{"id":"9dcc753a-be35-48cf-99b6-366bc5b1582a","arxiv_id":"2501.01262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MiJUN is a Mamba-inspired deep unfolding network that reconstructs 3D hyperspectral images from single 2D snapshots, outperforming prior methods in PSNR with fewer parameters.","lead":"This paper introduces MiJUN, a deep learning network that recovers 3D hyperspectral images from a single 2D snapshot by combining Mamba-style state space models with an accelerated unfolding framework. It reports state-of-the-art accuracy on standard benchmarks with about three times fewer parameters and less computation, which could make spectral imaging more practical in resource-limited settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"A-HQS update in Eqs. (4a)-(4c) is circular and Eq. (4b) is ill-posed as written; O(1/k^2) convergence is asserted without proof, so the paper's methodological claim is not reproducible from the text.","rationale":"The reader's weakest-assumption analysis is sound: the A-HQS iteration is the place where the paper's methodology is least secure. My reading agrees that Eqs. (4a)-(4c) are circular as written and that the convergence-rate claim is unsupported. The empirical core of the paper is credible: the comparison protocol follows the CASSI literature, the code is released, the parameter/FLOP counts are plausible, and the ablation study is consistent with the architectural story. However, the advertised 'accelerated' contribution is a central part of the paper's novelty, not a peripheral remark. A conditional verdict is appropriate because the flaw is in the written derivation, not necessarily in the trained system; the concrete test of inspecting the released implementation would distinguish a typographical indexing error from a genuine mismatch between the described algorithm and the deployed network. Since the reader already issued CONDITIONAL, my recommendation is unchanged.","tokens_in":14279,"tokens_out":4661,"duration_ms":40182,"concrete_test":"Download the released MiJUN repository and inspect the iterative update in the model's forward pass. Identify the exact expression used for x_{k+1}: does it use \\hat z_k from the previous stage or \\hat z_{k+1}? Also check the order of updates and whether z_{k+1} is computed as proxNet(x_{k+1}) before \\hat z_{k+1}. If the implementation matches the standard non-circular ordering, rerun the ablation with the corrected equations and verify Table 3 entries; if it matches the paper's circular equations, the network is not an A-HQS unfolding and the convergence claims should be removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest advertised methodological novelty is the 'accelerated unfolding network scheme' A-HQS with O(1/k^2) convergence. As printed, the scheme is not computable. Eq. (4a) defines x_{k+1} using \\hat z_{k+1}; Eq. (4c) defines \\hat z_{k+1} using z_{k+1}; Eq. (4b) defines z_{k+1} using x_{k+1}. This is a three-way circular dependency. Moreover, Eq. (4b) reads z_{k+1}=argmin_z \\mu/2\\|x_{k+1}-z_k\\|^2+\\tau R(z_k), where the optimization variable z appears only as a subscript; as written the objective is constant in z, so the argmin is undefined. The subsequent identification z_{k+1}=proxNet(x_{k+1}) in Eq. (9) is asserted rather than derived from (4b), and the O(1/k^2) rate and second-order ODE interpretation are stated without proof or a reference to a theorem. If the implementation uses the standard corrected ordering (x_{k+1} from \\hat z_k, then z_{k+1}=prox(x_{k+1}), then \\hat z_{k+1}=z_{k+1}+\\beta(z_{k+1}-z_k)), the empirical results are unaffected but the paper's equations need correction. If not, the implemented network differs from the described algorithm. Either way, the methodological contribution as described cannot be verified from the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MiJUN, a deep unfolding network for snapshot spectral compressive imaging (CASSI). The method combines an accelerated half-quadratic splitting (A-HQS) framework with a Mamba-inspired architecture that integrates global-local attention and mode-k tensor unfolding for multi-directional scanning. The authors report state-of-the-art PSNR/SSIM on ten synthetic scenes from the CAVE/KAIST protocols, improved generalization on the ICVL dataset, qualitative gains on real CASSI data, and lower parameter/FLOP counts than prior deep unfolding models. A reference implementation is provided.","tokens_in":14611,"tokens_out":5137,"duration_ms":47889,"significance":"If the empirical results hold, MiJUN is a meaningful step forward for CASSI reconstruction, offering strong accuracy with notably lower complexity than competing DUNs (e.g., Table 1: MiJUN-5stg at 0.56M params achieves 40.60 dB vs. RDLUF-MixS2-9stg at 1.89M params achieving 39.57 dB). The evaluation protocol is standard, the comparisons are comprehensive, and the ablations cover the main components. The code release is a further strength. However, the