{"id":"d8387534-76b2-482d-988e-dc4d2c7f52fb","arxiv_id":"2506.16353","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MambaHash uses a grouped, multi-directional Mamba backbone for deep supervised hashing and reports the highest mean average precision on CIFAR-10 and IMAGENET, with marginal gains on NUS-WIDE.","lead":"MambaHash converts images into compact codes so similar images get similar codes, enabling fast large-scale image search with a Mamba state space backbone instead of a Transformer or CNN. It is one of the first applications of Mamba to image retrieval and reports accuracy gains over prior hashing methods on standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim rests on NUS-WIDE MAP margins as small as 0.0001 over HybridHash, with no error bars or significance tests and all baselines taken from the authors' prior paper; single-seed differences cannot support 'superior performance'.","rationale":"Good-faith reading: MambaHash is a plausible combination of grouped Mamba scanning, channel interaction attention, and a multi-branch feature enhancement module; the loss is the standard pairwise likelihood plus quantization term; the code link is provided; ablations show each module helps. The central claim, however, is comparative and empirical: MambaHash outperforms SOTA. Everything needed for that claim to hold is that the Table 1 margins are genuine. They are not secured: margins of 0.0001 and 0.0005 on NUS-WIDE are far below typical run-to-run variation for deep hashing MAP, no seeds are reported, and the baseline numbers are not independently reproduced but copied from the authors' own prior paper [20]. The IMAGENET 48-bit result even goes against the headline. This is exactly the concern the reader identified. It is load-bearing because it is the only quantitative evidence for the 'superior performance' half of the central claim; the efficiency half is supported only by backbone FLOPs/parameters and is not what the reader's verdict hinges on. I would keep the conditional verdict: the architecture contribution is real, but acceptance should require the repeated-seed comparison and significance test described above. This read does not change the reader's verdict.","tokens_in":15028,"tokens_out":4841,"duration_ms":51835,"concrete_test":"Reproduce Table 1 with MambaHash and HybridHash under identical data splits, augmentation, optimizer, batch size, and loss, running at least 5 independent seeds per hash-bit setting. Report mean ± std MAP on NUS-WIDE (16/32/48/64) and IMAGENET (48), and compute a paired test over per-query average precision between the two models. If the 95% confidence interval for the difference includes 0 for the 16-bit or 32-bit NUS-WIDE setting, or if the IMAGENET 48-bit point remains negative, the central 'superior performance' claim should be weakened to 'comparable' and the manuscript should report the corrected Table 1 with error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the abstract's claim that MambaHash has 'superior performance' over state-of-the-art hashing methods, Table 1 must show real MAP differences rather than run-to-run noise. The weakest load-bearing condition is that the margins reported against HybridHash, the strongest baseline, are 0.0001, 0.0005, 0.0007, and 0.0014 on NUS-WIDE at 16/32/48/64 bits, and the IMAGENET 48-bit entry is actually lower (0.9087 vs. 0.9094). The paper reports a single run per setting, gives no variance or seed information, and states in Section 4.3 that all baseline numbers are derived from the authors' own prior paper [20], so the comparison inherits any protocol differences in that paper. Without repeated runs or significance testing, the observed differences are as consistent with optimization noise and evaluation-protocol drift as with a genuine architectural advantage. The architecture itself is internally coherent and the ablation in Table 2 shows both CIAM and AFEM contribute, so the residual risk is concentrated in the comparative claim, not in the existence of a working Mamba hashing model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaHash, a deep hashing model built on a visual state space backbone. The backbone uses grouped Mamba scanning in four directions, a channel interaction attention module (CIAM) for cross-channel communication, and an adaptive feature enhancement module (AFEM) that scales channel expansion with hash code length. The model is trained with a pairwise negative log-likelihood loss plus a quantization loss. Experiments on CIFAR-10, NUS-WIDE, and IMAGENET report MAP at 16, 32, 48, and 64 bits, and the authors claim that MambaHash achieves superior performance and efficiency compared with state-of-the-art deep hashing methods.","tokens_in":15270,"tokens_out":3876,"duration_ms":42623,"significance":"If the empirical claims hold, MambaHash would be the first Mamba-based deep hashing model and would show that selective state space backbones can be competitive for large-scale image retrieval while