{"id":"81c1e580-91f3-4bde-b8e5-9c6361650b7d","arxiv_id":"2411.12603","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A state-space model whose step sizes are coordinate differences improves point cloud classification from scratch and achieves 100 percent accuracy on DVS128 Gestures for one seed.","lead":"STREAM is a state-space model that encodes the gaps between neighboring points or successive events directly into its step-size parameter, letting sparse geometric data be processed as a sequence without conversion to frames. The authors report an improved point-cloud baseline on ModelNet40 and ScanObjectNN, and the model reaches 100 percent test accuracy on all 11 classes of the DVS128 Gestures dataset for at least one training run.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The point-cloud evidence for the claimed step-size inductive bias is confounded: STREAM changes ordering, sequence length, and Δ parameterization at once, with no matched ablation, so the causal role of coordinate-difference step sizes is not established.","rationale":"The paper's mathematical core (App. 6) is derived cleanly: solving the linear time-varying SSM with Dirac impulses yields h_k=e^{AΔ_k}h_{k−1}+B_ku_k and the pairwise kernel (6). The step-size injection is well-defined for both event time and sorted spatial coordinates. The problem is evidential: the point cloud experiments, which are the primary evidence for the spatial inductive-bias claim, do not isolate Δ. Section 4.1.1 explicitly says they replace Hilbert curves with XYZ sorting and increase token count from 2N to 3N. Therefore Tables 1–2 cannot identify which change caused the gains. Table 5 is a matched STREAM-vs-Mamba ablation, but only for event streams; it supports the temporal version of the idea, not the point-cloud/spatial version. The DVS128 100% headline is a best-seed maximum rather than a robust mean. None of this makes the method wrong, but it makes the central claim conditional on an additional controlled experiment and code release. The offered test would settle whether the concern actually lands.","tokens_in":15673,"tokens_out":7298,"duration_ms":84092,"concrete_test":"Run a matched ablation on ScanObjectNN PB-T50-RS with identical hyperparameters and seeds: (A) PointMamba: Hilbert ordering, 2N tokens, default Mamba Δ; (B) PointMamba with STREAM preprocessing: XYZ sorts, 3N tokens, default Mamba Δ; (C) STREAM as published. If B reaches D's accuracy (85.32% best / 84.4±0.7 mean) within roughly 0.2%, the gain is from preprocessing/sequence length and the inductive-bias claim is unsupported; if B stays near A and C retains most of the gain over B, the step-size encoding is the causal factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, §3.3, §4.1) is that explicitly injecting coordinate differences into the SSM step size, Δ_i=(t_i−t_{i-1})Ψ(δ), provides the inductive bias that improves PointMamba. Tables 1–2 compare STREAM (12.3M params) against the PointMamba baseline, but §4.1.1 changes three factors simultaneously: (i) sequence length goes from 2N to 3N; (ii) ordering changes from two Hilbert curves to three independent XYZ sorts, with additional learned scale/shift parameters per sort dimension; (iii) the Mamba Δ is replaced by coordinate-difference Δ. No point-cloud ablation varies only factor (iii). The text acknowledges the first two changes as preprocessing simplifications, so the measured +2.84% on ScanObjectNN PB-T50-RS and +0.3% on ModelNet40 cannot be attributed to the step-size encoding alone. Table 5 does isolate STREAM vs Mamba on DVS128/SSC, but those are temporal event streams where the independent variable is time, not spatial geometry; it does not settle the point-cloud/spatial-geometry claim. The DVS128 '100%' headline is also a best-of-five-seeds maximum (the closest variant in Table 5 has mean 99.2±0.3), and no code is released, but the decisive gap is the missing point-cloud ablation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"STREAM proposes to encode the relative coordinate differences of sparse geometric data (point clouds, event streams) directly into the step-size parameter ∆ of a Mamba-style selective state-space model. The paper derives the resulting interaction kernel from a linear time-varying state-space model, describes a modified CUDA kernel for efficient scan-based training, and reports experiments on point cloud classification (ModelNet40, ScanObjectNN) and event-based vision/audio (DVS128 Gestures, Spiking Speech Commands). The central claims are that the explicit geometric step-size parameterization is a powerful inductive bias, improving the PointMamba baseline by up to 2.84% on ScanObjectNN, and that STREAM achieves 100% test accuracy on all 11 classes of DVS128 Gestures.","tokens_in":15969,"tokens_out":3857,"duration_ms":40519,"significance":"If the central claim holds, the proposed geometric step-size encoding is a simple and potentially impactful way to inject inductive bias into state-space models for irregularly sampled data. The mathematical derivation in Appendix 6 is rigorous and standard, correctly identifying the kernel from the continuous-time solution. The event-stream experiments provide some isolated evidence for the benefit of coordinate-derived step sizes (Table 5), and the paper is the first to report 100% accuracy on the full DVS128 Gestures benchmark, albeit as a best-of-seeds result. The main weakness is that the point cloud experiments, which are the primary evidence for the claim on spatial geometry, change multiple factors at once and therefore do not establish the causal role of the proposed step-size encoding.","major_comments":[{"comment":"The point cloud comparison between STREAM and PointMamba changes three factors simultaneously: the sequence length is increased from 2N to 3N, the ordering is changed from two Hilbert curves to three independent XYZ sorts (with additional learned scale/shift parameters), and the ∆ parameterization is changed from a learned input-dependent function to the coordinate-difference form. The text acknowledges the first two as preprocessing simplifications, but no matched ablation is provided that varies only the step-size encoding. As a result, the reported improvements (+2.84% on ScanObjectNN PB-T50-RS, +0.3% on ModelNet40) cannot be attributed to the claimed inductive bias. I request an ablation that isolates factor (iii), for example PointMamba with the same 3N XYZ-sorted input but the default Mamba ∆, or STREAM with the original Hilbert-curve ordering.","section":"§4.1.1, Tables 1–2"},{"comment":"The abstract and contributions state that STREAM 'achieves 100% test accuracy on all 11 classes' of DVS128 Gestures, but the corresponding row in Table 5 (the variant with tk − t(k−1) and no softplus+Linear renormalization) reports a mean of 99.2% with a standard deviation of 0.3 over five seeds. The 100.0% in Table 3 is therefore a best-of-seeds maximum, not the average performance. The manuscript should either report the mean and variance alongside the maximum or qualify the abstract claim explicitly, since the current wording overstates the reliability of the result.","section":"§4.2.2, Tables 3 and 5"},{"comment":"The point cloud model concatenates three sequences sorted by X, Y, and Z into a single 3N sequence, and the step size is defined as ∆i = ti − ti−1 for the successive coordinates in that concatenated sequence. At the two concatenation boundaries, t jumps from the maximum value of one coordinate to the minimum value of the next (e.g., max X to min Y), producing a large, geometrically meaningless ∆. The paper does not describe any masking, state reset, or other mechanism to handle these boundary discontinuities. This is a concrete technical issue: unless the boundaries are handled explicitly, the claimed 'explicit encoding of geometric structure' is not actually implemented for a substantial fraction of the 3N positions, and the interpretation of the experimental results is unclear. Please clarify how the boundaries are treated.","section":"§3.4, §4.1.1"}],"minor_comments":[{"comment":"The phrase 'replacing the Mamba module in PointMamba with a STREAM module' is imprecise, because the comparison also changes the preprocessing, sequence length, and ordering; consider stating that the architecture is aligned to PointMamba while changing these additional components.","section":"§4.1.1"},{"comment":"The indices in the product range from i+1 to k, which is correct for the kernel, but the text says 'pairs of N points' while the sums run from 0 to N, giving N+1 points; please normalize the indexing notation.","section":"§3.2, equation (6)"},{"comment":"The paper states that the same hyperparameters as PointMamba are used, but the sequence length differs (3N vs 2N); please state explicitly whether training time or memory differs and