methodological contribution as written contains serious technical inconsistencies in the derivation of the A-HQS iteration, and the paper's central novelty claims are obscured by these issues.","major_comments":[{"comment":"The A-HQS iteration as written is circular: Eq. (4a) defines x_{k+1} using \\hat{z}_{k+1}; Eq. (4c) defines \\hat{z}_{k+1} from z_{k+1}; and Eq. (4b) defines z_{k+1} from x_{k+1}. In addition, Eq. (4b) has the optimization variable z appearing only in the subscript of z_k, making the objective constant in z and the argmin ill-posed. Consequently, the scheme cannot be executed as stated, and the subsequent identification z_{k+1} = proxNet_{τ,μ}(x_{k+1}) in Eq. (9) is not derived. The formula in Eq. (5) suggests the intended update uses \\hat{z}_k in place of \\hat{z}_{k+1}, and the z-subproblem should be z_{k+1} = argmin_z μ/2 ||x_{k+1} - z||^2 + τ R(z). The authors should correct these equations so that the algorithm is computable and consistent with the implementation in Fig. 3.","section":"Methodology, Eqs. (4a)–(4c) and Eq. (9)"},{"comment":"The paper claims that A-HQS attains O(1/k^2) convergence and can be interpreted as a second-order differential equation, but no proof, statement of assumptions, or reference to a supporting theorem is provided. Given the indexing errors in Eqs. (4a)–(4c), it is unclear what the actual iteration is whose convergence is being claimed. The authors should either prove the rate for the corrected update or cite the specific result they are invoking, and they should justify the second-order ODE interpretation with a derivation.","section":"Abstract and Introduction"},{"comment":"The ablation in Table 3 shows that adding the acceleration strategy alone (w/Acc) improves PSNR by only 0.01 dB over the baseline (38.59 to 38.60), while SSIM improves from 0.969 to 0.971. This marginal gain does not support the paper's strong claim that A-HQS 'effectively eliminates redundant information' and 'accelerates the convergence of iterations.' The authors should provide additional evidence, such as convergence curves, results at different stage counts, or an ablation where the acceleration is combined with each other component separately, to substantiate the role of A-HQS. If the acceleration's effect is negligible, the novelty claim should be revised accordingly.","section":"Ablation study, Table 3"}],"minor_comments":[{"comment":"The summation in Eq. (1) is garbled; it should be written as \\sum_{n=1}^{N_\\lambda}.","section":"Eq. (1)"},{"comment":"The statement that ΦΦ^T is an identity matrix interspersed with zeros is not self-evident for the CASSI model; a concise derivation or an explicit expression would help readers verify the simplification leading to Eq. (8).","section":"Eq. (7) and surrounding text"},{"comment":"There is a typo: 'Mamaba' should be 'Mamba' in the table rows, and 'Mamaba→ M-k Mamaba' should be 'Mamba → M-k Mamba'.","section":"Table 3"},{"comment":"The two subfigures of Fig. 4 are not fully described in the text; the notations (a) and (b) are not explicitly referenced in the paragraph on Mamba-inspired Transformer.","section":"Figure 4"},{"comment":"The claim that tensor mode-k unfolding is integrated into Mamba 'for the first time' should be supported by a citation search; recent works on Mamba for hyperspectral imaging may already use similar unfolding strategies.","section":"Related Work / Method"},{"comment":"The real-data comparison in Fig. 7 refers to 'Scene 4' without explaining how the scene was selected or what its characteristics are; adding context would improve the reproducibility of the qualitative evaluation.","section":"Experiments, Real data"}],"recommendation":"major_revision","confidential_remarks":"The central empirical contribution appears credible: the reported numbers are consistent with the field, the code is provided, and the comparisons are thorough. However, the mathematical formulation of A-HQS is currently not reproducible, and the convergence-rate claim is stated without proof. These issues are fixable in revision, so I recommend major revision rather than rejection. The near-zero effect of the acceleration strategy in the ablation is also worth close attention; if it persists, the authors should moderate their claims about the contribution of A-HQS."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a credible empirical paper for snapshot spectral compressive imaging. The architecture — deep unfolding with a Mamba-style backbone, a global-local attention branch, and mode-k tensor unfolding — is a sensible combination, and the gains on standard benchmarks (roughly 1 dB PSNR over RDLUF with about threefold fewer parameters and FLOPs) are plausible given the ablations. Code is released, and the authors test on synthetic, real, and unseen ICVL data. That is more than most papers in this subfield do.