{"id":"68c0dad7-4d7c-45b8-84a0-e9b417b0c267","arxiv_id":"2412.17366","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FlowMamba couples a bidirectional state space model with a feature-induced learned ordering to estimate point cloud scene flow, achieving 21.9% and 20.5% EPE3D reductions on FlyingThings3D and KITTI.","lead":"This paper introduces FlowMamba, a deep network that estimates 3D motion between two point clouds by propagating motion information globally through a state space model with a learned point ordering. It reports roughly 20% lower endpoint error than prior best methods on standard scene flow benchmarks, a meaningful gain for autonomous driving and robotics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FIO's learned ordering is non-differentiable as described: sorting by Eq. (6) yields zero gradient to the score MLP, so the ordering may be untrained, and the paper gives no soft-sort or straight-through estimator.","rationale":"The reader's weakest_assumption and my load-bearing concern coincide: the paper describes a sorting-based ordering without explaining how gradients reach the score MLP. This is a genuine omitted-method detail that is load-bearing because the FIO strategy is presented as a learned component and is credited for the improvements in Table 5. If the MLP is untrained, the ordering is a fixed random feature-dependent permutation, and the claimed 'feature-induced ordering' as a learned strategy does not exist. The concern is not that the reported numbers are impossible; the ablation tables and relative improvements are arithmetically consistent, and the plug-and-play experiments in Table 3, despite the outlier-column inconsistency, are encouraging. But without code or a clear differentiable surrogate, the mechanism cannot be verified. I do not see a reason to reject outright, because the authors may well have used a straight-through estimator or a soft-sort and simply omitted it from the paper; the correct verdict remains CONDITIONAL, requiring the authors to specify the gradient handling and release code. I also note a secondary factual issue: the abstract claims 'the first method to achieve millimeter-level precision on both datasets,' yet Table 1 shows DifFlow3D already reports 0.0078 m on KITTI, which is 7.8 mm; this overstatement should be corrected but is not the central technical concern. Overall, the conditionality of the reader's verdict is appropriate, and my analysis does not change it.","tokens_in":13262,"tokens_out":6249,"duration_ms":62625,"concrete_test":"Request or reimplement the code and inspect the backward pass through the sorting operation. Concretely, insert a hook on the scores from Eq. (6) and check whether the gradient magnitude is nonzero after the first training batch; if it is zero, the MLP is not trained. As an independent behavioral test, retrain FlowMamba on FlyingThings3D under the full protocol but detach the scores before sorting (stop-gradient to the MLP). If the final EPE3D stays near the reported 0.0089, the FIO improvement does not come from a learned ordering; if it degrades materially, the paper must specify the exact differentiable mechanism used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of the FIO strategy is under-specified to the point of likely being untrainable. In Section 3 ('Feature-induced Ordering Strategy'), Eq. (6) computes Score = tanh(MLP([cfl, mfl,n, hl,n−1])) and the points are sorted by this score to define the sequence fed into the bidirectional Mamba blocks. Sorting is a discrete operation: its gradient with respect to the scores is zero almost everywhere, and a native PyTorch implementation using torch.argsort would return integer indices and back-propagate no gradient to the MLP. The paper never describes a differentiable surrogate such as a soft sort, a perturbed optimizer, or a straight-through estimator. If no such mechanism is used, the score MLP is never trained, so the ordering is not 'learned' in any meaningful sense. This directly undermines the paper's claim that FIO 'leverages semantic-related and motion-related features' to construct causal dependencies, and it invalidates the interpretation of the Table 5 ablations: removing context, correlation, or hidden features changes the scores of a randomly initialized MLP, so the observed EPE3D differences do not demonstrate learned feature-induced ordering. The central claim that global motion propagation with a learned ordering yields state-of-the-art performance therefore rests on an unspecified and likely missing gradient path.