{"id":"3b12e8c6-e6e0-4ed9-a56a-78ef0669939b","arxiv_id":"2608.08965","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CoRe-UIE, a shared-backbone mixture-of-experts model routed by local degradation cues, achieves the best reported PSNR, SSIM, FSIM, FSIMC, and VSI on UIEB and LSUI and the best four no-reference metrics on U45.","lead":"This paper introduces CoRe-UIE, an underwater image enhancement model that restores different image regions with specialized expert networks for color, haze, texture, and lighting problems. It reports top scores on three standard underwater datasets and argues that region-aware expert mixing beats one global restoration network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The undefined mechanism cues P in Eqs. (4) and (8)-(9) leave the central claim of mechanism-aware region routing unverifiable and unreproducible.","rationale":"The reader's weakest assumption is the correct focal point. The central contribution is not a generic MoE but degradation-mechanism-aware expert collaboration; the only mechanism-specific information in the model is P, and it is never defined. The router in Eq. (4) and the response alignment loss in Eq. (9) both depend on P, so the mechanism-aware specialization claim cannot be implemented or tested from the paper. The empirical SOTA numbers do not resolve the gap because the ablations vary whole experts, not the cues. I find no internal contradiction beyond this omission; the architecture is plausible and the reported results are not impossible, so the reader's CONDITIONAL verdict stands. Secondary issues (Top-k apparently chosen on test set, the VSI discrepancy between Table 1 and Table 3) reinforce the need for code and a validation protocol but are not the primary load-bearing weakness.","tokens_in":11355,"tokens_out":7594,"duration_ms":74941,"concrete_test":"Request the exact formulas or code for Pc, Psc, Pt, Pl, then retrain CoRe-UIE on the same UIEB split with P replaced by (a) constant one-maps and (b) random noise maps, keeping all other settings fixed. If either control reproduces the reported 26.8815 dB PSNR and 0.9103 SSIM within run-to-run variance, the cues are not load-bearing; if both controls degrade substantially, the mechanism claim still requires the published formulas to be independently verifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central differentiator is mechanism-aware, region-adaptive expert routing, but the only mechanism-specific signal in the model is the input-derived cue set P={Pc,Psc,Pt,Pl}, and P is never defined. The text names the four cues (\"color imbalance, scattering-related contrast degradation, texture structure, illumination risk\") and says they are \"input-derived\", but gives no formulas, estimation procedure, normalization, or reference. This is load-bearing because both the router G=Softmax(R(F,P)) in Eq. (4) and the response-alignment loss in Eqs. (8)-(9) take P as input; Eq. (9) explicitly aligns each expert's response Norm(R_i) to Norm(P_i). Without P, the mechanism-aware specialization and the region-wise routing that define CoRe-UIE cannot be implemented, reproduced, or even tested for whether they correspond to actual degradations. If P is instead learned from F, then the claimed \"input-derived physical cues\" are not physical and the specialization story is circular. The reported benchmark numbers do not resolve the gap, because ablations (Table 3) remove whole experts but never vary P.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CoRe-UIE proposes an underwater image enhancement framework based on a content-preserving shared expert and four same-architecture routed experts intended for color correction, scattering suppression, texture recovery, and illumination protection. A region-adaptive Top-k router assigns experts to spatial regions using input-derived mechanism cues, and training adds a response alignment loss plus an HSIC-based representation disentanglement term. Experiments on UIEB, LSUI, and U45 report state-of-the-art or competitive results on full-reference and no-reference metrics, alongside ablations of individual experts, the HSIC term, and the Top-k choice.","tokens_in":11585,"tokens_out":3862,"duration_ms":37390,"significance":"If the missing details are supplied, the paper would offer a plausible and potentially useful way to induce degradation-aware specialization in homogeneous mixture-of-experts architectures for low-level vision. The external benchmark comparisons and the carefully structured ablations are strengths, and the reported gains on UIEB and LSUI are substantial enough to be interesting. However, the central claim of mechanism-aware region routing currently rests on an undefined cue set, and the benchmark tables show an internal inconsistency, so the contribution is not yet fully verifiable.","major_comments":[{"comment":"The four mechanism cues P={Pc, Psc, Pt, Pl} are never defined mathematically. The text only names them as color imbalance, scattering-related contrast degradation, texture