{"id":"2266dc3b-6303-4100-953e-3c9ff9310a05","arxiv_id":"2411.10951","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"TSFormer prunes attention tokens via Min-p sampling and a learned threshold gate, achieving competitive UHD restoration with 3.38M parameters and faster inference than prior models.","lead":"TSFormer is a lightweight Transformer for restoring ultra-high-definition images, combining Min-p token sampling with a claimed random-matrix-theory trust filter to prune attention tokens. It reports state-of-the-art restoration quality across six UHD benchmarks at about 3.4M parameters and near-real-time speed on a single GPU.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core RMT trust mechanism is undefined: Eq. 14 uses a scalar from an unspecified 'spectral density' to scale the threshold, so the claimed novelty cannot be implemented, tested, or falsified; this alone invalidates the paper's central contribution.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the RMT trust mechanism is the paper's core innovation, and it is undefined. I agree. The strongest claim has three parts: SOTA quality, real-time 4K speed, and a Min-p+RMT mechanism that enables both. The speed claim ('we infer that a 4K image requires 40 fps' in Section 6 from a 0.012s measurement at 1024×1024 in Table 7) is an unsupported extrapolation by roughly 8x pixel count; but that could, in principle, be patched by measuring actual 4K inference. The real-time claim is empirical and falsifiable, whereas the RMT claim is not even well-posed. Eq. 14 is presented as a mathematical statement, but \\hat{\\rho} is never defined, so there is no way to re-implement, verify, or falsify it. The paper's own ablations (Table 6, Table 9) are the only evidence, and they compare against straw-man proxies (FED, ISA) that are also not specified. The data mismatches (e.g., UHDFormer parameter count varying between 0.34M and 34.55M across tables) further undermine trust in the reported numbers, but the fundamental problem is that the core mechanism is a black box. Thus the central contribution cannot be accepted as a contribution; it is an assertion. The reader's verdict of REJECT is appropriate. No adjustment is needed.","tokens_in":12049,"tokens_out":5965,"duration_ms":56134,"concrete_test":"Obtain the official code or a precise mathematical definition of \\hat{\\rho} from the authors. Independently re-implement Eq. 14 using any mathematically consistent definition of spectral density (e.g., the mean of the eigenvalue distribution of M) and retrain the model on UHD-Haze with this gate. If the resulting PSNR does not match Table 6's 24.88 dB within ±0.1 dB, the reported ablation cannot be reproduced and the RMT mechanism is not the cause of the claimed gain. If no code or definition is available, the claim is untestable and must be rejected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.3, Eq. 14 defines the trusted threshold as \\hat{Threshold} = \\hat{\\rho} · Threshold, where \\hat{\\rho} is said to be generated from 'the spectral density \\rho of the random matrix M' through 'a mean pooling layer and a sigmoid function'. No formula for \\rho is given, no definition of the pooling/sigmoid composition, no learning rule or loss term, and no diagnostic that \\hat{\\rho} correlates with noise level. The claim 'If M is more random, it may contain more noise and therefore requires a higher threshold' is presented as a self-evident property without proof. The downsampling of M before computing \\rho is described only as 'bilinear interpolation' on an unspecified tensor shape; for an attention matrix, it is unclear whether the interpolation is spatial or across the token dimension. The ablation in Table 9 compares 'FED', 'ISA', and 'Ours', but FED and ISA are each described in two sentences with no implementation details, and the runtime numbers are implausible (0.45s for FED vs 0.012s for Ours). Because the entire 'trusted learning' contribution is this scalar gate, and the scalar is never defined, the central claim of RMT-guided token filtering is not a scientific claim; it is an opaque placeholder. The reader's weakest assumption is therefore exactly the right diagnosis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TSFormer, a lightweight transformer for UHD image restoration that combines Min-p sampling with a so-called trusted learning mechanism based on random matrix theory (RMT) to sparsify attention. The architecture is an encoder-decoder with Trusted Sparse Blocks containing Min-p Sparse Attention (MSA), frequency-domain attention computation, and