{"id":"e72d3491-ac4b-4fba-b9ea-f98b891dd551","arxiv_id":"2601.08341","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A transformer that starts each token with local-plus-sparse-global neighbors, then expands and prunes that candidate list layer by layer, reports consistent state-of-the-art PSNR/SSIM at matched FLOPs on classical and lightweight SR.","lead":"Sparse-attention transformers for super-resolution usually lock each pixel to fixed local windows; this paper gives every pixel its own evolving list of far-away pixels to borrow texture from. The method reports PSNR gains of up to 0.33 dB over the previous best at matched compute, if the published numbers hold.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Protocol mismatch undermines SOTA attribution: IET uses Muon + two-stage training while baselines are published AdamW numbers; the 0.17–0.33 dB gains may be training-recipe effects rather than the IEA mechanism.","rationale":"The reader's verdict is CONDITIONAL, and the protocol mismatch is listed as one of the fragilities, but the reader's weakest_assumption focuses on the transitivity/expander premise. I see the protocol mismatch as the single most load-bearing concern because it directly attacks the central claim (SOTA at comparable complexity) without depending on the internal rationale. If the expansion mechanism's transitivity assumption is false, the method might still be empirically superior—only the explanation would be wrong. But if the training-recipe confound is real, the reported superiority over PFT/ATD is unproven, and the abstract's claim collapses. The ablation study (Table 1) is internally controlled and lends credibility to the mechanism, but it uses IET-light trained for 250k iterations on DIV2K only; it cannot rule out the optimizer/schedule confound in the full comparisons. The concrete test directly controls for this confound in the closest competitor. I therefore retain UNCHANGED verdict: the paper remains CONDITIONAL until the protocol-matching experiment is run or code is released to enable it. The manuscript's own limitation—the code link is a placeholder—further supports keeping the verdict conditional, though that is secondary to the scientific check.","tokens_in":18537,"tokens_out":3126,"duration_ms":31147,"concrete_test":"Retrain the strongest competitor PFT (and ATD) with IET's exact training protocol on the same DF2K data: Muon+AdamW hybrid, two-stage 50×50→75×75 patches, 300k+250k iterations, dilation 2→3, matching the authors' hyperparameters. If PFT's ×2 Urban100 PSNR rises to within 0.05 dB of IET's 35.07, the architecture-specific advantage is not established. Complementary check: train IET under the standard AdamW recipe used to generate PFT's published numbers; if the margin over PFT shrinks below significance, the SOTA claim is a training-recipe artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IET achieves state-of-the-art performance at comparable computational complexity (Abstract; §4.3). The evidence in Tables 4–5 shows gains over PFT/ATD, e.g., 0.17 dB on Urban100 ×2, 0.33 dB for IET-light over PFT-light, and 0.43 dB over ATD-light. However, all baselines are cited published numbers trained with the standard AdamW recipe. IET is trained with a different protocol: a Muon+AdamW hybrid optimizer, two-stage training with 50×50 then 75×75 patches, 300k+250k iterations, and dilation 2→3 (main text §4.1; Supp. F). No baseline is retrained under this protocol. In super-resolution, optimizer choice and patch-size schedules of this magnitude routinely shift PSNR by several tenths of a dB, so the reported margins fall well within the range attributable to training recipe alone. The 'comparable computational complexity' claim only matches FLOPs, not training protocol, and the performance comparison is therefore not controlled for the optimizer/schedule confound. The internal ablation (Table 1) isolates the expansion mechanism under a fixed protocol, but it does not establish that the same gain transfers to the headline comparison against prior work. A second concern—the unverified transitivity assumption of §3.1—would only affect the explanation if the empirical gain survives, whereas the protocol mismatch directly threatens the central empirical claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Individualized Exploratory Transformer (IET) for single-image super-resolution. IET replaces grouped attention (window-based or category-based) with an individualized exploratory attention (IEA) mechanism in which each token maintains its own asymmetric set of attention candidates. Candidates are initialized via a Dense-Local Sparse-Global (DLSG) scheme, then expanded by linking to two-hop neighbors of the most similar one-hop neighbors, and pruned by a sparsification step. A