maintaining linear complexity. The architecture is internally coherent, the ablations in Table 2 demonstrate that CIAM and AFEM each contribute, and the authors release source code. However, the central 'superior performance' claim currently rests on single-run comparisons with very small margins against the strongest baseline, and one reported cell is actually below the baseline; the significance is therefore conditional on a substantially strengthened evaluation protocol.","major_comments":[{"comment":"The claim that MambaHash has 'superior performance' over state-of-the-art deep hashing methods is not supported by the reported NUS-WIDE margins against HybridHash: 0.0001, 0.0005, 0.0007, and 0.0014 at 16, 32, 48, and 64 bits. The paper reports a single run, provides no variance or seed information, and states in §4.3 that all baseline results are 'derived from [20]', the authors' own prior paper. Differences of this magnitude are within typical run-to-run noise and also inherit any protocol differences in [20]. Moreover, the IMAGENET 48-bit entry is lower than HybridHash (0.9087 vs. 0.9094), directly contradicting the assertion that MambaHash consistently outperforms. Please report multiple seeds with means and standard deviations, perform significance tests where appropriate, and ideally re-run the strongest baselines under the same protocol.","section":"§4.3, Table 1"},{"comment":"The adaptive channel enhancement rule λ = 2^(μK+b), with μ=1/16 and b=0, is introduced without sensitivity analysis. Since AFEM is presented as a main contribution, the adaptive rule should be validated against fixed ratios or alternative mappings (e.g., a linear mapping) to show that the gain comes from adaptivity rather than simply from increased channel capacity. Please add ablations over μ, b, and the CIAM kernel size k.","section":"§3.4, Eq. (16)"},{"comment":"The hyperparameter written as 'a' in §4.2 is never defined; presumably it is η in Eq. (18), the weight of the quantization loss. No cross-validation range or sensitivity analysis is reported for this hyperparameter, which is necessary to reproduce the trade-off between the pairwise NLL loss and the quantization loss. Please specify the exact hyperparameter and its tuning procedure.","section":"§4.2, Eq. (18)"}],"minor_comments":[{"comment":"There is a typo in the text: 'Sicne' should be 'Since'.","section":"§3.5"},{"comment":"The activation function is written as 'Sigmod' and should be 'Sigmoid'.","section":"§3.3.3, Eq. (11)"},{"comment":"The caption contains 'HybirdHash', which should be 'HybridHash'.","section":"Table 3 caption"},{"comment":"The HybridHash row for IMAGENET is typeset without spaces: '0.88860.90940.9110' should read '0.8886 0.9094 0.9110'.","section":"Table 1"},{"comment":"The hidden feature dimensions are listed as 64, 128, 348, and 512; 348 is unusual and is likely a typo for 384 or 256. Please clarify.","section":"§4.4"},{"comment":"The phrase 'well efficiency and superior performance' is ungrammatical; consider 'good efficiency and superior performance'.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The experimental protocol is the main weakness. All baseline numbers are taken from the authors' own prior paper [20], and the margins over that baseline on NUS-WIDE are at the level of numerical noise. Even if the architecture is sound, the manuscript's central claim of superiority needs to be backed by repeated runs and, ideally, independent re-running of the strongest baselines. The paper also needs careful proofreading for typos and formatting errors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MambaHash is the first Mamba-based deep supervised hashing model I know of, and that alone makes it worth a look. The architecture is coherent: grouped Mamba scanning borrowed from GroupMamba, a channel interaction attention module (a reasonable ECA-style variant), a multi-branch feature enhancement module, and a standard pairwise NLL plus quantization loss. The ablations show each module contributes, and the efficiency numbers are credible: 38.99M parameters and 7.53G FLOPs versus HybridHash's 55.57M and 14.17G, with roughly matching MAP on CIFAR-10 and IMAGENET. That is a genuinely useful datapoint for the retrieval subfield.\n\nThe soft spot is the comparative claim. All baseline numbers come from the authors' own prior paper [20], there are no error bars or multiple seeds, and the NUS-WIDE margins over HybridHash are 0.0001, 0.0005, 0.0007, and 0.0014 across bit lengths. At 48 bits on IMAGENET, MambaHash is actually lower (0.9087 vs 0.9094). Those differences are within run-to-run noise, so the abstract's 'superior performance' language is not supported. The architecture itself is internally coherent and the ablation shows both modules help, so the risk is concentrated in the comparative claim, not in the existence of a working Mamba hashing model. The adaptive enhancement ratio formula (exponential in hash code length, with mu=1/16 and b=0) is a free heuristic with no sensitivity analysis; that is a minor issue but worth flagging. Code is linked but not versioned, so reproducibility is only partial.