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FlowMamba proposes a supervised point-cloud scene-flow network built on the MSBRN coarse-to-fine backbone, replacing the GRU update with an Iterative SSM-based Update (ISU) module that uses bidirectional Mamba blocks to propagate global motion information, and adding a Feature-Induced Ordering (FIO) strategy that sorts points by a learned score before sequence modeling. On FlyingThings3D and KITTI the method reports state-of-the-art EPE3D values (0.0089 and 0.0062 m, reductions of 21.9% and 20.5% over DifFlow3D), with additional experiments showing that ISU improves existing iterative methods and ablations of the ISU and FIO components.","tokens_in":13553,"tokens_out":4461,"duration_ms":43363,"significance":"If the results are reproducible, the paper makes a solid empirical contribution: it is, to my knowledge, the first application of bidirectional Mamba/SSM sequence modeling to point-cloud scene flow, and the reported gains are large and consistent across both synthetic and real-world benchmarks. The paper includes extensive comparisons (Tables 1 and 2), ablations (Tables 4 to 6), a runtime comparison (Table 7), and a plug-and-play study (Table 3). The arithmetic in the headline EPE3D reductions checks out. The two issues below -- the unspecified gradient path for the learned ordering and the inconsistent Outliers values in Table 3 -- are load-bearing and need to be fixed before the claims can be fully trusted.","major_comments":[{"comment":"The FIO strategy computes Score = tanh(MLP([cfl, mfl,n, hl,n-1])) and then sorts all points by this score to define the sequence fed into the bidirectional Mamba blocks. Sorting is a discrete, non-differentiable operation: a naive implementation (e.g., torch.argsort) gives zero gradient to the score MLP. The paper does not describe any differentiable surrogate such as a soft sort, a perturbed optimizer, or a straight-through estimator, nor does it state that gradients are stopped or that the permutation is treated as a fixed preprocessing step. This matters because the FIO ablations in Table 5 are interpreted as evidence that the ordering is learned from the three features; with an untrained score MLP, the differences between ablations would reflect fixed random permutations rather than a learned feature-induced ordering. The authors must specify the exact gradient path used in training, or present evidence (e.g., before/after training analysis of the scores) that the ordering is actually learned.","section":"Section 3, 'Feature-induced Ordering Strategy', Eq. (6)"},{"comment":"The reported Outliers value for MSBRN on FlyingThings3D, 0.0243, duplicates the Ours value in Table 1 for the same metric and contradicts Table 1's MSBRN Outliers value of 0.0560. Moreover, MSBRN(w/ISU) shows Outliers 0.0892 on FlyingThings3D, which is worse than both possible baselines (0.0560 or 0.0243), even though the same row reports a 28.6% EPE3D improvement. This inconsistency directly undermines the claim that ISU improves baseline accuracy across all metrics. Please correct the table and re-evaluate the plug-and-play claim for the Outliers metric on FlyingThings3D; if the improvement is not consistent, the claim should be qualified accordingly.","section":"Table 3 (Universality of proposed modules)"}],"minor_comments":[{"comment":"The conclusion states that the proposed method shows 'a strong university'; this should read 'universality'.","section":"Conclusion"},{"comment":"In Eq. (2), the symbol B is used both for the continuous projection matrix and for the discretized matrix; using \\bar{B} or another symbol for the discretized version would avoid confusion.","section":"Preliminaries, Eq. (2)"},{"comment":"There are several grammatical issues, for example 'There has been a significantly increasing focus' and 'the same number of iterations as reporting in (Cheng and Ko 2023)'; these should be fixed in a polish pass.","section":"Abstract and Introduction"},{"comment":"The FIO ablation differences are small (EPE3D ranging from 0.0096 to 0.0089); reporting error bars or multiple seeds would make the conclusions more robust.","section":"Table 5"},{"comment":"Since the FIO gradient path is central to the method, a short pseudocode block or a statement about the differentiable sorting mechanism would improve reproducibility.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The two major issues -- the missing gradient description for FIO and the Table 3 Outliers inconsistency -- are the main barriers. If the authors can clarify the training procedure and correct the table, I would support publication. I recommend asking for code release or a precise implementation description to verify the FIO gradient path."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know about this paper? It's the first to apply bidirectional Mamba/SSM to point cloud scene flow, and it delivers: 21.9% and 20.5% EPE3D reduction over DifFlow3D on FlyingThings3D and KITTI, with consistent gains across metrics and a clean plug-and-play ablation replacing GRUs with their ISU module. The core design—bidirectional SSM for global motion propagation plus a feature-induced ordering (FIO) to make the point sequence more spatially coherent—is genuinely new in this task. If you work on scene flow, the architecture is worth understanding.