Similarity-Fused FFN further mixes features of similar tokens. Experiments on Set5, Set14, BSD100, Urban100, and Manga109 report consistent PSNR/SSIM improvements over recent window- and category-based methods at comparable FLOPs, for both a full-size model and a lightweight variant.","tokens_in":18802,"tokens_out":6261,"duration_ms":53205,"significance":"If the reported gains hold under controlled comparisons, IEA is a useful alternative to grouped attention for SR, and the paper's systematic evaluation across five benchmarks and two model scales is a strength. The paper also provides an inference-time comparison and visualizations of the candidate-refinement process. However, the central 'state-of-the-art under comparable computational complexity' claim is currently weakened by a training-protocol mismatch with the published baselines, and the main ablation table is misreported in the text, attributing the largest single-component gain to the wrong component. These issues need to be resolved before the headline claim can be accepted. The paper does not include machine-checked proofs; the code link is a placeholder rather than a resolved repository.","major_comments":[{"comment":"The text and Table 1 disagree. The text states that 'the second model introduces the Expansion mechanism' and attributes about 0.58 dB improvement on Urban100 to it. In the table, the second row is DLSG only (26.75 dB) and the third row is Expansion only (26.26 dB). Thus DLSG alone gives +0.58 dB over the baseline, while Expansion alone gives only +0.09 dB. This reversal undermines the conclusion that 'directly connecting two-hop neighbors effectively expands the content-aware receptive field and brings substantial performance gains.' Please correct the text or the table and re-evaluate the role of expansion relative to DLSG in the final model.","section":"§4.2, Table 1"},{"comment":"The headline SOTA claim is not controlled for training protocol. IET is trained with a Muon+AdamW hybrid optimizer, a two-stage schedule (50×50 then 75×75 patches, 300k+250k iterations), and a dilation change from 2 to 3, while all baselines are cited from published numbers obtained with the standard AdamW recipe. In SR, optimizer choice and patch-size schedules of this magnitude can shift PSNR by several tenths of a dB, which is exactly the range of the reported margins (e.g., 0.17 dB over PFT on Urban100 ×2; 0.33 dB for IET-light over PFT-light). Please retrain at least one strong baseline (e.g., PFT or ATD) under the identical protocol, or provide evidence that the training recipe does not affect the relative ranking.","section":"§4.1, Supp. F, and §4.3, Tables 4–5"},{"comment":"The expansion mechanism rests on the transitivity premise that if A is similar to B and B is similar to C in the preceding layer, then A is likely similar to C in subsequent layers. This premise is not verified. The appeal to Expander Graphs requires a 'unique neighborhood' property, but the paper provides no measurement of whether the learned attention graphs satisfy this property, nor what fraction of expanded two-hop candidates survive sparsification and are actually used. This is not fatal to the empirical results, but it is needed to support the 'content-aware exploratory' explanation. Please add a quantitative analysis of the expanded candidates, or soften the graph-theoretic framing.","section":"§3.1–3.2, Eqs. (5)–(8)"}],"minor_comments":[{"comment":"In the lightweight SR table, the rows for MambaIRv2-light at ×3 and ×4 are labeled with Scale = ×2. The scale column should read ×3 and ×4 respectively.","section":"Table 5"},{"comment":"The supplement text says 'in Section A ... Section B ... Section C', but the actual section headings are F, G, H. Please renumber for consistency.","section":"Supplementary Material"},{"comment":"The code link is given as 'here' with no actual URL. Please provide a working repository link.","section":"Abstract"},{"comment":"The sentence 'IET-light outperforms SwinIR by 0.02dB on Set5 and 0.19dB on Urban100' is ambiguous: it refers to the non-light SwinIR from Table 4, not SwinIR-light from Table 5. Please clarify to avoid conflation of the two models.","section":"§4.3"},{"comment":"In SF-FFN, the selection of 'the most similar neighbors I_highest' is based on updated neighbor indices. It would be helpful to state whether this selection is performed per head or per token, and whether it adds meaningful computation on top of the already-computed attention indices.