how the increased sequence length is handled.","section":"§4.1.1"},{"comment":"The claim of 'for the first time' 100% on DVS128 is made relative to previous work, but the comparison in Table 3 includes methods with different evaluation protocols (e.g., TENNs-PLEIADES on 10 of 11 classes); this is noted, but the headline claim should also be qualified by the seed-selection issue raised above.","section":"§1 and §4.2.2"}],"recommendation":"major_revision","confidential_remarks":"The theoretical derivation and the event-stream ablation are solid, and the paper is within scope for a computer vision venue. The decisive gap is the missing point-cloud ablation that isolates the step-size parameterization; without it, the central contribution (geometric inductive bias for point clouds) is not empirically supported. The DVS128 100% claim also needs to be reported with its variance. These are fixable with additional experiments and careful rewording, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on the STREAM paper.\n\nThe genuinely new thing is the parameterization: feed actual coordinate or timestamp differences into the Mamba step size, Δ_i = (t_i − t_{i−1})Ψ(δ). The derivation in Appendix 6 is clean, standard linear-systems theory — solving the ODE with Dirac-delta inputs gives exactly that kernel, no free parameters, no circularity. That part is solid. The CUDA extension for irregular step sizes is a useful engineering contribution, and the event-stream experiments in Table 5 isolate the Δ choice against Mamba, showing consistent gains on both DVS128 Gestures and Spiking Speech Commands. Those are honest, controlled comparisons.\n\nThe soft spot is the point-cloud story. The claimed up-to-2.84% gain over PointMamba on ScanObjectNN changes three things at once: sequence length (2N to 3N), ordering (two Hilbert curves to three XYZ sorts with learned scale/shift), and the Δ parameterization. The paper never runs a matched ablation that varies only Δ for point clouds, so the causal role of the step-size encoding for spatial geometry is not established. The text acknowledges the first two changes as preprocessing simplifications, which makes the gap load-bearing: the improvement could come from longer sequences or a friendlier ordering. The event-stream ablation is temporal, not spatial, so it doesn't settle the point-cloud claim. This is a real, addressable flaw, not a nitpick, because the abstract's central claim is about the inductive bias for sparse geometric data.\n\nMinor points: the 100% on DVS128 is the best of five seeds; the table's best variant has mean 99.2±0.3. The paper does phrase it as 'maximum accuracy', so it's not deceptive, but the headline should carry the mean. No code is released, which limits reproducibility.\n\nWho's this for? Researchers using SSMs for point clouds or event streams. The idea is simple enough to try in other backbones, and the derivation makes it easy to adapt. It deserves a serious referee: the core idea is sound, the derivation is rigorous, and the event-stream comparisons are meaningful. But the point-cloud section needs a matched ablation that changes only the Δ parameterization, plus code, before the central claim is credible.\n\nMy recommendation: send it to peer review, and make the missing point-cloud ablation the primary revision request.","headline":"A clean derivation and a plausible parameterization, but the point-cloud evidence is confounded and the spatial-geometry claim is not yet isolated.","tokens_in":16508,"tokens_out":4144,"would_cite":true,"duration_ms":33721,"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":"Coordinate-difference step sizes give one state-space model a strong inductive bias for sparse geometric data, delivering state-of-the-art point-cloud accuracy and, for the first time, 100% on all 11 DVS128 Gesture classes.","keywords":["state-space models","point clouds","event-based vision","irregular step sizes","Mamba","DVS128 Gestures","sparse geometric data","selective scan"],"falsifier":"Run the STREAM point-cloud model on ScanObjectNN's PB-T50-RS split with the step size $\\Delta_k$ forced to a constant (or set by PointMamba's input-dependent rule) while keeping the 3N XYZ-sorted sequence and all other hyperparameters fixed; if accuracy stays near 85.32%, the geometric inductive bias is not what produces the reported gain.","tokens_in":15499,"feed_emoji":"🎯","tokens_out":9932,"duration_ms":78737,"temperature":0.7,"pith_summary":"The