{"id":"2c4d7610-448c-4503-9576-dca6f945072a","arxiv_id":"2507.04369","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A camera-LiDAR 3D detector built around a hybrid local-global Mamba block with height-fidelity LiDAR encoding reports 75.0 NDS on nuScenes validation, outperforming prior transformer-based fusion methods.","lead":"A new fusion method for self-driving perception combines camera and LiDAR data using a Mamba-based linear-attention block, claiming a state-of-the-art 75.0 NDS score on the nuScenes benchmark. If the results hold up, it suggests efficient global fusion without the quadratic cost of standard attention can be competitive for 3D object detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The height-fidelity mechanism is not isolated: switching to continuous coordinates (Eqs. 8–10) changes x, y, and z and the voxel set, so attributing the gains to height alone is underdetermined; a z-only ablation is needed.","rationale":"I read the paper as an empirical architecture paper whose central claim is that a pure Mamba fusion block plus height-fidelity LiDAR encoding reaches 75.0 NDS on nuScenes val while being fast. The strongest direct evidence for the fusion claim is Tab. 7, where a window-transformer and a hybrid-Mamba variant are compared in the same Image+BEV continuous setting, giving +1.3 mAP/+1.4 NDS for Mamba; this is a reasonably controlled comparison and supports the core claim. The paper also gives component ablations (Tabs. 4–6), qualitative projections, and robustness tests. What is least secure is the causal attribution in Sec. 1: the height-fidelity encoding is a bundle of three changes (continuous x,y,z centroids, continuous downsampled centroids, and a conflict test), all of which alter Mamba's token ordering and voxel selection. No experiment isolates height. This is a real soft spot because height fidelity is the paper's central novelty and the stated reason why linear attention fails, but it does not by itself invalidate the reported SOTA numbers. The reader's CONDITIONAL verdict already captures the lack of reproducibility details (no seed variance, half-dataset ablations, code without commit hash), and my concern sharpens the missing test rather than overturning the empirical result; therefore I recommend UNCHANGED.","tokens_in":14893,"tokens_out":7819,"duration_ms":90320,"concrete_test":"Run a z-only HFL variant: keep the original discrete x,y voxel centers from LION, set only z to the continuous ScatterMean value from Eqs. (8)–(9), and disable the Eq. (10) conflict test. Train this on the same 50% nuScenes protocol as Tab. 7 row ⑥ (Hybrid Mamba, Discrete, Image+BEV) and compare mAP/NDS to rows ⑥ (70.3/73.1) and ⑦ (71.9/74.3). If the z-only variant recovers most of the gain, the height-fidelity story is supported; if it stays near the discrete row, the gain is due to x,y changes in the Hilbert/local ordering or the conflict test, not height fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's distinct novelty is height-fidelity LiDAR encoding (Sec. 3.4), and the causal story in Sec. 1 is that loss of height during multi-modal alignment causes vanilla Mamba's degradation. The evidence, however, compares discrete versus continuous coordinate computations as a bundle. Equations (8)–(9) replace voxel centroids with ScatterMean coordinates in all three axes, not just z; Eq. (10) changes the set of generated voxels. These continuous coordinates feed the Hilbert index (Eq. 3) and the local region partitions (Eq. 6), so token order changes in both Mamba stages. Tab. 7's ⑥→⑦ contrast therefore mixes height fidelity with serialization order and voxel-set changes. The paper does not measure height error before/after HFL, and does not run a condition in which only the z coordinate is corrected while x,y remain on the discrete grid. If the observed 71.9/74.3 gain comes from smoother Hilbert ordering or from the conflict test removing duplicate voxels, the height-fidelity contribution claimed as the core mechanism is not established. The benchmark result could still be correct; what is underdetermined is why it works and whether the mechanism transfers to other backbones and linear-attention blocks as claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MambaFusion, a camera-LiDAR 3D object detection framework in which fusion is performed entirely with Mamba/linear-attention blocks rather than windowed transformers. The authors propose a Hybrid Mamba Block that combines local and global Mamba stages, and a Height-Fidelity LiDAR Encoding (HFL) that computes voxel coordinates in continuous space to preserve height information during multi-modal alignment. On