\n\nWho is this for? People working in deep hashing who want to know whether Mamba backbones are viable. It is a reasonable incremental contribution, not a breakthrough. The central claim needs revision: either soften the superiority language or add repeated runs, significance tests, and independent baseline reproduction. I would send it to peer review — a serious referee can push for those additions — and conditional acceptance seems right, not desk rejection.","headline":"First Mamba-based deep hashing model with a coherent architecture and real efficiency gains, but the 'superior performance' claim rests on single-run margins as small as 0.0001 against the authors' own HybridHash baseline.","tokens_in":15810,"tokens_out":1875,"would_cite":true,"duration_ms":21303,"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":"A state-space deep hashing model, MambaHash, claims the highest reported mean average precision on three image retrieval benchmarks across 16, 32, 48, and 64-bit hash codes.","keywords":["deep hashing","image retrieval","Mamba","visual state space model","channel interaction attention","adaptive feature enhancement","binary hash codes","large-scale image retrieval"],"falsifier":"Re-run MambaHash and HybridHash on NUS-WIDE 16-bit and IMAGENET 16-bit and 48-bit with identical splits and at least five random seeds; if the MAP gaps (0.0001, 0.0105, and -0.0007, respectively) fall within one standard deviation, the claimed superiority over the strongest baseline is not established.","tokens_in":14773,"feed_emoji":"🔎","tokens_out":9759,"duration_ms":96226,"temperature":0.7,"pith_summary":"Large-scale image retrieval needs binary hash codes that are compact enough for fast search and accurate enough to return relevant images. This paper argues that Mamba, a state-space architecture with linear-time scanning, can power a deep hashing model if the network is structured to keep both local and global information. To support that claim, the authors build MambaHash, whose backbone splits channels into four groups, scans each group in a different direction, and adds a channel-interaction attention module and an adaptive feature-enhancement module. Trained with pairwise likelihood and quantization losses, MambaHash reports mean average precision (MAP) that is the highest among the compared methods on CIFAR-10 and NUS-WIDE at all four code lengths and on IMAGENET at 16, 32, and 64 bits, with the 48-bit IMAGENET result within 0.0007 of the best baseline. If the reported margins survive repeated runs, MambaHash would show that linear-complexity state-space models can replace quadratic self-attention in retrieval pipelines.","feed_headline":"MambaHash tops image-retrieval benchmarks with state-space hashing","feed_subtitle":"A linear-complexity Mamba backbone with multi-direction scanning matches or beats transformer-based hashing at lower cost.","key_machinery":"The load-bearing mechanism is the grouped multi-directional Mamba scan. The backbone divides its channels into four groups and runs a VSSS block (the visual selective-scan building block) on each group in a distinct direction — left-to-right, right-to-left, top-to-bottom, bottom-to-top — then concatenates the results, giving global context with linear complexity in sequence length. The channel interaction attention module (CIAM) repairs the cross-channel communication that grouping cuts off: global average pooling feeds one $k\\times k$ 1D convolution for local channel context and one linear layer for global context, and their sum is passed through a sigmoid to score channels. The adaptive feature enhancement module (AFEM) expands the channel dimension by the ratio $\\lambda = 2^{\\mu K + b}$, with $\\mu=1/16$ and $b=0$, applies $1\\times1$, $3\\times3$, and $5\\times5$ depthwise convolutions to capture multi-scale detail, and restores the original width. A hash layer with tanh outputs binary codes, trained by pairwise negative log-likelihood plus a quantization error term.","core_discovery":"The paper's central claim is that Mamba, despite being designed for long-sequence autoregressive tasks, is suitable for image retrieval when the network is built around grouped, multi-directional selective scans. MambaHash is that construction: a stage-wise backbone that splits channels into four groups, scans them left-to-right, right-to-left, top-to-bottom, and bottom-to-top, then reconnects the groups with a channel-interaction attention module and diversifies features with an adaptive enhancement module. Trained on pairwise images with a negative log-likelihood objective plus quantization error, it produces binary hash codes that the authors report as superior to the compared deep supervised hashing methods on CIFAR-10, NUS-WIDE, and IMAGENET. The ablations support the design logic: removing either module lowers MAP everywhere, and the Mamba-only