structure, and illumination risk, and calls them 'input-derived,' but gives no formulas, estimation procedures, normalization, or references. Both the router G=Softmax(R(F,P)) in Eq. (4) and the response alignment loss in Eq. (9) depend on P, with Eq. (9) explicitly aligning Norm(R_i) to Norm(P_i). Without a precise definition of P, the method cannot be implemented, reproduced, or tested for whether the cues actually correspond to local degradations. Please provide exact definitions or clearly state that P is learned from F; if P is learned, the 'physical cue' interpretation must be revised.","section":"Region-Adaptive Top-k Routing (Eqs. 4 and 8-9)"},{"comment":"The reconstruction loss L_rec is described only as combining 'pixel-wise fidelity terms with structural consistency constraints' and is never written out. Since the quantitative benchmark results are the main evidence for the method's effectiveness, the exact form of L_rec (e.g., L1, L2, SSIM, perceptual terms, and their weights) is needed for reproducibility and for assessing whether the improvement comes from the proposed routing or from an ad hoc reconstruction objective. Please specify L_rec explicitly and report the values of lambda_align and lambda_hsic.","section":"Loss Functions (Eq. 7)"},{"comment":"The default model's VSI on UIEB is reported as 0.9855 in Table 1 but as 0.9821 in Table 3 for 'Ours(ToP-2)', which is the same default configuration. These values cannot both be correct. Moreover, no error bars or significance tests are reported anywhere, so it is unclear whether the differences from the second-best methods are stable across runs. Please reconcile the inconsistent VSI values and provide variance information or repeated-run statistics.","section":"Tables 1 and 3"},{"comment":"The ablations remove whole experts or the HSIC term but never vary or ablate the mechanism cues P. Consequently, the experiments do not demonstrate that routing decisions correspond to the intended degradation mechanisms; they only show that having more experts helps. To support the central claim of mechanism-aware region-adaptive routing, please include analyses that visualize or quantify routing maps, cue maps, or per-expert response heatmaps, and, if possible, a variant in which P is replaced by randomized or ablated cues.","section":"Ablation Study (Table 3)"}],"minor_comments":[{"comment":"The author line 'Ziheng Cao1 Guanying Huo1∗' is missing a comma between the names 'Ziheng Cao1' and 'Guanying Huo1'.","section":"Author line, page 1"},{"comment":"The label 'Ours(ToP-2)' in Table 3 should be 'Ours (Top-2)' to match the terminology used in the text.","section":"Table 3 and text"},{"comment":"The Related Work section states that CoRe-UIE introduces 'four structurally differentiated routed experts,' but the Methodology consistently says all routed experts share the same architecture with independent parameters; please align this wording.","section":"Related Work"},{"comment":"Method names are inconsistent across tables: 'SSUIE' appears in the tables while 'SS-UIE' is used in the text; please unify the naming.","section":"Tables 1-3"},{"comment":"The expert block in Eq. (3) does not specify the nonlinearity sigma or the kernel sizes and normalization for the convolutional layers C1_i and C2_i; please provide these implementation details in the setup section.","section":"Equation (3)"},{"comment":"The spatial normalization Norm(·) used in the response alignment loss is not defined; please specify whether it is per-channel or per-region and how it is computed in practice.","section":"Equation (9)"}],"recommendation":"major_revision","confidential_remarks":"The undefined mechanism cues P are the main technical gap; this is fixable but requires a substantive addition, not just cosmetic editing. The inconsistent VSI values between Tables 1 and 3 should be investigated before acceptance, as they raise doubts about the reliability of the other reported numbers. If the authors can supply exact definitions of P and L_rec, reconcile the table inconsistency, and add routing-map evidence, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper is a solid, workmanlike application of mixture-of-experts to underwater image enhancement. What's new: it puts four shared-backbone experts behind a region-adaptive Top-k router, adds a shared content expert, and uses HSIC to decorrelate expert features. The benchmark results are competitive — best PSNR/SSIM/FSIM/FSIMC/VSI on UIEB and LSUI, best no-reference scores on U45 — and the ablations show each expert and the HSIC term earn their keep. The comparison is external, so there is no derivation-level circularity. The writing is clear and the motivation (spatially heterogeneous, coexisting degradations) is sensible.