feature fusion blocks. Experiments are reported on five UHD tasks (low-light enhancement, dehazing, deblurring, deraining, desnowing) with claims of state-of-the-art PSNR/SSIM and real-time 4K inference (40 fps) at 3.38M parameters. The paper also claims that MSA can be plugged into other transformer-based restoration models to improve efficiency.","tokens_in":12331,"tokens_out":6547,"duration_ms":60370,"significance":"If the claims were substantiated, the combination of Min-p sparsification and an RMT-based trusted filter could offer a useful efficiency/quality trade-off for UHD restoration, and the plug-and-play MSA study would broaden its impact. The paper contains a wide set of experiments across five UHD benchmarks and a generalization study on three existing models, which is a strength. However, the central RMT mechanism is not actually defined (Eq. 14 leaves the spectral density unspecified), and the experimental tables contain internal inconsistencies. As a result, the significance of the work cannot be assessed from the current text: the core novelty is a placeholder, and the empirical validation is unreliable in places.","major_comments":[{"comment":"The trusted learning mechanism is not defined. Equation (14) sets \\hat{Threshold} = \\hat{\\rho} \\cdot Threshold, where \\hat{\\rho} is said to be generated from 'the spectral density \\rho of the random matrix M' through 'a mean pooling layer and a sigmoid function'. No formula for \\rho is given, no matrix ensemble is specified, the pooling/sigmoid composition is not described, and no learning rule or loss term is provided for \\hat{\\rho}. The claim that 'If M is more random, it may contain more noise and therefore requires a higher threshold' is an assertion without derivation. Because this mechanism is the paper's claimed novelty for robustness and generalization, the central contribution is not a well-defined scientific claim, and the experiments cannot be interpreted as testing it.","section":"Section 3.3, Eq. (14)"},{"comment":"The real-time 4K claim of 40 fps is not supported by the reported measurements. Table 7 gives an inference time of 0.012 s at 1024×1024 resolution, which corresponds to about 83 fps at that resolution. A 4K frame (3840×2160) has roughly 8.3 times as many pixels, so a linear extrapolation gives about 10 fps, not 40 fps. The paper's statement 'Without considering I/O operations, we infer that a 4K image requires 40 fps' needs a direct measurement on 4K inputs or a careful scaling analysis accounting for patch-based processing and resolution-dependent components.","section":"Section 6 and Table 7"},{"comment":"The models in Table 10 appear to be mislabeled, undermining the plug-and-play claims in Section 7. 'UHDFormer' is listed with 34.55M parameters and 36.61 dB PSNR, but Table 1 reports UHDformer with 0.34M parameters, and Tables 5 and 8 report UHDformer achieving 36.61 dB on UHD-Rain and UHD-Snow. 'UHDFour' is listed with 41.56 dB PSNR, a value that matches UHDDIP in Tables 5 and 8, not UHDFour's 26.23 dB on UHD-LL in Table 1. These inconsistencies make it impossible to verify the claimed efficiency gains and minimal performance loss when integrating MSA.","section":"Table 10"},{"comment":"The reported performance gains over the strongest baseline UHDDIP are small: 0.01 dB on UHD-Blur, 0.19 dB on UHD-Haze, 0.22 dB on UHD-Rain, and 0.26 dB on UHD-Snow. No error bars, confidence intervals, or repeated-run statistics are provided. Differences of this magnitude are within typical training variation for image restoration models, so the claim of state-of-the-art robustness and generalization is not statistically substantiated.","section":"Tables 2, 4, 5, 8"},{"comment":"The ablation against FED and ISA is not reproducible. FED and ISA are each described in two sentences, with no specification of patch size, how the eigenvalue threshold \\tau is chosen, how the strategies are integrated into the MSA module, or how the runtime is measured. The runtime difference between FED (0.45 s) and the proposed method (0.012 s) is over 37×, but no complexity analysis is given to explain where this speed-up comes from. Without implementation details, the comparison cannot be verified or reproduced.","section":"Section 5.2 and Table 9"}],"minor_comments":[{"comment":"The sentence 'A decoder is a mirrored operation of a decoder' should probably read 'a mirrored operation of the encoder'.","section":"Section 3.2"},{"comment":"The shapes of QFFT, KFFT, and M are not specified, so it is unclear whether the FFT is applied along the spatial dimension or the token