{"id":"bf54cbba-f0bf-46e6-aae8-5087dd7bbd05","arxiv_id":"2411.15380","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Nd-BiMamba2 is a single architecture that flattens 1D, 2D, and 3D inputs and runs forward and backward Mamba2 passes to capture bidirectional context.","lead":"This paper proposes one neural network design, Nd-BiMamba2, meant to process 1D time series, 2D images, and 3D volumes with a bidirectional version of the Mamba2 model. A generalist might read it to see a lightweight, deployable alternative to specialized architectures, though the paper reports only speed and model size, not task accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No task-level accuracy is reported: the claims that bidirectional modeling and adaptive padding improve feature extraction rest only on FLOPs/timing tables, so the central functional claim is empirically unsupported.","rationale":"The reader's weakest assumption concerns whether flattening preserves spatial structure. I agree that this is the key technical risk, but I would sharpen it: the paper provides no task-level accuracy experiment at all, so even the preconditions for evaluating that assumption are absent. The runtime and export experiments are real evidence that the code runs on multiple platforms, which supports the deployment claims. However, the central functional claims, especially bidirectional feature extraction and preservation of multi-dimensional structure, are supported only by FLOPs, timing, and padding-size tables. The paper also has internal inconsistencies, such as Section III being empty and the convolution formulas in Section IV-B-2 not appearing in Algorithm 1, but these are secondary. The decisive issue is the missing accuracy evaluation. The appropriate verdict remains REJECT; no adjustment is needed, hence UNCHANGED.","tokens_in":8373,"tokens_out":4533,"duration_ms":43357,"concrete_test":"Run a controlled CIFAR-10 classification benchmark with matched d_model and parameter counts: (1) Nd-BiMamba2 exactly as in Algorithm 1, (2) a unidirectional control with the backward path removed, and (3) an order-shuffled control in which the 2D flattening in Step 1 uses a fixed random permutation. Train all three with identical schedules and three seeds, and report top-1 accuracy. If (2) matches (1), the bidirectional addition contributes no measurable feature-quality benefit; if (3) matches (1), the specific scan ordering is irrelevant and the model is not exploiting spatial structure. This directly tests the missing evidence under the weakest assumption about flattening.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Nd-BiMamba2 efficiently handles 1D, 2D, and 3D data while capturing bidirectional information. The architecture in Algorithm 1 is simple: pad, reshape/flatten, map channels, pass through forward and backward Mamba2, sum, map back, and trim. This construction will run, and the open-source code plus ONNX/TorchScript exports do support the portability claim. What is not established is that the resulting features are useful. Section V-B states that 'comparative experiments across different data dimensions' validated an improvement in feature representation from bidirectional modeling, but Table III contains only FLOPs, time, and parameter counts, not accuracy on any task. Section V-C claims adaptive padding 'maintains high accuracy,' but Table IV contains only padded sizes. No benchmark, no downstream task, no comparison to unidirectional Mamba2, Vision Mamba, BiLSTM, or any baseline, and no error bars appear anywhere. Because the 'flatten to 1D' step (Section IV-B-1, Algorithm 1 Step 1) determines what spatial context the SSM can see, the usefulness of the architecture hinges on that ordering; without accuracy experiments one cannot distinguish a model that exploits 2D/3D structure from one that merely flattens it away. The authors characterize the bidirectionality result as validation, but the evidence offered does not measure feature quality. This gap is load-bearing because the claimed advantage of Nd-BiMamba2 over existing Mamba variants is architectural effectiveness, not just runtime.