variant remains competitive, which the authors take as evidence that Mamba itself can carry an image retrieval pipeline.","pith_inferences":["The authors take all baseline numbers from their own earlier HybridHash paper rather than re-running them; if the two models are trained in the same code with multiple seeds, the margins on NUS-WIDE (as small as 0.0001 MAP) may not hold, making the efficiency advantage the more durable result.","The reported FLOP and parameter counts compare backbone architectures, not end-to-end retrieval latency; a straightforward extension is to measure query time and memory on a million-scale database, where Mamba's linear complexity should widen the gap.","Because $\\mu$ and $b$ in the adaptive expansion rule were fixed once and applied across all experiments, tuning them per dataset is an obvious extension that could either enlarge or shrink the observed gains."],"forward_implications":["At 48 bits, MambaHash reports 38.99M parameters and 7.53G FLOPs, against 55.57M parameters and 14.17G FLOPs for the hybrid Transformer-CNN baseline, so the linear-complexity backbone is the cheaper option in the comparison.","The ablation results show both added modules pull weight: removing CIAM or AFEM lowers MAP on every dataset and bit length, while the plain Mamba variant still holds its own.","The same architecture ranks first or second on every benchmark configuration in Table 1, which suggests the design transfers across single-label and multi-label retrieval.","The hash-length-dependent expansion rule $\\lambda = 2^{\\mu K + b}$ offers a fixed recipe for scaling feature capacity with code length, avoiding per-bit manual tuning."],"supporting_citations":[{"why":"Supplies the grouped-Mamba layer, four-direction scanning, and VSSS block adopted in the backbone.","marker":"[1]"},{"why":"Defines the retrieval splits and evaluation protocol reused for all three datasets.","marker":"[8]"},{"why":"Introduces Mamba's selective state-space mechanism, whose linear complexity the paper builds on.","marker":"[15]"},{"why":"Is the main baseline and the stated source of every baseline MAP number in Table 1.","marker":"[20]"},{"why":"Contributes the pairwise negative log-likelihood objective used to preserve semantic similarity.","marker":"[31]"},{"why":"Provides the SS1D scan and stage-wise visual Mamba design that the VSSS block is based on.","marker":"[38]"},{"why":"Supplies the ECA channel-attention approach used in the local branch of CIAM.","marker":"[47]"},{"why":"Contributes the pairwise quantization loss that reduces binary-code quantization error.","marker":"[60]"}],"fun_headline_variants":["MambaHash: state-space hashing for faster large-scale image retrieval","Linear-time Mamba backbone elevates deep image hashing","Multi-directional Mamba scans improve image hashing accuracy","State-space hashing with Mamba tops image-retrieval benchmarks","MambaHash: efficient state-space model for scalable image retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported MAP differences over the strongest baseline, several of which are only 0.0001 to 0.005, reflect real model gains rather than run-to-run variation, given that all baseline numbers were copied from an earlier paper and no repeated trials or error bars are reported.","fun_headline_variants_meta":{"raw":{"variants":["MambaHash: state-space hashing for faster large-scale image retrieval","Linear-time Mamba backbone elevates deep image hashing","Multi-directional Mamba scans improve image hashing accuracy","State-space hashing with Mamba tops image-retrieval benchmarks","MambaHash: efficient state-space model for scalable image retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000876,"raw_usage":{"total_tokens":3801,"prompt_tokens":971,"completion_tokens":2830,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":2743}},"tokens_in":587,"tokens_out":2830,"duration_ms":23015,"temperature":1.0,"reasoning_tokens":2743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:44:08.259564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run MambaHash and HybridHash on NUS-WIDE 16-bit and IMAGENET 16-bit and 48-bit with identical splits and at least five random seeds; if the MAP gaps (0.0001, 0.0105, and -0.0007, respectively) fall within one standard deviation, the claimed superiority over the strongest baseline is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the retrieval splits and evaluation protocol reused for all three datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the main baseline and the stated source of every baseline MAP number in Table 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the pairwise negative log-likelihood objective used to preserve semantic similarity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ECA channel-attention approach used in the local branch of CIAM."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the pairwise quantization loss that reduces binary-code quantization error."}],"review_version":1}