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The authors overlap with PFT [27] and ATD [44], and the CUDA sparse-matrix multiplication framework is reused from PFT. This is not by itself a problem, but the incremental contribution over those works should be stated more explicitly. The protocol mismatch and the ablation mislabeling are the key concerns; if the authors can supply a controlled retraining of at least one baseline and correct the ablation interpretation, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a legitimately new attention mechanism for SR, and the paper's internal evidence is decent, but the headline \"state-of-the-art at comparable complexity\" is strained because the comparisons are not protocol-controlled. The per-token index-matrix propagation (IEA) with two-hop expansion and sparsification is genuinely new relative to PFT and ATD. The ablation in Table 1 isolates each component and shows consistent gains; the 0.17–0.33 dB improvement over PFT across benchmarks is plausible if the mechanism works. SF-FFN is a reasonable add-on.\n\nThe main soft spot is the training-recipe confound. IET uses Muon+AdamW hybrid, two-stage schedule, and different patch sizes; baselines are published AdamW numbers. In SR, these choices can shift PSNR by a few tenths of a dB, so the claimed margins are within the range that the recipe alone could explain. The internal ablation controls for this, but the headline comparison does not. That's fixable: retrain PFT and ATD under the same recipe, or at least show an AdamW-trained IET variant for the main benchmarks. Until then, the claim should be \"competitive under our training recipe,\" not unqualified SOTA.\n\nThe transitivity assumption in §3.1 is asserted with an expander-graph analogy but never verified on the trained attention graphs. This weakens the explanation, not the empirical result — the ablation still shows the expansion helps. The code link is a placeholder, which isn't fatal but makes independent verification impossible. The \"9.1% complexity\" sentence is ambiguous: it matches full SwinIR, not SwinIR-light, which is the model listed in the table; the text should say so explicitly.\n\nOverall, the paper is a coherent, honest-incremental contribution. Nothing here is incoherent or circular; the numbers are what they claim to be under a fixed protocol. The main fix is to address the protocol confound before taking the SOTA claim at face value. I'd send it to peer review — it's worth referee time — and the reviewer should ask for the retrained baselines and the code. I'd bring it to the reading group; the mechanism may transfer to other restoration or dense-prediction tasks.","headline":"A genuinely new per-token expansion attention mechanism for SR, with good internal ablations, but the SOTA claim overreaches because baselines aren't trained under the same recipe.","tokens_in":19430,"tokens_out":4707,"would_cite":true,"duration_ms":42061,"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":"This paper argues that token-adaptive asymmetric attention candidate selection outperforms fixed-group attention for image super-resolution at matched compute.","keywords":["single image super-resolution","transformer","attention mechanism","token-adaptive attention","asymmetric similarity","graph expansion","sparsification","individualized exploratory attention"],"falsifier":"Train the same IET architecture with expansion replaced by uniformly random long-range candidates at the same k and FLOPs; if random candidates close the 0.58 dB gap on Urban100 ×4, the two-hop transitivity assumption is not what drives the gain. Equivalently, inspect the learned attention graphs: if most two-hop promoted candidates are pruned immediately or receive near-zero attention weights, the expansion rationale is not supported.","tokens_in":18305,"feed_emoji":"🔍","tokens_out":6773,"duration_ms":54067,"temperature":0.7,"pith_summary":"The paper tries to establish that super-resolution Transformers do not need fixed groupings—windows or semantic categories—to keep attention cheap. Instead, each token can maintain its own directed list of attention candidates, chosen by content similarity, and that list can be refined across layers: nearby candidates seed the search, two-hop neighbors of high-similarity tokens are promoted to direct candidates, and low-similarity ones are pruned. The proposed IET model reports state-of-the-art PSNR/SSIM on standard SR benchmarks under comparable FLOPs to recent window- and category-based methods, with the largest gains on texture-heavy datasets like Urban100. A sympathetic reader would care because it suggests the inductive bias of fixed local windows is not necessary for efficient global attention in low-level vision.","feed_headline":"Per-token attention candidates beat fixed windows in super-resolution","feed_subtitle":"Each image token expands and prunes its own neighbor list layer by layer, gaining up to 0.43 dB at matched FLOPs.","key_machinery":"The central machinery is the Individualized Exploratory Attention (IEA) mechanism, built on an index matrix