dimension, and what the resulting attention matrix size is.","section":"Section 3.3, Eqs. (10)-(11)"},{"comment":"The operation in Eq. (13) is a deterministic hard threshold, not the stochastic 'Min-p sampling' used in LLM decoding; the terminology should be clarified to avoid confusion.","section":"Section 3.3, Eq. (13)"},{"comment":"The 'cumulative probability distribution plots' are not explained; the axes are not labeled, and the relation of the plots to the sampling methods is unclear.","section":"Figure 7"},{"comment":"Several references are inconsistent or appear to use placeholder-like arXiv numbers: [15] is 'arXiv:2303.12345', [22] is 'arXiv:2301.12345', and two different entries ([5] and [30], [21] and [22]) describe the same methods (SFNet and DehazeFormer) with different authors and titles. The reference list should be checked.","section":"References"},{"comment":"The heading 'MSA of Potential' is likely a typo and should be revised.","section":"Section 7 heading"}],"recommendation":"reject","confidential_remarks":"The reference list contains entries that look like placeholders or are mutually inconsistent (e.g., two different SFNet references and two different DehazeFormer references); the editor should verify citation integrity. The paper's core contribution is an undefined 'random matrix theory' gate, and the experimental tables contain mismatches that undermine the reported results. These issues go beyond what can be fixed in a routine revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper's stated novelty—random-matrix-theory trusted filtering—is not actually in the paper. Eq. (14) defines the trusted threshold as rho times the Min-p threshold, where rho comes from a mean-pooling layer and sigmoid applied to the 'spectral density' of the attention matrix. No spectral density is ever computed or defined; no eigenvalue analysis appears; no theorem is invoked. The stress-test note is right: this is an opaque placeholder, not a mechanism. Second, the actual system is a reasonably clean combination of Min-p sampling (borrowed from LLM decoding) with FFT-based attention, on a lightweight U-shaped net. That part is coherent, and the plug-in experiments in Table 10, if the numbers could be trusted, would be a useful engineering data point.\n\nWhat the paper does well: it runs five UHD restoration tasks, compares against a broad set of baselines, and the MSA idea is simple enough to be dropped into other transformers. The writing is readable. The architecture figure is clear. The Min-p attention sparsification is a legitimate idea, and the ablation in Table 6 shows it helps over top-k by a couple tenths of a dB.\n\nNow the soft spots, in proportion. The central flaw is load-bearing: if the RMT claim is removed, the paper is an incremental engineering entry—gains over UHDDIP are under 0.3 dB, often under 0.1, and TSFormer uses about four times the parameters. The 4K 40fps claim is extrapolated from a 1024x1024 timing on a 2080Ti; naive scaling gives roughly 10–20fps at 4K, and no 4K timing is reported. The tables have internal inconsistencies: Table 10 gives UHDFormer 34.55M parameters while Table 1 and Table 4 list UHDformer at 0.34M; the names UHD, UHDFour, and UHDformer are shuffled. The FED and ISA ablations in Table 9 get two sentences each and implausible runtimes. No code, no data, and a handful of references look unreliable. These are not minor nits; they make the empirical claims hard to verify.\n\nWho is this for? A reader focused on efficient UHD restoration might pick up the Min-p + sparse attention idea from this paper and try it in their own model, but they would want code and honest numbers before trusting it.\n\nRecommendation: If I were an editor, I would not desk reject outright—there is a real architectural idea here—but I would send it to review only with the expectation that the RMT claim be removed or actually implemented, code be released, and the timing claim be remade on real 4K data. In its current form, I would not accept or recommend anyone build on it.","headline":"The RMT 'trusted filtering' is never defined—Eq. 14 is a learned scalar with no spectral density—so the paper's central novelty is opaque, and the empirical gains are incremental.","tokens_in":12894,"tokens_out":5794,"would_cite":false,"duration_ms":67685,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TSFormer claims to restore 4K images at 40 fps with just 3.38M parameters while beating far larger models on multiple UHD restoration benchmarks.","keywords":["UHD image restoration","token sparsification","Min-p sampling","random