\n\nThe soft spots are real but not fatal. The FIO sorts points by a score from an MLP (Eq. 6), and the paper never says how gradients pass through the sort. Sorting is non-differentiable; with a native argsort the score MLP would get zero gradient, so the ordering would be a function of whatever the input features happen to be, not a trained ordering. The paper lacks any mention of soft-sort or straight-through estimators. That doesn't necessarily sink the method—the features feeding the score come from trained modules, so the ordering still evolves during training—but it does undercut the claim that FIO 'learns' causal dependencies, and it makes the Table 5 ablations hard to interpret. Removing a feature changes the input to the score MLP, so the EPE differences don't cleanly demonstrate that the ordering was learned. This is the biggest issue and it needs an explicit answer from the authors.\n\nSecond, Table 3 has a clear inconsistency: MSBRN's Outliers on FlyingThings3D is 0.0560 in Table 1 but 0.0243 in Table 3, and MSBRN(w/ISU) jumps to 0.0892. That's a copy-paste or labeling error, not a load-bearing flaw, but it makes you trust the numbers a bit less. Also, no code is released, so independent verification is impossible right now.\n\nThe arithmetic in the headlined gains checks out, and the ablations are otherwise internally consistent. The 'millimeter-level' claim is a metric-units framing artifact, not a technical achievement—0.0089 m is 8.9 mm, which is nice but not a different regime.\n\nWho is this for? Anyone working on point cloud scene flow or SSM applications to 3D understanding. It deserves a serious referee: the empirical result is strong enough that the FIO gradient question and table typo should be resolved, not desk-rejected. My recommendation: send it to review, but require the authors to specify the gradient path through the ordering (or provide code), and fix Table 3.","headline":"FlowMamba delivers strong scene flow gains via bidirectional Mamba, but the FIO ordering's gradient path is unspecified and Table 3 has a clear number bug.","tokens_in":14072,"tokens_out":3484,"would_cite":false,"duration_ms":31436,"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":"FlowMamba claims that replacing local recurrent updates with a bidirectional state-space model guided by a learned point ordering cuts point-cloud scene flow error by roughly a fifth and reaches sub-centimeter accuracy on standard…","keywords":["scene flow","point cloud","state space model","Mamba","global motion propagation","iterative update","learned ordering","ill-posed regions"],"falsifier":"Train FlowMamba with FIO replaced by a fixed, non-learned ordering such as lexicographic sorting by 3D coordinates, or a random permutation, keeping all else fixed; if the EPE3D on FlyingThings3D remains at 0.0089 meters, the learned ordering is not the source of the improvement.","tokens_in":13058,"feed_emoji":"🎯","tokens_out":4913,"duration_ms":43264,"temperature":0.7,"pith_summary":"This paper proposes FlowMamba, a network that estimates point-cloud scene flow—the 3D motion of each point between two frames—by propagating global motion information through a bidirectional state space model instead of relying only on local neighborhoods. The central claim is that its Iterative SSM-based Update module (ISU) resolves the ambiguity of flat, occluded, and slender regions where local evidence is insufficient, and that its Feature-Induced Ordering (FIO) makes the point sequence spatially coherent so the SSM can actually propagate information. On FlyingThings3D and KITTI the method reports EPE3D of 0.0089 and 0.0062 meters, reductions of 21.9% and 20.5% over DifFlow3D, and the first millimeter-level accuracy on both benchmarks. The authors also claim the ISU is a plug-and-play replacement for GRU update cells in existing iterative scene flow networks, improving their accuracy as well.","feed_headline":"Global motion propagation nets first millimeter-level scene flow","feed_subtitle":"FlowMamba cuts EPE3D by 21.9% on FlyingThings3D and 20.5% on KITTI.","key_machinery":"Two components carry the argument. The ISU (Iterative SSM-based Update) takes context, motion, and hidden features, orders the points according to a learned score, applies several bidirectional Mamba blocks to the sorted sequence, and adaptively fuses the output with the previous hidden state via a sigmoid gate—this is what propagates global motion patterns. The FIO (Feature-Induced Ordering) computes Score = tanh(MLP[cfl, mfl,n, hl,n-1]) for each point and sorts all points by that score, aiming to turn the irregular point cloud into a sequence with spatial continuity