nuScenes, the method reports 75.0 NDS / 72.7 mAP on validation and 75.9 NDS / 73.2 mAP on test with 4.7 FPS, surpassing prior methods including IS-FUSION and SparseLIF. The paper also reports ablations, robustness studies under sensor degradation, and an integration into the FusionAD planner.","tokens_in":15103,"tokens_out":4837,"duration_ms":52691,"significance":"If the reported results are reproducible, the paper makes a meaningful advance: it demonstrates that pure Mamba-based fusion can reach state-of-the-art accuracy on a standard multi-modal 3D detection benchmark while being faster than recent transformer-based competitors. The central conceptual contribution is the height-fidelity LiDAR encoding, which is claimed to fix a height-information loss that degrades vanilla Mamba. The paper is generally well written and the experimental suite is broad, including component ablations, robustness to sensor degradation, ERF visualizations, and an end-to-end planning evaluation. I credit the authors for releasing code and for comparing against external benchmarks. However, the load-bearing mechanistic claim about height information is not isolated in the experiments, and the ablation results are single-run, half-dataset numbers with small deltas. The benchmark result could still be correct, but the reason it works is underdetermined as presented.","major_comments":[{"comment":"The paper's central claim is that height-information loss causes the degradation of vanilla Mamba and that Height-Fidelity LiDAR Encoding fixes it. However, the discrete-to-continuous change implemented by Eqs. (8)–(9) replaces voxel centroids with ScatterMean coordinates in all three axes, not only z, and Eq. (10) changes the set of generated voxels. These changes also alter the Hilbert indices in Eq. (3) and the local region partitions in Eq. (6), so the improvement from row ⑥ to row ⑦ in Tab. 7 (and from ③ to ⑤ in Tab. 4) is not attributable to height fidelity alone. I would need an ablation in which only the z coordinate is computed continuously while x and y remain on the discrete grid (or an equivalent isolation of height information) to support the height-specific mechanism.","section":"§3.4, Eqs. (8)–(10), Tab. 7"},{"comment":"The causal story in Sec. 1—that Mamba's performance drop is caused by height-information loss during multi-modal alignment, 'leading to deviations in sequence order'—is asserted rather than demonstrated. The comparison in Tab. 7 from row ④ to row ⑦ changes the architecture (vanilla Mamba vs. hybrid local/global Mamba), the fusion space (Image+Frustum vs. Image+BEV), and the coordinate discretization simultaneously. No experiment holds model capacity, optimizer, sequence ordering, or architecture fixed while correcting only height information. Please add such an experiment, or explicitly weaken the causal claim to a design observation rather than a demonstrated mechanism.","section":"§1, §4.3.4, Tab. 2"},{"comment":"All component ablations are performed on half of the nuScenes validation set with a single run, and the reported deltas are small (e.g., 71.4→71.9 mAP and 73.6→74.3 NDS in Tab. 4 between ⑥ and ⑦; 71.5→71.9 mAP in Tab. 5(a)). Without multiple seeds or error bars, these differences may be within run-to-run variation, which undermines the component-level conclusions and the claim that HFL is the key enabler. Please report mean and standard deviation over at least three seeds for the main ablations, or clearly present the numbers as preliminary with the attendant uncertainty.","section":"§4.3, Tabs. 4–6"}],"minor_comments":[{"comment":"There are typos: 'top-tire NDS' should be 'top-tier NDS', and 'adapt pure linear attnetion' should be 'adapt pure linear attention'.","section":"Abstract and Sec. 2"},{"comment":"The notation for the conflict test in Eq. (10) is hard to parse; please rewrite it with explicit set-builder notation and define the membership condition '∈ CP_s' more clearly.","section":"§3.4, Eq. (10)"},{"comment":"The '1.5× faster' claim relative to IS-FUSION is ambiguous: Tab. 3 gives 5.4 FPS for MambaFusion-Lite (a 1.69× speedup) and 4.7 FPS for MambaFusion-Base (a 1.47× speedup) against IS-FUSION's 3.2 FPS; please specify which configuration is being compared.","section":"§4.2, Tab. 3"},{"comment":"Figure 5 is qualitative; reporting a quantitative projection or height error before and after HFL would directly support the height-loss narrative and complement the mAP/NDS tables.","section":"§3.4, Fig. 5"},{"comment":"The 'vanilla' linear-attention baseline in Tab. 2 is only described as replacing UniTR's fusion module; please clarify whether the exact same tokenizers, training