\n\nThe soft spot is the methodology section. Equations (4a)-(4c) are circular as printed: x_{k+1} uses \\hat z_{k+1}, which uses z_{k+1}, which uses x_{k+1}. And (4b) is ill-posed because the optimization variable doesn't appear in the objective. The O(1/k^2) convergence rate and the second-order ODE interpretation are asserted without proof. I believe the implementation uses the standard corrected ordering — x from \\hat z_k, then z = prox(x), then \\hat z = z + beta(z - z_k) — and the empirical results are not affected. But the text as it stands is not reproducible, and the theoretical claims are unsupported. The authors need to fix the indexing and either prove the rate or clearly mark it as a heuristic.\n\nNone of this undercuts the experimental claim. The ablation study shows each component adding something, and the comparisons are fair by the field's standards. The main value is the architecture and the measured performance, not the theory. So I'd send this to peer review, expecting major revision on the theory section. Someone working on SCI will want to cite the empirical comparison. Not a breakthrough, but a useful step.","headline":"MiJUN is a credible empirical extension with released code; the A-HQS equations are circular as printed and the O(1/k^2) claim is unproven, but the experimental contribution justifies peer review.","tokens_in":15165,"tokens_out":3579,"would_cite":true,"duration_ms":33922,"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":"MiJUN reconstructs hyperspectral detail from single 2D snapshots with a Mamba-Transformer hybrid.","keywords":["snapshot spectral compressive imaging","deep unfolding network","Mamba","state space model","tensor mode-k unfolding","hyperspectral image reconstruction","half-quadratic splitting","global-local attention"],"falsifier":"Run the literal iteration from Eqs. (4a)-(4c) with a random initialization and standard CASSI matrices; if no step can be evaluated because $x_{k+1}$ requires $\\hat{z}_{k+1}$ which requires $z_{k+1}$ which requires $x_{k+1}$, or if the reported PSNR gains vanish when the indexing is corrected, the methodological claim of an accelerated convergent scheme collapses.","tokens_in":14040,"feed_emoji":"🌈","tokens_out":6422,"duration_ms":54960,"temperature":0.7,"pith_summary":"The paper proposes MiJUN, a deep unfolding network for snapshot spectral compressive imaging that recovers a 3D hyperspectral image from one 2D coded measurement. MiJUN combines an accelerated half-quadratic splitting scheme, a Mamba-inspired Transformer block with global-local attention, and tensor mode-k unfolding to scan features in twelve directions. On ten simulated scenes, MiJUN-5stg reaches 40.60 dB average PSNR while using 0.56M parameters, outperforming the previous state-of-the-art RDLUF-MixS2-9stg (39.57 dB, 1.89M parameters) and showing qualitative improvements on real datasets. The central claim is that this joint design captures long-range dependencies and local textures at lower computational cost than Transformer-based unfolding networks.","feed_headline":"Mamba-inspired network beats SOTA at one-third the cost","feed_subtitle":"MiJUN reconstructs fine details from single 2D snapshots using tensor mode-k unfolding and 12 scanning directions.","key_machinery":"The load-bearing object is the iterative unfolding scheme A-HQS: Eqs. (4a)-(4c) define an accelerated half-quadratic splitting that alternates a closed-form $x$-update (simplified via the Sherman-Morrison-Woodbury identity to element-wise division using $\\Phi\\Phi^T$) with a learned proximal operator. The proximal operator is the MMB network: (1) Mamba-iT, which fuses a selective state space model branch with a global-local attention branch and is described as a variant of the Transformer, and (2) M-k Mamba, which applies tensor mode-$k$ unfolding along each mode of the 3D HSI to produce 12 scanning directions and emphasize low-rankness before feeding into the SSM. The paper also asserts that A-HQS corresponds to a second-order differential equation with $O(1/k^2)$ convergence.","core_discovery":"MiJUN establishes that a deep unfolding network can beat Transformer-based competitors in CASSI reconstruction while using roughly one-third the parameters and computation. The key is a Mamba-inspired prior extractor (MMB) that acts as the proximal operator in an accelerated half-quadratic splitting loop: the Mamba-iT module reformulates selective state space models as linear attention and adds a global-local attention mechanism, while the M-k Mamba module unfolds the 3D tensor along each mode so that low-rank structure is exposed and twelve scanning sequences are generated. The paper reports that this yields sharper details (hair, bird wings) and higher PSNR/SSIM than prior methods on both synthetic and real measurements, and that the acceleration scheme converges with rate $O(1/k^2)$.","pith_inferences":["The claimed acceleration rests on an iteration whose equations are circular as printed: $x_{k+1}$ depends on $\\hat{z}_{k+1}$, which is built from $z_{k+1}$, which