so that the SSM's causal dependencies are meaningful.","core_discovery":"The paper's core discovery, on its own terms, is that scene flow accuracy in ill-posed regions is limited by the local receptive field of the recurrent update cell, and that a bidirectional structured state space model with a learned ordering can supply the missing global context. FlowMamba replaces the GRU in a coarse-to-fine MSBRN-style architecture with an ISU that sorts points by a learned score, passes hidden states through bidirectional Mamba blocks, and fuses the globally refined hidden state with the previous iteration's state using a learned gate. This combination achieves state-of-the-art end-point error on both the synthetic FlyingThings3D and real-world KITTI benchmarks without fine-tuning on KITTI, and the ablation table attributes the gain specifically to the ISU and FIO components.","pith_inferences":["We infer, though the paper does not test it, that the FIO style of learned ordering could transfer to other point-cloud sequence tasks such as object segmentation or completion, where spatial continuity in a scan order may similarly help SSMs.","An implication left implicit is that any cheap spatial ordering that already encodes locality might capture part of FIO's gain; ablating against such orderings would isolate what the learned order adds.","The claim that ISU propagates global information could be tested directly by probing how much the hidden state of a distant point changes after one ISU pass, a diagnostic the paper does not report.","If the ordering is not trainable end-to-end, the observed gains might come from the Bi-Mamba block alone on a fixed order; the paper's own unidirectional-Mamba ablation suggests directionality matters more than the score."],"forward_implications":["If correct, global sequence-based motion propagation is a viable fix for ill-posed scene flow regions, not just local cost volumes.","Iterative scene flow methods can drop in the ISU in place of GRU and gain accuracy without retraining from scratch.","Millimeter-level accuracy on KITTI without fine-tuning implies synthetic-to-real generalization is strong enough for practical deployment.","Mamba's linear complexity makes the global receptive field affordable on 8k-point clouds.","Fewer iterations reach baseline accuracy, suggesting faster inference options."],"supporting_citations":[{"why":"Base coarse-to-fine architecture that FlowMamba extends and the baseline for the plug-and-play experiment.","marker":"(Cheng and Ko 2023)"},{"why":"Defines the Mamba selective state space model that the ISU's Bi-Mamba blocks are built on.","marker":"(Gu and Dao 2023)"},{"why":"DifFlow3D, the previous state-of-the-art whose EPE3D is reduced by 21.9% and 20.5%.","marker":"(Liu et al. 2024a)"},{"why":"Provides the FlyingThings3D synthetic training and test set.","marker":"(Mayer et al. 2016)"},{"why":"Provides the KITTI real-world benchmark used for zero-shot generalization.","marker":"(Geiger et al. 2013)"},{"why":"Source of the bidirectional SSM design used in the Bi-Mamba blocks.","marker":"(Zhu et al. 2024)"}],"fun_headline_variants":["FlowMamba: first millimeter-level scene flow via global propagation","Global motion state-space model achieves first mm-level scene flow","FlowMamba cuts scene flow error by 21.9% and 20.5% on two benchmarks","Scene flow goes global: FlowMamba's state-space unit achieves record accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ordering is claimed to be learned, but the paper does not explain how gradients pass through the discrete sorting step to train the MLP that produces the scores, so the FIO may not actually be learning anything.","fun_headline_variants_meta":{"raw":{"variants":["FlowMamba: first millimeter-level scene flow via global propagation","Global motion state-space model achieves first mm-level scene flow","FlowMamba cuts scene flow error by 21.9% and 20.5% on two benchmarks","Scene flow goes global: FlowMamba's state-space unit achieves record accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00067,"raw_usage":{"total_tokens":3048,"prompt_tokens":932,"completion_tokens":2116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2033}},"tokens_in":548,"tokens_out":2116,"duration_ms":13802,"temperature":1.0,"reasoning_tokens":2033,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:32:24.861254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FlowMamba with FIO replaced by a fixed, non-learned ordering such as lexicographic sorting by 3D coordinates, or a random permutation, keeping all else fixed; if the EPE3D on FlyingThings3D remains at 0.0089 meters, the learned ordering is not the source of the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Base coarse-to-fine architecture that FlowMamba extends and the baseline for the plug-and-play experiment."}],"review_version":1}