schedule, and fusion space are used across all rows so that the comparison to UniTR is clean.","section":"§3.1, Tab. 2"},{"comment":"The table's checkmarks are not defined there; please add a legend defining Baseline-LC, HMB-M, HMB-R, HMB-B, and HFL in the caption for readability.","section":"§4.3.1, Tab. 4"}],"recommendation":"major_revision","confidential_remarks":"This is a strong systems paper with impressive benchmark numbers, but the main mechanistic claim needs an isolation experiment. If the authors can provide a z-only ablation (or otherwise separate height correction from changes in x/y coordinates and voxel set) and add variance estimates to the small-delta ablations, I would be inclined to support acceptance. The current recommendation is major_revision rather than reject because the central result is plausible and the missing evidence is obtainable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: MambaFusion is a credible engineering contribution. The headline result—pure Mamba/linear-attention fusion hitting 75.0 NDS on nuScenes val at competitive speed—is the kind of thing the 3D detection community will care about. The Hybrid Mamba Block and the continuous-space voxel encoding are sensible pieces, and the paper does more than most: it tests three linear attention families, reports ERFs, includes robustness under sensor degradation, and shows the fusion module helps in an end-to-end planner. That is real work.\n\nWhat is actually new is the combination, not the components. Mamba for 3D detection isn't new, and height-preserving voxel tricks exist in spirit. But the specific package—continuous-space ScatterMean coordinates for voxel centroids, conflict test in LION, and the local+global Mamba fusion in raw and BEV space—is not something I've seen. The ablations are internally consistent and the gains are in the right direction.\n\nThe soft spots are real but manageable. The biggest is the causal story. The paper says vanilla Mamba fails because height information is lost during alignment, and height-fidelity encoding fixes it. But as the stress-test notes, switching from discrete to continuous coordinates changes x, y, and z together, changes the Hilbert ordering, and changes the voxel set via the conflict test. Attributing the 71.9/74.3 to 'height' alone is underdetermined. A z-only ablation—keep x,y discrete, only correct z—would settle it. That matters because the paper sells height as the load-bearing idea.\n\nThe experimental transparency is also thin in places. No error bars, single runs, component choices made on half the validation set, and the code URL has no commit or configs. For a benchmark-driven paper, that's exactly where a referee should push. The FPS numbers are hard to trust without a reproducible setup.\n\nI'd also downweight the 'first to show pure linear fusion can be SOTA' framing. It's true within the narrow comparison table, but it's the kind of claim that doesn't age well and isn't the real contribution.\n\nOverall: this paper deserves a serious referee. The empirical result is likely real, the method is well-motivated, and the flaws are fixable with more experiments and better reporting. I would accept it for review, with the expectation of heavy revision.","headline":"Solid empirical paper with a real but under-isolated mechanism claim; worth refereeing with requests for proper ablations and code.","tokens_in":15659,"tokens_out":2108,"would_cite":true,"duration_ms":23021,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MambaFusion claims the first pure-Mamba fusion block to reach state-of-the-art camera-LiDAR 3D object detection, with 75.0 NDS on nuScenes validation.","keywords":["3D object detection","multi-modal fusion","Mamba","state space model","camera-LiDAR fusion","height-fidelity encoding","dense global fusion","nuScenes"],"falsifier":"Compare the continuous-space height-fidelity encoding against discrete-space coordinates with identical model capacity and training schedule; if the discrete version with extra height channels or height-aware positional embedding reaches the same NDS, then coordinate precision is not the active ingredient. Conversely, randomize the Hilbert ordering while keeping height-fidelity coordinates; if performance does not collapse, sequence order is not what carries the gain.","tokens_in":14627,"feed_emoji":"🚗","tokens_out":7153,"duration_ms":70337,"temperature":0.7,"pith_summary":"This paper sets out to show that a fusion module built purely from linear-complexity state-space models, specifically Mamba, can outperform windowed-transformer fusion for camera-LiDAR 3D object detection, at lower inference cost. The starting