paper argues that sparse geometric data — point clouds and event-camera streams — should be treated as pulses whose coordinate differences set the step size of a state-space model. It claims this explicit geometric encoding is a strong inductive bias: replacing the Mamba module in PointMamba with this parameterization lifts trained-from-scratch accuracy by up to 2.84% on ScanObjectNN and by 0.3% on ModelNet40. The same model, applied directly to raw event streams without frames or 2D convolutions, is claimed to reach 100% test accuracy on all 11 classes of the DVS128 Gestures dataset, reported for the first time. If the paper is right, one state-space backbone with coordinate-difference step sizes is a competitive universal sequence model for sparse geometric data.","feed_headline":"State-space model scores 100% on all 11 event-gesture classes","feed_subtitle":"Feeding coordinate gaps into the model's step size lifts point-cloud accuracy and perfects event-gesture recognition.","key_machinery":"The central object is the state-space recurrence $h_k = e^{A_k \\Delta_k} h_{k-1} + B_k u_k$ with step size $\\Delta_k$ set to the coordinate difference $t_k - t_{k-1}$ between consecutive points. This makes the transition matrix an exponentially oscillating kernel that integrates the history of earlier points, and a modified scan primitive parallelizes the irregularly spaced recurrence in $O(\\log N)$ time. STREAM is built on Mamba, a selective state-space layer whose step size is normally a learned function of the input; the paper's change is to make that step size carry the geometry explicitly.","core_discovery":"The central discovery is that the recurrence step $\\Delta_k = t_k - t_{k-1}$ of a selective state-space model can act as a geometric operator. Setting $\\Delta_k$ to the true coordinate difference between successive points produces a pairwise interaction kernel $\\Phi(t_k,t_i) = C_k \\prod_{j=i+1}^k \\exp(A_j \\Delta_j) B_i$ that is computed for all $N$ points in $O(N)$ steps. STREAM replaces Mamba's input-dependent step size with $(t_i - t_{i-1})\\Psi(\\delta)$, decouples $B_i$ from $\\Delta_i$, and sorts point clouds by the X, Y, and Z coordinates instead of using space-filling curves. The paper claims this explicit parameterization of geometry is what drives the point-cloud gains and enables fully event-based processing that attains 100% accuracy on all 11 DVS128 Gesture classes.","pith_inferences":["A matched ablation varying only the step-size encoding would determine how much of the ScanObjectNN gain comes from geometry rather than from the longer 3N sequence or the simpler XYZ ordering.","The coordinate-difference step-size idea likely extends beyond point clouds and events to any irregularly sampled signal, such as LiDAR sweeps, physiological recordings, or financial tick streams.","The 100% DVS128 result is achieved with a specific token encoding and CutMix-style event mixing, so rerunning without those augmentations would reveal what part of the perfect score is due to the model's geometric parameterization.","The kernel in equation (6) casts STREAM as a continuous convolution with an exponentially oscillating kernel, which connects it to a broader family of kernel-based geometric operators and suggests principled ways to tune the state dimension."],"forward_implications":["A single state-space backbone with coordinate-difference step sizes competes with specialized point-cloud and event-vision models on the same architecture.","Point-cloud sequence models can drop space-filling curve preprocessing and sort by raw X, Y, or Z coordinates without sacrificing accuracy.","Event-based recognition without frames or 2D convolutions can reach 100% test accuracy on all 11 DVS128 Gesture classes.","The recurrence runs asynchronously in $O(1)$ work per new point or event and trains in $O(\\log N)$ parallel steps on long streams.","The same parameterization transfers to event-based audio, reaching 86.3% on the Spiking Speech Commands dataset."],"supporting_citations":[{"why":"PointMamba: the point-cloud state-space baseline whose Mamba module STREAM replaces, providing the architecture and from-scratch accuracy numbers it improves on.","marker":"[20]"},{"why":"Mamba: the selective state-space model and scan kernel that STREAM modifies to support irregular coordinate-difference step sizes.","marker":"[9]"},{"why":"DVS128 