{"id":"671ca9a9-b7ca-4dd9-968d-d011ef10651e","arxiv_id":"2507.20224","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MambaMap fuses four previous frames of BEV features and instance queries via gated state space layers, beating prior HD map construction methods on nuScenes and Argoverse2.","lead":"MambaMap uses a state space model to combine current and historical camera views into vectorized HD maps for self-driving. It reports top scores on nuScenes and Argoverse2, but the efficiency claim is partially contradicted by its own FPS tables.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table III gives the smallest SOTA margin as 0.4 mAP (Argoverse2 new split), contradicting the text's 'minimum gain of 3.1 mAP'; with no error bars or code, the SOTA claim is not yet established.","rationale":"I read the paper's strongest claim as an empirical SOTA claim. The architecture is coherent and the ablations are internally consistent, but the load-bearing condition for the SOTA claim is that the benchmark numbers are accurate, fairly compared, and statistically meaningful. That condition is not met at the level claimed. The paper's own Table III shows a minimum margin of 0.4 mAP on the Argoverse2 disjoint split, directly contradicting the text's 'minimum gain of 3.1 mAP,' and no seed variance is reported. Without released code or verifiable baseline reproduction, the small margins could be noise or configuration differences. The instance-query L2 matching issue identified by the reader is a plausible technical risk, but it is less load-bearing than the empirical verification problem: even imperfect matching could still yield competitive mAP, whereas an unsupported point estimate cannot establish SOTA. The reader's rationale did mention small margins and missing error bars, but their formal weakest_assumption focused on query matching; therefore I mark partial agreement. The conditional verdict remains appropriate: accept the SOTA claim only after multi-seed validation, code release, and correction of the 3.1 mAP statement.","tokens_in":985,"tokens_out":954,"duration_ms":97394,"concrete_test":"After releasing the MambaMap code and exact baseline reproduction configurations, rerun the Argoverse2 geographically-disjoint split at least three seeds for MambaMap, StreamMapNet, SQD-MapNet, and MemFusionMap, reporting mean and standard deviation. If the 61.0 vs 60.6 mAP gap does not exceed one standard deviation, the 'outperforms SOTA' claim should be softened to 'comparable'; if it persists across seeds and exceeds the measured variance, the claim is supported. As a preliminary analytical check, recompute the minimum pairwise gain in Table III: the minimum is 0.4 mAP on Argoverse2, not 3.1 as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MambaMap outperforms prior SOTA across data splits and perception ranges. That claim rests entirely on the point estimates in Tables I-III. The least secure link is the correctness and statistical significance of the reported margins. Section IV-B states after Table III: 'our MambaMap outperforms all competing methods with a minimum gain of 3.1 mAP.' This is contradicted by the paper's own numbers: on the Argoverse2 geographically disjoint split, MambaMap (61.0 mAP) exceeds both SQD-MapNet and MemFusionMap (60.6 mAP) by only 0.4 mAP; on the nuScenes disjoint split, the smallest gain is 2.1 mAP over MemFusionMap (38.0 to 40.1), not 3.1. Thus the 'substantial margins' claim is not supported, and at least one numeric statement in the text is internally inconsistent. All experiments appear to be single runs with no variance estimates. Several baselines are quoted from other papers, and the rest are described only as 'reproduced using official public codes,' with no code or reproduction configuration provided. Because the closest reported margin (0.4 mAP) is within typical run-to-run variation for these benchmarks, the SOTA claim is not yet established. This is an evidence concern about the empirical argument, not an architectural objection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaMap, an online vectorized HD map construction method that fuses temporal information at both BEV and instance-query levels using a state-space model. A memory bank stores recent BEV features and instance queries; BEV Mamba Fusion applies gated state-space blocks with multi-directional scanning, while Instance Mamba Fusion matches queries across frames by L2 distance and scans spatial-temporal sequences. Experiments on nuScenes and Argoverse2 report improved mAP over prior methods, including on geographically disjoint splits, with ablations over fusion levels, scanning strategies, matching costs, SSM variants, and memory-bank size.","tokens_in":12335,"tokens_out":4967,"duration_ms":55288,"significance":"If the reported results are reproducible, the paper makes a useful contribution: it demonstrates that state-space temporal fusion can improve long-range temporal consistency in online vectorized HD map construction, and the ablations are unusually complete, covering fusion levels, scanning strategies, matching costs, SSM variants, and memory-bank size. The fixed-size memory bank with gated SSM is a clean way to avoid quadratic sequence cost. However, the central SOTA claim is currently supported only by single-run point estimates and is weakened by an internal contradiction in the text, so the empirical evidence needs strengthening before the claim is established.","major_comments":[{"comment":"In the paragraph after Table III, the text