puzzle is that swapping Mamba into an existing fusion framework degrades accuracy, and the paper attributes the drop to height information being lost when LiDAR voxels are compressed into a shared coordinate space, which corrupts the order in which Mamba reads the fusion sequence. The proposed remedy is a height-fidelity LiDAR encoding that computes voxel coordinates in continuous space and filters out conflict-prone generated voxels, followed by a Hybrid Mamba Block that mixes local-window and bidirectional global Mamba over a Hilbert-curve serialization. The reported result is a nuScenes validation NDS of 75.0, surpassing previous best methods even those using higher-resolution images, and roughly 1.5x faster inference than a recent top method.","feed_headline":"Pure Mamba fusion hits 75.0 NDS on nuScenes 3D detection","feed_subtitle":"Linear-complexity fusion beats windowed transformers once LiDAR height is preserved; runs 1.5x faster.","key_machinery":"The load-bearing mechanism is the pairing of height-fidelity LiDAR encoding with the Hybrid Mamba Block. Height-fidelity encoding replaces the discrete-space centroid of each voxel with a continuous-space ScatterMean over the raw point coordinates inside that voxel, and adds a conflict test that suppresses newly generated voxels whose merged position would be ambiguous; this keeps the z-coordinate precise enough that LiDAR features projected into image space land on the objects they describe. The Hybrid Mamba Block then processes these aligned features at two scales: a local Mamba over non-overlapping windows captures fine structure, and a bidirectional global Mamba over tokens serialized with a Hilbert curve, a space-filling curve that preserves spatial locality, captures full-scene context. The block is placed in a shared modality-aligner role, in raw-space fusion, and in BEV fusion, so it both harmonizes the two modalities and performs the final fused representation.","core_discovery":"The central claim is that pure linear operations for fusion, not only quadratic attention, can reach state-of-the-art performance in camera-LiDAR 3D detection. The paper's key discovery is that the failure of vanilla Mamba in this setting is caused by height-information loss during multimodal alignment, and that this loss is fixable: by computing voxel coordinates through scatter-mean averaging in continuous 3D space rather than quantized discrete voxel grids, and by rejecting generated voxels that would merge ambiguously with neighbors, the LiDAR features project onto image features with far fewer mismatches. Once alignment is accurate, a Hybrid Mamba Block, composed of local Mamba for fine detail and bidirectional global Mamba for whole-scene context with tokens ordered along a Hilbert curve, delivers dense global fusion. On nuScenes this yields 75.0 NDS and 72.7 mAP on validation and 75.9 NDS on test, beating UniTR, IS-FUSION, and SparseLIF while running substantially faster.","pith_inferences":["The causal story is not isolated: if optimization instability of SSM layers, rather than height loss, drives vanilla Mamba's drop, the height-fidelity encoding might still help without being the load-bearing mechanism; a controlled capacity-matched comparison would separate these.","Continuous-space coordinate fixes could plausibly improve other projection-based fusion tasks, such as radar-camera, LiDAR-maps, or point-text alignment, wherever discrete quantization corrupts correspondences.","The conflict test removes voxels likely to merge, which also reduces point density in cluttered regions; a density-matched baseline would reveal whether part of the gain is de-noising rather than height preservation.","A testable extension is to evaluate with higher-resolution images: if the margin over windowed transformers shrinks, height-fidelity matters most when image-LiDAR alignment is coarse; if it grows, the benefit lies in the global modeling itself."],"forward_implications":["Pure linear fusion blocks can replace windowed transformers inside a BEV detector without sacrificing accuracy, so latency-critical autonomy systems can spend the saved compute on other stages.","Height-fidelity encoding transfers: applying it to the windowed-transformer variant also improves performance, meaning the alignment fix is not specific to Mamba.","The fused representation generalizes beyond detection: plugging the method into FusionAD reduces average displacement error from 0.81 to 0.70 and collision rate from 0.12 to 0.10.","Robustness to degraded sensors improves, with the largest gain under unsynchronized LiDAR-camera timestamps, where global context compensates for broken local