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Nd-BiMamba2, a unified architecture based on the Mamba2 state-space model, intended to process 1D, 2D, and 3D data by flattening inputs, running forward and backward Mamba2 modules, fusing their outputs, and undoing padding. The authors report computational cost (FLOPs, time, parameters) across 1D/2D/3D sizes, demonstrate ONNX/TorchScript export, and test on CPU/GPU; code is open-source. The central claim is that this bidirectional design captures multi-dimensional information efficiently.","tokens_in":8612,"tokens_out":7619,"duration_ms":64097,"significance":"If the architecture were shown to improve task-level accuracy on 1D/2D/3D benchmarks, it would be a useful contribution given its modular design and cross-platform deployment. The open-source code and export support are strengths for reproducibility. However, the paper as written provides no accuracy experiments; the claimed feature-representation improvement rests solely on increased FLOPs and runtime, which by themselves do not establish utility. The related work also incorrectly states that existing Mamba modules lack bidirectionality while citing bidirectional variants. Thus the scientific contribution is currently limited to a deployment demo, not a validated model.","major_comments":[{"comment":"The claim that 'comparative experiments... validated the improvement in feature representation by bidirectional modeling' is unsupported: Table III reports only FLOPs, time, and parameter counts, with no accuracy measure. Increased FLOPs, time, and an 89% parameter increase (150.8k -> 285.21k) are costs, not evidence of improved feature quality. Please add task-level evaluations (e.g., classification or regression on 1D, 2D, and 3D datasets) comparing Nd-BiMamba2 against unidirectional Mamba2 and existing bidirectional Mamba baselines, with standard metrics and error bars.","section":"V-B, Table III"},{"comment":"The statement that adaptive padding 'maintains high accuracy across different input sizes' is not supported by Table IV, which lists only input dimensions, auto-padded sizes, and a boolean indicator of whether the padding matches Mamba2's equal padding. No accuracy is reported anywhere in the paper. Please either remove accuracy claims or provide the corresponding measurements.","section":"V-C, Table IV"},{"comment":"The related-work section asserts that 'existing Mamba modules primarily focus on unidirectional feature modeling' and lack bidirectional modeling, yet the same section cites [20] (VSSD, non-causal state space duality) and [22] (Vision Mamba, bidirectional state space model). This internal contradiction overstates the novelty of Nd-BiMamba2. The manuscript should accurately position its contribution relative to these existing bidirectional Mamba works and empirically compare against them.","section":"II-D"},{"comment":"The design flattens 2D/3D tensors into a 1D sequence (Section IV-B-1, Algorithm 1 Step 1), which determines the spatial context available to the state-space model. The paper does not justify that this flattening preserves useful spatial structure, nor does it test alternative scan orders. Without accuracy experiments, the reader cannot tell whether the model exploits multi-dimensional structure or merely discards it. Please include ablations over scan orders and comparisons with 2D/3D-aware Mamba methods.","section":"IV-B-1 and Algorithm 1 Step 1"}],"minor_comments":[{"comment":"Section III ('Algorithm Design') appears to be empty; the text jumps from Section II directly to Section IV. Please either merge the sections or add the missing content.","section":"III"},{"comment":"The unit 'GMac' in the FLOPs column denotes multiply-accumulate operations, not floating-point operations; please use a consistent metric (e.g., MACs or FLOPs).","section":"Table III"},{"comment":"The text says the parameter-count growth is 'relatively small,' but the number of parameters nearly doubles from 150.8k to 285.21k (about 89% increase); please revise this characterization.","section":"V-B"},{"comment":"The padding formula as typeset is ambiguous: it should be \\max(0, \\lceil ((D_i - 1)s_i + k_i - 1)/2 \\rceil) or similar. Please clarify the numerator and parentheses.","section":"Equation (7)"},{"comment":"There are several typos, including 'nd-BiMamba2' (lowercase) in Section V-D, 'B IMAMBA2' in the index terms, and the unrendered section headers. A thorough proofread is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a short technical note with a useful open-source implementation, but its scientific claims are not yet backed by task-level evidence. The paper overstates novelty, given the existing bidirectional Mamba works that it cites. For a general machine-learning journal, I would require substantial revision with real benchmarks; the paper may be better suited to a systems or software-track venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is exactly what it looks like: a straightforward wrapper that flattens 1D/2D/3D tensors, runs forward and backward Mamba2, adds the two passes, and trims padding. The math is standard, the code is open source, and the ONNX/TorchScript exports with CPU/GPU/mobile timing tables support the portability claim. That part is real, and it deserves credit.