matrix theory","lightweight transformer","real-time 4K","low-light enhancement","image dehazing"],"falsifier":"Fix all hyperparameters and replace the learned trust scalar $\\hat{\\rho}$ with a constant, or inject controlled white noise into the attention map and check whether $\\hat{\\rho}$ rises with noise variance; if the restored-image PSNR does not drop when $\\hat{\\rho}$ is frozen, or if the scalar does not track injected noise, the claimed random-matrix-theory gate is not the cause of the robustness gains.","tokens_in":11795,"feed_emoji":"🖼️","tokens_out":7320,"duration_ms":68992,"temperature":0.7,"pith_summary":"The paper introduces TSFormer, a transformer for restoring ultra-high-definition images that deliberately lets very few tokens move through the network. Its central claim is that a dynamic token filter, Min-p sampling scaled by a trust scalar derived from the spectral density of the attention map, keeps only high-confidence features, so a 3.38M-parameter model can restore a 4K image in real time at about 40 fps while still matching or beating much larger restoration models. The paper reports state-of-the-art PSNR/SSIM on five UHD restoration tasks and shows that plugging the same filter into other transformer-based restorers cuts their FLOPs and runtime with minimal quality loss. If these claims hold, high-resolution restoration no longer requires heavy models, which matters for real-time applications on a single GPU.","feed_headline":"TSFormer claims 4K restoration at 40 fps with 3.38M params","feed_subtitle":"A sparse attention gate that keeps only high-confidence tokens is claimed to beat far larger models on five UHD tasks.","key_machinery":"Min-p Sparse Attention (MSA), a sparse attention mechanism in which the attention map $M$ is computed by multiplying FFT-transformed query and key patches, then sparsified by retaining entries above $\\text{Threshold} = p_{\\text{base}}\\max(M)$. Before thresholding, $M$ is treated as a random matrix, downsampled, and its spectral density is reduced to a scalar $\\hat{\\rho}$ by mean pooling and a sigmoid, rescaling the threshold via $\\hat{\\text{Threshold}} = \\hat{\\rho} \\cdot \\text{Threshold}$. The scalar is meant to raise the threshold when the attention matrix is noisy or random, so that unstable tokens are filtered. MSA sits inside Trusted Sparse Blocks together with a feed-forward network, and the whole U-shaped network uses Feature Fusion Blocks to merge encoder-decoder features.","core_discovery":"The central discovery, as the authors present it, is that most tokens in an attention map for UHD restoration are not worth attending to, and that a probability-based threshold with a random-matrix-theory trust adjustment selects the rest. Concretely, attention scores are computed via FFT, then any score below $\\text{Threshold} = p_{\\text{base}} \\cdot \\max(M)$ is zeroed, and the threshold is multiplied by $\\hat{\\rho}$, a scalar obtained by downsampling the attention matrix, computing its spectral density, applying mean pooling and a sigmoid. The result, Min-p Sparse Attention, repeats inside every block of a U-shaped encoder-decoder. Across UHD low-light, dehazing, deblurring, deraining, and desnowing benchmarks, the model reports the best PSNR and SSIM among the compared methods, with 3.38M parameters and a reported 4K inference speed near 40 fps, and the same MSA unit accelerates LLFormer, UHDformer, and UHDFour when added to them.","pith_inferences":["One immediate independent test is to replace the learned scalar $\\hat{\\rho}$ with a constant while keeping everything else fixed; if the PSNR gain on UHD-Haze largely disappears, the Min-p threshold alone may be doing most of the work.","The sparse-attention gate could transfer to other high-resolution dense prediction tasks, such as UHD video super-resolution or compression artifact removal, where throughput is the main bottleneck.","A stricter version of the trusted mechanism would derive the threshold from a formal confidence bound on the eigenvalue distribution rather than a sigmoid of a mean-pooled spectral density; such a version would make the random-matrix-theory connection testable in isolation."],"forward_implications":["A 4K image can be restored in about 25 ms on a single GPU, which makes real-time UHD video restoration feasible.","Other transformer-based restoration models can adopt MSA and cut FLOPs and runtime by roughly 20-25 percent while approximately maintaining PSNR and SSIM, as demonstrated on LLFormer, UHDformer, and UHDFour.","Restoration quality does not need to scale with parameter count