depends on $x_{k+1}$. If the indexing is taken literally, the scheme is not executable; the empirical gains likely come from the learned proximal modules rather than from the acceleration step itself.","The tensor mode-$k$ unfolding idea could transfer to other 3D inverse problems (e.g., video snapshot compressive imaging or multi-spectral medical imaging) where low-rank structure along specific tensor modes is known to exist.","The global-local attention branch that compensates Mamba's local-texture loss suggests that any sequential state space model for images might benefit from an explicit local attention companion, not just a convolutional patch."],"forward_implications":["CASSI hardware could reconstruct higher-quality hyperspectral images in real time, since MiJUN-5stg needs only 0.56M parameters and 40.98 GFLOPs.","The Mamba-as-Transformer reformulation suggests that selective state space models can replace costly $O(N^2)$ attention in image reconstruction without losing long-range context.","Tensor mode-$k$ unfolding gives a principled way to feed 3D image tensors into sequence-based state space models, exposing low-rank structure along each mode.","The accelerated unfolding scheme, if valid, would reduce the number of unfolding stages needed to reach a given accuracy, cutting memory and training time.","Similar joint unfolding designs could be applied to other snapshot compressive imaging problems, such as video or medical spectral imaging."],"supporting_citations":[{"why":"Provides the previous state-of-the-art RDLUF-MixS2 baseline on simulated scenes, the main comparison MiJUN must beat; its parameters and PSNR are the numbers the paper directly outperforms.","marker":"Dong et al. 2023"},{"why":"Introduces the selective state space model (Mamba) that MiJUN reinterprets as a Transformer variant and uses as the backbone of the prior extractor.","marker":"Gu and Dao 2023"},{"why":"Supplies the VMamba cross-scanning strategy (four 1D traversal directions) that MiJUN extends to twelve directions via tensor mode-$k$ unfolding.","marker":"Liu et al. 2024"},{"why":"Provides the linear-attention perspective on Mamba that the paper relies on to justify combining SSM with a global-local attention mechanism.","marker":"Han et al. 2024"},{"why":"Defines the degradation-aware unfolding transformer (DAUHST) used as a baseline and supplies the half-shuffle attention paradigm MiJUN contrasts with.","marker":"Cai et al. 2022c"},{"why":"Provides the pixel-adaptive deep unfolding transformer (PADUT) baseline whose stage count and performance are compared against MiJUN.","marker":"Li et al. 2023"},{"why":"Provides the real CASSI datasets used for real-data experiments and the end-to-end baseline that MiJUN is compared against.","marker":"Meng, Ma, and Yuan 2020"}],"fun_headline_variants":["Mamba-inspired network: sharper details at one-third the cost","MiJUN: Mamba-based unfolding beats SOTA with one-third compute","12-scan Mamba cuts compute and sharpens snapshot spectral imaging","MiJUN: Transformer-like Mamba with 12 scan directions wins in detail","One-third cost, faster convergence, sharper details: MiJUN for spectral imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The A-HQS iteration in Eqs. (4a)-(4c) is assumed to be well-posed and to converge with the stated $O(1/k^2)$ rate, but as written the updates reference each other cyclically, so the acceleration mechanism is not actually demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-inspired network: sharper details at one-third the cost","MiJUN: Mamba-based unfolding beats SOTA with one-third compute","12-scan Mamba cuts compute and sharpens snapshot spectral imaging","MiJUN: Transformer-like Mamba with 12 scan directions wins in detail","One-third cost, faster convergence, sharper details: MiJUN for spectral imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":2145,"prompt_tokens":981,"completion_tokens":1164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1066}},"tokens_in":597,"tokens_out":1164,"duration_ms":9779,"temperature":1.0,"reasoning_tokens":1066,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:31:08.557129+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the literal iteration from Eqs. (4a)-(4c) with a random initialization and standard CASSI matrices; if no step can be evaluated because $x_{k+1}$ requires $\\hat{z}_{k+1}$ which requires $z_{k+1}$ which requires $x_{k+1}$, or if the reported PSNR gains vanish when the indexing is corrected, the methodological claim of an accelerated convergent scheme collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the previous state-of-the-art RDLUF-MixS2 baseline on simulated scenes, the main comparison MiJUN must beat; its parameters and PSNR are the numbers the paper directly outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the real CASSI datasets used for real-data experiments and the end-to-end baseline that MiJUN is compared against."}],"review_version":1}