\n\nThe soft spot is load-bearing and you already spotted it. Section V-B says bidirectional modeling was “validated” to improve feature representation, but Table III only lists FLOPs, time, and parameters. Higher FLOPs do not mean better features; interpreting them that way is a category error. Section V-C claims adaptive padding “maintains high accuracy,” yet Table IV reports padded sizes only. There is no benchmark, no downstream task, no baseline (unidirectional Mamba2, Vision Mamba, BiLSTM, anything), and no error bars. The related work also contradicts itself: it says existing Mamba modules lack bidirectionality while citing Vision Mamba and VSSD, both of which are explicitly bidirectional or non-causal.\n\nThe flattening ordering in Section IV-B-1 and Algorithm 1 is the hinge of the whole approach, and the paper never tests whether that ordering preserves spatial structure. This is not a minor omission; without accuracy experiments you cannot distinguish a model that exploits 2D/3D structure from one that flattens it away and hopes for the best.\n\nAll that said, the paper is not incoherent. The architecture is described clearly, the equations are correct, and the code exists. The problem is that the authors overclaim what the experiments show. This would be an acceptable workshop systems note or a code-release README, but as a research paper claiming improved feature representation, it does not support its central conclusion.\n\nWho gets value? Someone looking for a reusable Mamba2 multi-dimensional building block might use the code. A reader interested in evidence standards might use this as a case study in why FLOPs are not accuracy. For peer review: I would desk reject the current version. The fix is not minor—the authors would need to run actual tasks (image classification, segmentation, time-series forecasting) and compare against unidirectional and bidirectional baselines. Then it could become a modest but useful engineering contribution.","headline":"A clean, honest engineering wrapper for bidirectional Mamba2 across 1D/2D/3D with shipped code and portability tests—but the claim that it improves feature representation rests on a category error: FLOPs and timing are treated as evidence of accuracy, and no task-level experiment appears anywhere.","tokens_in":9165,"tokens_out":2215,"would_cite":false,"duration_ms":21703,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single Mamba2-based architecture, Nd-BiMamba2, processes 1D, 2D, and 3D data with bidirectional context and deploys across CPU, GPU, and mobile.","keywords":["Mamba2","state space models","bidirectional modeling","multi-dimensional data processing","adaptive padding","ONNX","TorchScript","cross-platform deployment"],"falsifier":"Train Nd-BiMamba2 and a standard 2D-specific model on the same image classification benchmark, such as CIFAR-10, with matched budgets, and also train Nd-BiMamba2 with row-major versus randomly shuffled flattening orders; if accuracy is similar under shuffling, or far below the 2D-specific baseline, the flatten-then-scan design is not capturing spatial geometry.","tokens_in":8143,"feed_emoji":"🔁","tokens_out":9566,"duration_ms":77987,"temperature":0.7,"pith_summary":"Nd-BiMamba2 proposes a unified architecture built on the Mamba2 state-space module for processing 1D time series, 2D images, and 3D volumes with the same code path. The design runs a forward Mamba2 pass and a backward Mamba2 pass over the reversed input, adds the two feature maps, and uses padding sizes adapted to the input dimension so that shapes stay regular. The paper argues this captures bidirectional context in higher-dimensional data while keeping the linear-time efficiency of state-space models and avoiding the quadratic cost of attention. If correct, a single model family could replace separate CNN, RNN, and Transformer designs for different data dimensions and be exported to ONNX and TorchScript for CPU, GPU, and mobile deployment.","feed_headline":"Bidirectional Mamba2 variant unifies 1D, 2D, and 3D processing","feed_subtitle":"Forward and backward Mamba2 passes plus adaptive padding aim to unify time series, images, and volumes.","key_machinery":"The central object is the bidirectional Mamba2 block: a forward Mamba2 pass over the flattened input and a second pass over the reversed input, with the two hidden representations added and then reordered. Mamba2 is a selective state-space layer that models sequences in linear time; running it in both directions is what supplies bidirectional context. The adaptive padding step, with padding size computed from kernel size and stride per dimension, keeps tensors aligned to fixed multiples so the same convolution and state-space code can handle 1D, 2D, and 3D inputs. The flatten-and-reshape step is the load-bearing device that lets one recurrence sweep across all dimensions.","core_discovery":"The central claim is that the Mamba2 module can be extended into a general bidirectional engine for data of any rank by uniformly representing inputs as a tensor with dummy dimensions for missing axes, flattening the tensor to a sequence, running