on these benchmarks: 3.38M parameters outperforms models with many times more parameters.","The same architecture handles multiple degradations, including low light, haze, blur, rain, and snow, without task-specific heads, suggesting a single all-in-one model can serve UHD restoration."],"supporting_citations":[{"why":"Defines Min-p sampling, the dynamic probability-based threshold that MSA adapts to sparsify attention tokens.","marker":"[20]"},{"why":"Supplies the probabilistic sparsification theory that motivates Min-p's high-confidence feature selection.","marker":"[1]"},{"why":"Random matrix theory source cited for eigenvalue-based analysis of feature stability in the trusted filtering step.","marker":"[7]"},{"why":"Random matrix theory textbook cited to justify treating attention matrices as random matrices and using their spectral density.","marker":"[23]"},{"why":"Introduces the UHD-Blur and UHD-Haze datasets used to train and evaluate the reported deblurring and dehazing results.","marker":"[25]"},{"why":"Restormer is the main general-purpose restoration transformer baseline that TSFormer compares against and that MSA is shown to accelerate.","marker":"[38]"},{"why":"UHDformer is a strong lightweight UHD baseline on several tasks, and MSA integration into it demonstrates the plug-and-play efficiency claim.","marker":"[27]"},{"why":"UHDDIP is the prior state-of-the-art on deraining and desnowing benchmarks whose PSNR and SSIM values TSFormer is claimed to exceed.","marker":"[28]"}],"fun_headline_variants":["TSFormer: real-time 4K restoration with only 3.38M parameters","Sparse attention enables TSFormer to restore 4K at 40fps","TSFormer: Trusted sparse attention for fast UHD restoration","Min-p sparse attention: TSFormer hits 40fps on 4K with 3.38M params","TSFormer: 3.38M-param model does real-time 4K via token filtering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that the scalar computed from the spectral density of the downsampled attention matrix, after mean pooling and a sigmoid, genuinely measures token uncertainty and noise, so that multiplying the Min-p threshold by it suppresses noise and improves robustness; if that scalar is not a real noise measure, the trusted-filtering results are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["TSFormer: real-time 4K restoration with only 3.38M parameters","Sparse attention enables TSFormer to restore 4K at 40fps","TSFormer: Trusted sparse attention for fast UHD restoration","Min-p sparse attention: TSFormer hits 40fps on 4K with 3.38M params","TSFormer: 3.38M-param model does real-time 4K via token filtering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000465,"raw_usage":{"total_tokens":2320,"prompt_tokens":940,"completion_tokens":1380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1266}},"tokens_in":556,"tokens_out":1380,"duration_ms":11700,"temperature":1.0,"reasoning_tokens":1266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:06:33.244813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix all hyperparameters and replace the learned trust scalar $\\hat{\\rho}$ with a constant, or inject controlled white noise into the attention map and check whether $\\hat{\\rho}$ rises with noise variance; if the restored-image PSNR does not drop when $\\hat{\\rho}$ is frozen, or if the scalar does not track injected noise, the claimed random-matrix-theory gate is not the cause of the robustness gains.","supporting_citations":[{"cited_title":"Estimating error probabilities of sparse representations","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic sparsification theory that motivates Min-p's high-confidence feature selection."},{"cited_title":"Eigenvalues and condition numbers of ran- dom matrices","cited_arxiv_id":null,"evidence_quote":"Random matrix theory source cited for eigenvalue-based analysis of feature stability in the trusted filtering step."},{"cited_title":"Topics in Random Matrix Theory","cited_arxiv_id":null,"evidence_quote":"Random matrix theory textbook cited to justify treating attention matrices as random matrices and using their spectral density."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Restormer is the main general-purpose restoration transformer baseline that TSFormer compares against and that MSA is shown to accelerate."},{"cited_title":"Uhdformer: A lightweight model for ultra-high-definition image restoration","cited_arxiv_id":null,"evidence_quote":"UHDformer is a strong lightweight UHD baseline on several tasks, and MSA integration into it demonstrates the plug-and-play efficiency claim."}],"review_version":1}