I ∈ R^{N×k} that explicitly stores each token's top-k attention candidates. The paper treats token similarities as a directed graph and applies two graph operations: expansion, which promotes two-hop neighbors (tokens similar to tokens that are similar to the query) into direct candidates via Eqs. 5–8, and sparsification, which prunes low-similarity candidates via top-k selection (Eq. 4). This layered expansion-and-sparsification is the mechanism that lets attention grow from local to adaptive global under a fixed computational budget.","core_discovery":"The central claim is that attention candidates for super-resolution should be individual, asymmetric, and content-aware rather than group-defined. Concretely, the paper introduces Individualized Exploratory Attention (IEA): each token starts with a Dense-Local Sparse-Global (DLSG) initialization—dense attention within a small region plus uniformly sampled distant tokens—then in subsequent layers expands its candidate set by gathering the top one-hop neighbors of its own top one-hop neighbors (two-hop expansion) while sparsifying away low-attention tokens. The index matrix that records each token's candidates is optimized implicitly through this expansion-and-sparsification process. The paper","pith_inferences":["A testable extension would measure what fraction of two-hop-expanded candidates actually survive sparsification and contribute to the output; if most are pruned or contribute little, the 0.58 dB gain might come from having more candidates rather than from the two-hop rationale.","The paper invokes expander-graph properties to justify DLSG initialization but does not verify that trained attention graphs satisfy the required uniqueness-of-neighborhood condition; checking the spectral gap or the overlap of one-hop neighborhoods of real models would ground or refute that justification.","A same-training-protocol comparison would make the reported margins directly attributable to the attention mechanism, since the baselines' published numbers come from a different optimizer and schedule.","The same expansion-and-sparsification loop could be applied to other token-based tasks, such as segmentation or denoising, where long-range but content-selective context is valuable—an easy extension of the code."],"forward_implications":["If IET's reported gains hold, the fixed-window inductive bias in SR Transformers can be replaced by a per-token, content-driven candidate search without increasing FLOPs.","The ablation attributes about 0.58 dB on Urban100 ×4 to the expansion mechanism, implying that two-hop similarity propagation is the main source of the improvement over a local-only baseline.","IET-light matches or exceeds much larger models on several benchmarks while using a fraction of the FLOPs, suggesting the candidate-selection principle transfers to lightweight settings.","The paper's visualization shows attention candidates spreading beyond a 32×32 window across blocks, indicating the mechanism produces genuinely long-range content-aware receptive fields.","The authors state the design could transfer to other high-level vision tasks and NLP, where adaptive similarity modeling may help."],"fun_headline_variants":["Token-custom attention candidates for super-resolution","Per-token candidate selection boosts SR efficiency","Adaptive attention candidates outperform fixed SR groups","IET: each token explores its own attention path"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that token similarity is transitive across layers: if token A is similar to B and B is similar to C in one layer, then A is similar to C in the next layer—without this transitivity, the two-hop expansion mechanism reduces to adding generic distant context rather than content-aware candidates.","fun_headline_variants_meta":{"raw":{"variants":["Token-custom attention candidates for super-resolution","Per-token candidate selection boosts SR efficiency","Adaptive attention candidates outperform fixed SR groups","IET: each token explores its own attention path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1081,"prompt_tokens":700,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":444,"tokens_out":381,"duration_ms":4299,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:51:07.953958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same IET architecture with expansion replaced by uniformly random long-range candidates at the same k and FLOPs; if random candidates close the 0.58 dB gap on Urban100 ×4, the two-hop transitivity assumption is not what drives the gain. Equivalently, inspect the learned attention graphs: if most two-hop promoted candidates are pruned immediately or receive near-zero attention weights, the expansion rationale is not supported.","supporting_citations":[],"review_version":1}