\n\nThe problem is the degradation cues P. Equations (4) and (9) depend on P for routing and for aligning expert responses, but P is never defined. The paper names four cues (color imbalance, scattering contrast, texture structure, illumination risk) and calls them input-derived, but gives no formulas, no estimation procedure, no normalization, no reference. Without P, the core claim — that experts specialize by degradation mechanism — cannot be implemented, reproduced, or even tested. This is load-bearing, not cosmetic. If P is learned from the features F, the 'physical cue' story collapses into circularity; if P is hand-crafted, the authors need to say exactly how. This is the one thing that keeps the paper from being acceptable.\n\nThere are also smaller but real issues. Table 1 lists Ours VSI=0.9855 on UIEB; Table 3 lists Ours(ToP-2) VSI=0.9821 for the same default model. That is a direct inconsistency. The reconstruction loss L_rec is described only as 'fidelity plus structural consistency,' not specified. k=2 is chosen by comparing Top-1/2/3/4 on the test sets, i.e., test-set selection. No error bars are reported. No code is released.\n\nThe gains over the runner-up are roughly 0.4 dB PSNR on UIEB, so the empirical impact is modest. Still, the architecture is a plausible new combination for a real subfield problem, and the ablations are informative. If the authors define P, fix the tables, and release code, this could be a solid contribution. As it stands, the central mechanism-aware routing claim is unsubstantiated.\n\nMy recommendation: send it to peer review, but require major revision. The referee should ask for a precise definition of P, corrected tables, and either code or a detailed experimental protocol. It's not a desk reject — the idea is worth examining — but it is not acceptable in this form.","headline":"A competent MoE-for-underwater-enhancement paper whose central mechanism-awareness claim rests on undefined degradation cues P; the empirical results are real but the paper needs major revision before acceptance.","tokens_in":12091,"tokens_out":2825,"would_cite":false,"duration_ms":25516,"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":"This paper claims that underwater image enhancement should be modeled as region-wise collaboration among mechanism-specialized experts rather than a single global restoration mapping, and reports top scores on UIEB, LSUI, and U45.","keywords":["underwater image enhancement","mixture of experts","region-adaptive routing","shared-backbone experts","HSIC representation disentanglement","degradation cues","image restoration"],"falsifier":"Retrain CoRe-UIE on UIEB with the four mechanism cues replaced by random or constant per-pixel maps, keeping all other components fixed; if PSNR stays close to 26.88 dB, then cue-guided routing and response alignment are not carrying the claimed specialization, while a large drop would confirm that the missing cue formulas matter.","tokens_in":11191,"feed_emoji":"🌊","tokens_out":7456,"duration_ms":65259,"temperature":0.7,"pith_summary":"Underwater photos are degraded in ways that vary from region to region: one area may have a green color cast, another residual haze, a third washed-out texture, and a fourth dark shadows, with several of these problems present in the same local patch. CoRe-UIE argues that a single network mapping applied uniformly to the whole image is the wrong tool for such compound, spatially mixed degradation. The paper proposes an expert-collaboration network: a shared expert preserves content everywhere, while four routable experts with identical architecture but independent weights specialize in color correction, scattering suppression, texture recovery, and illumination protection, selected per region by a Top-2 router. It reports the best scores among the compared methods on UIEB, LSUI, and U45, with UIEB PSNR of 26.88 dB, and supports the design with ablations that remove each expert, the shared expert, and the HSIC term. A sympathetic reading is that mechanism-aware region-wise routing, not a more powerful unified backbone, is the reason for the gains.","feed_headline":"Four region-routed experts top underwater image benchmarks","feed_subtitle":"Instead of one global mapping, it routes each region to specialized experts and tops UIEB, LSUI, and U45.","key_machinery":"The load-bearing mechanism is a region-adaptive Top-2 router over four shared-backbone routed experts, each a residual block with independent parameters, plus a shared expert active everywhere. The router combines learned features with four input-derived mechanism cues $\\{P_c, P_{sc}, P_t, P_l\\}$ for color imbalance, scattering contrast, texture structure, and illumination risk; a response alignment loss forces each expert's normalized spatial response to resemble its cue, and an HSIC-based loss penalizes statistical dependence among all routed-expert representation pairs. This is what carries the argument that same-architecture experts can specialize to different coexisting degradations without hand-designed heterogeneous structures.","core_discovery":"CoRe-UIE claims that representing underwater restoration as region-adaptive collaboration among degradation-specific experts is both feasible and superior to uniform restoration. The framework keeps a