alignment.","Because the Hybrid Mamba Block is compatible with RetNet and RWKV as well, the design is a template for linear-attention fusion rather than a Mamba-only trick."],"supporting_citations":[{"why":"Supplies the windowed-transformer fusion baseline whose fusion module is replaced by Mamba and whose Image+Frustum alignment defines the initial setup.","marker":"[47]"},{"why":"The selective state-space model used as the pure linear fusion operator that the paper adapts into local and global blocks.","marker":"[8]"},{"why":"The LiDAR backbone whose voxel generation and top-k salient voxel creation introduce height deviation; the paper adds a conflict test against its generated voxels.","marker":"[33]"},{"why":"HiPPO initialization is cited as the reason Mamba captures long-range dependencies without explicit spatial priors, justifying the choice over other linear attentions.","marker":"[9]"},{"why":"nuScenes is the dataset and benchmark where all NDS and mAP claims are measured.","marker":"[2]"},{"why":"Defines the BEV-space alignment and robustness-evaluation protocol that the paper inherits as its dense-local fusion baseline.","marker":"[26]"},{"why":"The recent state-of-the-art method used for the direct speed comparison (1.5x faster inference).","marker":"[53]"},{"why":"The Hilbert curve is cited as the space-filling serialization that preserves spatial locality for Mamba global fusion.","marker":"[15]"}],"fun_headline_variants":["Mamba fusion preserves height, hits 75.0 NDS on nuScenes","Pure Mamba fusion reaches 75.0 NDS, beats windowed attention","Height-fidelity Mamba scores 75.0 NDS for 3D detection","Linear Mamba fusion: 75.0 NDS and faster inference on nuScenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's causal explanation, that height-information loss rather than optimization instability or the serialization itself is what makes vanilla Mamba degrade, is an attribution supported by ablations but not isolated in a controlled experiment that fixes capacity, optimizer, and ordering.","fun_headline_variants_meta":{"raw":{"variants":["Mamba fusion preserves height, hits 75.0 NDS on nuScenes","Pure Mamba fusion reaches 75.0 NDS, beats windowed attention","Height-fidelity Mamba scores 75.0 NDS for 3D detection","Linear Mamba fusion: 75.0 NDS and faster inference on nuScenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2969,"prompt_tokens":1012,"completion_tokens":1957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1878}},"tokens_in":628,"tokens_out":1957,"duration_ms":15095,"temperature":1.0,"reasoning_tokens":1878,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:48:37.366992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the continuous-space height-fidelity encoding against discrete-space coordinates with identical model capacity and training schedule; if the discrete version with extra height channels or height-aware positional embedding reaches the same NDS, then coordinate precision is not the active ingredient. Conversely, randomize the Hilbert ordering while keeping height-fidelity coordinates; if performance does not collapse, sequence order is not what carries the gain.","supporting_citations":[{"cited_title":"Unitr: A unified and efficient multi-modal transformer for bird’s-eye-view repre- sentation","cited_arxiv_id":null,"evidence_quote":"Supplies the windowed-transformer fusion baseline whose fusion module is replaced by Mamba and whose Image+Frustum alignment defines the initial setup."},{"cited_title":"Hippo: Recurrent memory with optimal polynomial projections","cited_arxiv_id":null,"evidence_quote":"HiPPO initialization is cited as the reason Mamba captures long-range dependencies without explicit spatial priors, justifying the choice over other linear attentions."},{"cited_title":"nuscenes: A mul- timodal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"nuScenes is the dataset and benchmark where all NDS and mAP claims are measured."},{"cited_title":"Bevfusion: A simple and robust lidar-camera fusion framework","cited_arxiv_id":null,"evidence_quote":"Defines the BEV-space alignment and robustness-evaluation protocol that the paper inherits as its dense-local fusion baseline."},{"cited_title":"Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection","cited_arxiv_id":null,"evidence_quote":"The recent state-of-the-art method used for the direct speed comparison (1.5x faster inference)."},{"cited_title":"¨Uber die stetige abbildung einer linie auf ein fl ¨achenst¨uck","cited_arxiv_id":null,"evidence_quote":"The Hilbert curve is cited as the space-filling serialization that preserves spatial locality for Mamba global fusion."}],"review_version":1}