32 block; the low-resolution frame is expanded into a canvas E and written into those quadrants, while the high-resolution fragments fill the other three quadrants, giving V_hat = E·M + V_tilde·(1−M). This keeps the fused tensor at the fragment-only resolution, so semantic content is added at zero marginal compute. The second piece is the Vision Mamba encoder, a bidirectional selective state-space encoder whose linear complexity in sequence length, combined with spatial and temporal position embeddings and a regression token, produces the final quality score.","core_discovery":"The paper claims to be the first to apply state-space models successfully to VQA, and argues that prior efficient sampling fails because fragment-based sampling discards semantics while resize/crop discard distortion. USDS fixes this by sampling high-resolution fragments for distortion and a low-resolution full-frame copy for semantics, then fusing them with a binary mask that reserves the bottom-right quadrant of each block for the low-resolution view. The MVQA encoder is a bidirectional Vision Mamba with spatial-first, temporal-next scanning and a regression token whose output is mapped by an MLP to a quality score. On LSVQ and three small datasets, MVQA-middle reports SROCC 0.898 on LSVQ test and 0.901 average after fine-tuning, while MVQA-tiny reports 34 GFLOPs and 0.028 s per 150-frame clip, roughly twice as fast as FAST-VQA. The paper takes these results to show that a single state-space branch can carry both technical and semantic quality.","pith_inferences":["The paper leaves untested whether the 3:1 mask ratio is optimal; rotating the mask or varying the semantic fraction could show that the accuracy gain is specific to the bottom-right placement.","One extension the paper does not explore is making the semantic quadrant content-aware, for example weighting the low-resolution view by a CLIP affinity score, which could matter for videos where semantics dominate quality.","Because Mamba scales linearly with sequence length, a natural next test is longer clips than the 32-frame input; the efficiency advantage would likely grow, but accuracy at long temporal horizons is unmeasured.","The same mask-fusion idea might transfer to other quality-related tasks beyond VQA, such as image aesthetic scoring, since it provides a cheap global-context channel alongside local detail."],"forward_implications":["If the reported numbers hold, a single SSM branch can replace the multi-branch design that current top VQA models use, where a technical quality branch is paired with a separate semantic or aesthetic branch.","USDS is architecture-independent: the paper's cross-architecture test shows Video Swin Transformer also improves with USDS over Fragments, so the sampling scheme could be dropped into other VQA backbones.","Because Mamba's complexity is linear in sequence length, the efficiency gap should widen for longer and higher-resolution videos, making the approach more attractive for 2K/4K streaming quality control.","MVQA-tiny's fixed 34 GFLOPs and 0.028 s per 150-frame clip, independent of input resolution, put real-time or batch VQA screening on a single GPU within reach."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that forms MVQA's backbone and provides linear-complexity sequence modeling.","marker":"[11]"},{"why":"FAST-VQA's fragment sampling (GMS) is the distortion sampling baseline that USDS extends and the main accuracy and efficiency comparison.","marker":"[70]"},{"why":"Vision Mamba's bidirectional scanning block is the encoder architecture that MVQA adopts.","marker":"[90]"},{"why":"VideoMamba contributes the spatial-first, temporal-next scan order and positional embedding setup used by MVQA.","marker":"[35]"},{"why":"FasterVQA is the strongest single-branch efficient baseline in the efficiency and accuracy comparisons.","marker":"[71]"},{"why":"DOVER is the multi-branch aesthetic and technical baseline that MVQA is compared against on accuracy.","marker":"[76]"},{"why":"CLiF-VQA is a semantic-branch VQA baseline with the strongest average fine-tuned SROCC among multi-branch comparators and motivates USDS's semantic channel.","marker":"[48]"},{"why":"MBVQA is the multi-branch baseline with the highest LSVQ-test SROCC among comparators and a key accuracy target in Tables 1 and 3.","marker":"[69]"},{"why":"MRET's multi-resolution sampling is one of the four sampling baselines in the ablation study that USDS must beat.","marker":"[29]"},{"why":"Patch-VQ/LSVQ provides the pretraining corpus and the intra-dataset test sets used in the main experiments.","marker":"[83]"}],"fun_headline_variants":["Mamba VQA: 2x speed, 1/5 memory, same accuracy","First Mamba VQA: 2x faster, same accuracy","Mask-fused sampling gives Mamba VQA speed and fidelity","Unified sampling for Mamba VQA: fast, light, accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a fixed 3:1 blend, three quadrants of high-resolution fragments plus one bottom-right quadrant of downsampled full-frame content, captures both distortion and semantics well enough at every resolution, since that ratio and mask position are hand-chosen and never ablated.","fun_headline_variants_meta":{"raw":{"variants":["Mamba VQA: 2x speed, 1/5 memory, same accuracy","First Mamba VQA: 2x faster, same accuracy","Mask-fused sampling gives Mamba VQA speed and fidelity","Unified sampling for Mamba VQA: fast, light, accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3226,"prompt_tokens":1001,"completion_tokens":2225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":617,"tokens_out":2225,"duration_ms":16304,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:12:38.877022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MVQA-tiny on LSVQ with the USDS mask inverted so the low-resolution semantic view occupies the top-left quadrant of each block instead of the bottom-right, keeping everything else identical, and compare SROCC on LSVQ test; if the gain over Fragments disappears or reverses, the reported accuracy advantage is tied to the specific mask placement rather than to unified semantic-and-distortion sampling.","supporting_citations":[{"cited_title":"Fast- vqa: Efficient end-to-end video quality assessment with frag- ment sampling","cited_arxiv_id":null,"evidence_quote":"FAST-VQA's fragment sampling (GMS) is the distortion sampling baseline that USDS extends and the main accuracy and efficiency comparison."},{"cited_title":"Videomamba: State space model for efficient video understanding","cited_arxiv_id":null,"evidence_quote":"VideoMamba contributes the spatial-first, temporal-next scan order and positional embedding setup used by MVQA."},{"cited_title":"Neigh- bourhood representative sampling for efficient end-to-end video quality assessment","cited_arxiv_id":null,"evidence_quote":"FasterVQA is the strongest single-branch efficient baseline in the efficiency and accuracy comparisons."},{"cited_title":"Exploring video quality assessment on user generated contents from aesthetic and technical perspectives","cited_arxiv_id":null,"evidence_quote":"DOVER is the multi-branch aesthetic and technical baseline that MVQA is compared against on accuracy."},{"cited_title":"CLiF-VQA: Enhancing video quality assess- ment by incorporating high-level semantic information re- lated to human feelings","cited_arxiv_id":null,"evidence_quote":"CLiF-VQA is a semantic-branch VQA baseline with the strongest average fine-tuned SROCC among multi-branch comparators and motivates USDS's semantic channel."},{"cited_title":"Modular blind video quality assess- ment","cited_arxiv_id":null,"evidence_quote":"MBVQA is the multi-branch baseline with the highest LSVQ-test SROCC among comparators and a key accuracy target in Tables 1 and 3."},{"cited_title":"Mret: Multi-resolution transformer for video quality assessment","cited_arxiv_id":null,"evidence_quote":"MRET's multi-resolution sampling is one of the four sampling baselines in the ablation study that USDS must beat."},{"cited_title":"Patch-vq:’patching up’the video quality problem","cited_arxiv_id":null,"evidence_quote":"Patch-VQ/LSVQ provides the pretraining corpus and the intra-dataset test sets used in the main experiments."}],"review_version":1}