content-preserving shared expert active for all spatial regions and four shared-backbone routed experts active only where their mechanism is relevant; a region-adaptive Top-2 router assigns normalized weights based on learned features plus input-derived mechanism cues. Two additional objectives make structurally identical experts diverge: a response alignment loss matches each expert's spatial response to its corresponding degradation cue, and an HSIC-based loss reduces statistical dependence among expert feature pairs. On UIEB the full model reports 26.88 dB PSNR, 0.9103 SSIM, 0.9666 FSIM, 0.9554 FSIMC, and 0.9855 VSI, and it reports the best compared scores on LSUI and U45 as well, so the paper's claim is that the routing-plus-disentanglement design, not a larger or differently shaped network, drives the improvement.","pith_inferences":["If the same recipe is transferred to other compound-degradation settings such as hazy low-light scenes, rain plus fog, or mixed noise and blur, the four mechanism cues would need to be redefined for each domain; the paper does not test this transfer.","Because the cue formulas are missing, the cleanest isolation experiment is one the paper does not run: compare input-derived cues against purely learned routing predictors with identical capacity and losses; that comparison would show how much of the gain is truly cue-guided.","A consequence the authors leave implicit is interpretability: response alignment should make each expert's spatial activation map readable as a where-is-this-degradation-happening signal, which could aid failure diagnosis in deployment, but the paper does not quantify this."],"forward_implications":["On UIEB, the full model reaches 26.88 dB PSNR, 0.9103 SSIM, 0.9666 FSIM, 0.9554 FSIMC, and 0.9855 VSI, the best values among the compared methods.","On LSUI, CoRe-UIE also tops the compared methods on every full-reference metric, so the advantage is not tied to a single benchmark's degradation distribution.","On U45, which has no ground-truth references, CoRe-UIE gets the best NIQE, BRISQUE, CEIQ, and PIQE scores among the compared methods.","Ablations show each routed expert contributes; dropping the scattering expert causes the largest PSNR drop, and removing the HSIC term also degrades all reported metrics.","Top-2 routing outperforms Top-1, Top-3, and Top-4, supporting the claim that locally coexisting degradations need more than one expert but not dense activation of all experts."],"supporting_citations":[{"why":"Supplies the UIEB benchmark and reference pairs used for the main full-reference evaluation.","marker":"Li et al. 2020"},{"why":"Supplies the LSUI dataset used for the second full-reference evaluation.","marker":"Peng, Zhu, and Bian 2023"},{"why":"Provides the Hilbert–Schmidt Independence Criterion on which the expert disentanglement loss is built.","marker":"Gretton et al. 2005"},{"why":"Establishes the sparsely-gated mixture-of-experts routing idea that the region-adaptive Top-k router adapts.","marker":"Shazeer et al. 2017"},{"why":"Shows how sparse expert routing can be applied to visual tokens, a direct precursor of the region-level routing used here.","marker":"Riquelme et al. 2021"},{"why":"Provides the reconstruction-loss practice that the paper follows for pixel-wise fidelity and structural consistency.","marker":"Zamir et al. 2022"}],"fun_headline_variants":["Region-routed expert collaboration lifts underwater image quality","Specialized experts fix underwater images per region","CoRe-UIE routes regions to dedicated underwater restoration experts","HSIC-disentangled experts outperform on three underwater benchmarks","Per-region expert routing sets new bar for underwater enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's routing and alignment losses are built entirely on four input-derived mechanism cues whose formulas are never given, so if those cues cannot be computed reliably from an input image, the mechanism-aware specialization has no definable foundation.","fun_headline_variants_meta":{"raw":{"variants":["Region-routed expert collaboration lifts underwater image quality","Specialized experts fix underwater images per region","CoRe-UIE routes regions to dedicated underwater restoration experts","HSIC-disentangled experts outperform on three underwater benchmarks","Per-region expert routing sets new bar for underwater enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1564,"prompt_tokens":947,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":563,"tokens_out":617,"duration_ms":5895,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:18:29.429597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain CoRe-UIE on UIEB with the four mechanism cues replaced by random or constant per-pixel maps, keeping all other components fixed; if PSNR stays close to 26.88 dB, then cue-guided routing and response alignment are not carrying the claimed specialization, while a large drop would confirm that the missing cue formulas matter.","supporting_citations":[],"review_version":1}