separate forward and backward Mamba2 paths, and fusing them by addition. Nd-BiMamba2 wraps this with dimension-adaptive convolution kernels—$(k,1,1)$ for 1D, $(k_1,k_2,1)$ for 2D, $(k_1,k_2,k_3)$ for 3D—and an adaptive padding rule that aligns each dimension to a multiple of the model's block size (4 for 1D and 3D, 8 for 2D). The output is trimmed back to the original shape, making the block a drop-in layer for inputs of different dimensionalities. The paper's evidence is computational and portability focused: it reports FLOPs, latency, and parameter counts on CPU, GPU, and mobile, with bidirectional processing roughly doubling FLOPs and parameters while the layer count remains fixed.","pith_inferences":["The flattening order, whether row-major, snake, or Hilbert-style, is an implicit design choice that could matter more than the bidirectional fusion itself, and the paper does not test it.","The same forward-plus-backward-then-add recipe could be applied to other linear-time sequence models, such as linear attention, to isolate whether the benefit comes from bidirectionality or from Mamba2's state compression.","If the architecture's generality holds, dimensionality could become a hyperparameter rather than an architectural commitment, allowing one pretrained module to be adapted across audio, image, and volumetric tasks."],"forward_implications":["One architecture can serve the three most common data ranks, so a project spanning time series, images, and volumes no longer needs separate model families.","The block preserves input and output shapes, so it can slot into existing networks at any layer without dimension-specific rewiring.","Because the model exports to ONNX and TorchScript, the same trained weights can run in production on CPUs, GPUs, and mobile devices.","Bidirectional processing increases FLOPs and computation time about twofold, but the added parameter count is independent of input size.","Adaptive padding reduces memory waste compared with fixed equal padding, particularly for non-square 2D and non-cubic 3D inputs."],"supporting_citations":[{"why":"Supplies the Mamba2 selective state-space module that Nd-BiMamba2 wraps with forward and backward passes.","marker":"[19]"},{"why":"Establishes Mamba as a linear-time selective state-space sequence model, the efficiency basis the architecture inherits.","marker":"[5]"},{"why":"Demonstrates bidirectional state-space scanning for visual representation, the pattern Nd-BiMamba2 extends to N dimensions.","marker":"[22]"},{"why":"Shows a non-causal state-space duality variant for vision, supporting the move from unidirectional to bidirectional Mamba processing.","marker":"[20]"},{"why":"Provides the bidirectional LSTM baseline whose parallelization and deployment difficulties motivate the new design.","marker":"[11]"},{"why":"Supplies the self-attention mechanism whose quadratic complexity Nd-BiMamba2 aims to avoid in high-dimensional settings.","marker":"[4]"}],"fun_headline_variants":["Bidirectional Mamba2 tames 1D, 2D, and 3D data","Unified Mamba2 bidirectional layer for 1D, 2D, 3D","One architecture for time series, images, and volumes","Mamba2 bidirectional fusion across data dimensions","Nd-BiMamba2: unified bidirectional Mamba for any data rank"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that flattening 2D and 3D tensors into 1D sequences, in the order the code chooses, preserves enough spatial structure for the state-space model to learn useful features.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional Mamba2 tames 1D, 2D, and 3D data","Unified Mamba2 bidirectional layer for 1D, 2D, 3D","One architecture for time series, images, and volumes","Mamba2 bidirectional fusion across data dimensions","Nd-BiMamba2: unified bidirectional Mamba for any data rank"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000621,"raw_usage":{"total_tokens":2910,"prompt_tokens":1008,"completion_tokens":1902,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":1804}},"tokens_in":624,"tokens_out":1902,"duration_ms":12745,"temperature":1.0,"reasoning_tokens":1804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:21:23.732280+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Nd-BiMamba2 and a standard 2D-specific model on the same image classification benchmark, such as CIFAR-10, with matched budgets, and also train Nd-BiMamba2 with row-major versus randomly shuffled flattening orders; if accuracy is similar under shuffling, or far below the 2D-specific baseline, the flatten-then-scan design is not capturing spatial geometry.","supporting_citations":[{"cited_title":"Bidirectional lo ng short- term memory networks for relation classification,","cited_arxiv_id":null,"evidence_quote":"Provides the bidirectional LSTM baseline whose parallelization and deployment difficulties motivate the new design."}],"review_version":1}