Gestures: the event-based vision dataset on which STREAM claims 100% test accuracy across all 11 classes.","marker":"[1]"},{"why":"Event-SSM: prior event-by-event state-space approach for asynchronous streams, providing the direct event-based baseline and the CutMix-style augmentation used in training.","marker":"[38]"},{"why":"S7: concurrent event-based state-space work reporting 99.2% on DVS128 Gestures, the reference STREAM surpasses.","marker":"[41]"},{"why":"TENNs-PLEIADES: prior work reporting 100% on 10 of the 11 DVS128 classes, which frames STREAM's all-11-class claim.","marker":"[27]"},{"why":"ScanObjectNN: the real-scanned point-cloud benchmark whose hardest split (PB-T50-RS) shows the largest reported gain over PointMamba.","marker":"[43]"},{"why":"ModelNet40: the standard clean point-cloud classification dataset used for the second point-cloud evaluation.","marker":"[51]"}],"fun_headline_variants":["STREAM embeds coordinate differences into state-space step size","Geometric step sizes give state-space model 100% on all DVS128 classes","O(N) pairwise interactions by injecting geometry into step size","State-space model with geometric step sizes perfects event recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The point-cloud accuracy gains are credited to the geometric step-size encoding, but the comparison against PointMamba changes the sequence length, the ordering, and the step-size parameterization at the same time, so no matched test isolates the step-size encoding as the cause.","fun_headline_variants_meta":{"raw":{"variants":["STREAM embeds coordinate differences into state-space step size","Geometric step sizes give state-space model 100% on all DVS128 classes","O(N) pairwise interactions by injecting geometry into step size","State-space model with geometric step sizes perfects event recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2973,"prompt_tokens":1010,"completion_tokens":1963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1890}},"tokens_in":626,"tokens_out":1963,"duration_ms":15121,"temperature":1.0,"reasoning_tokens":1890,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:20:58.643378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the STREAM point-cloud model on ScanObjectNN's PB-T50-RS split with the step size $\\Delta_k$ forced to a constant (or set by PointMamba's input-dependent rule) while keeping the 3N XYZ-sorted sequence and all other hyperparameters fixed; if accuracy stays near 85.32%, the geometric inductive bias is not what produces the reported gain.","supporting_citations":[{"cited_title":"Pointmamba: A simple state space model for point cloud analysis","cited_arxiv_id":null,"evidence_quote":"PointMamba: the point-cloud state-space baseline whose Mamba module STREAM replaces, providing the architecture and from-scratch accuracy numbers it improves on."},{"cited_title":"Mamba: Linear-time sequence mod- eling with selective state spaces, 2024","cited_arxiv_id":null,"evidence_quote":"Mamba: the selective state-space model and scan kernel that STREAM modifies to support irregular coordinate-difference step sizes."},{"cited_title":"A low power, fully event-based gesture recognition system","cited_arxiv_id":null,"evidence_quote":"DVS128 Gestures: the event-based vision dataset on which STREAM claims 100% test accuracy across all 11 classes."},{"cited_title":"Scalable 10 event-by-event processing of neuromorphic sensory signals with deep state-space models","cited_arxiv_id":null,"evidence_quote":"Event-SSM: prior event-by-event state-space approach for asynchronous streams, providing the direct event-based baseline and the CutMix-style augmentation used in training."},{"cited_title":"S7: Selective and simplified state space layers for sequence modeling, 2024","cited_arxiv_id":null,"evidence_quote":"S7: concurrent event-based state-space work reporting 99.2% on DVS128 Gestures, the reference STREAM surpasses."},{"cited_title":"Revisiting point cloud classification: A new benchmark dataset and classifi- cation model on real-world data","cited_arxiv_id":null,"evidence_quote":"ScanObjectNN: the real-scanned point-cloud benchmark whose hardest split (PB-T50-RS) shows the largest reported gain over PointMamba."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"ModelNet40: the standard clean point-cloud classification dataset used for the second point-cloud evaluation."}],"review_version":1}