states that 'our MambaMap outperforms all competing methods with a minimum gain of 3.1 mAP,' but the table's own numbers contradict this: on the Argoverse2 geographically disjoint split, MambaMap (61.0 mAP) is only 0.4 mAP ahead of SQD-MapNet and MemFusionMap (both 60.6), and on the nuScenes disjoint split the smallest gain is 2.1 mAP over MemFusionMap (40.1 vs 38.0), not 3.1. This internal inconsistency directly undermines the 'substantial margins' claim and must be corrected, with the reported margins restated consistently.","section":"§IV-B, Table III"},{"comment":"The SOTA claim rests entirely on point estimates from what appear to be single training runs; no error bars, standard deviations, or repeated-seed statistics are reported for any table. The closest reported margin (0.4 mAP on the Argoverse2 new split) is within typical run-to-run variation for this benchmark, so without variance estimates the superiority claim is not statistically established. In addition, the paper mixes published baseline numbers (marked '*') with baselines 'reproduced using official public codes' but gives no reproduction configuration or code release, making the comparison hard to audit. Please report multiple-seed statistics and release code/checkpoints or detailed reproduction settings.","section":"§IV-B, Tables I-III"},{"comment":"The instance-level temporal fusion assumes that L2 distance between query vectors is a reliable proxy for the identity of the same map element across frames, and the Hungarian matching in Eq. (10) reorders history under that proxy. Table VI compares three matching costs by final mAP but does not analyze how often the matching is correct or whether query vectors carry stable instance identity. This is a correctness-risk concern: if the L2 cost frequently mismatches elements, the temporal fusion could swap features across instances and corrupt predictions. A concrete test would be to measure the ground-truth-instance agreement of the Hungarian matches on a sample of frames, or to compare against matching by predicted geometry; at minimum, the paper should discuss this limitation.","section":"§III-D, Eqs. (9)–(11)"}],"minor_comments":[{"comment":"The checkmark-only rows should be explicitly labeled; the first row has no checkmarks, and the rows with BMF or IMF alone are ambiguous without a legend.","section":"Table IV"},{"comment":"The text label 'Fused Quries' contains a typo and should read 'Fused Queries'.","section":"Figure 2"},{"comment":"The hyperparameter list uses both 'Nnum = 100' and the symbol 'N_q' from Section III-D without defining their relationship; please unify the notation.","section":"§IV-A"},{"comment":"The FPS column shows MambaMap is slower than StreamMapNet (12.7 vs 14.9 on nuScenes; 13.6 vs 15.9 on Argoverse2), but the text emphasizes computational efficiency and does not comment on this overhead; please add a sentence discussing the speed-accuracy trade-off.","section":"Tables I and II"},{"comment":"The abstract and conclusion claim superiority 'across various splits and perception ranges,' but the geographically disjoint-split experiments in Table III are reported only at the 30 m range; the 50 m range on the new splits is missing. Please either add those results or restate the claim to match the experiments.","section":"§IV-B and Table III"}],"recommendation":"major_revision","confidential_remarks":"The empirical-evidence concern should weigh heavily: the 0.4 mAP margin combined with the 3.1 mAP text error suggests the paper should not be accepted without a careful revision and additional experiments. I do not see evidence of misconduct, but the reporting discipline needs tightening before the SOTA claim can be trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MambaMap is a sensible engineering contribution and the experiments are substantial, but the paper overstates its own margins and the SOTA claim is not yet fully secured. Worth engaging, not worth trusting at face value.\n\nWhat is new: the paper adapts gated state space layers (DSS + GSS-style gating) and VMamba-style multi-directional scanning to temporal fusion for vectorized HD map construction, at both BEV feature and instance-query levels. That combination, plus the multi-frame memory bank with size N, is genuinely new for this task. The ablations are thorough: they compare scanning strategies, matching costs, SSM variants, and memory size. The architecture is coherent and the reported gains over StreamMapNet on nuScenes (3.9 and 3.0 mAP) are consistent with the design. That is real work and the paper is honestly situated relative to prior temporal modeling work.\n\nSoft spots. First, the text says after Table III 'minimum gain of 3.1 mAP' but the paper's own numbers give 0.4 mAP on the Argoverse2 disjoint split (61.0 vs 60.6). That is an internal inconsistency in the central empirical claim. Second, all experiments look like single runs; no seeds, no error bars. On this benchmark, 0.4 mAP is within run-to-run noise, so the 'outperforms all competing methods' claim needs variance estimates or code. Third, the efficiency story is muddier than claimed: MambaMap runs at 12.7 FPS while StreamMapNet is 14.9 and SQD-MapNet 14.4 on the same A30. The SSM linear-complexity argument is about asymptotic scaling, not actual speed here. Minor: instance matching via L2 query distance is plausible but the ablation only shows aggregate mAP; a few failure cases would help.\n\nThe citation pattern looks fine. Self-citations to MGMap and the uncertainty-instructed paper appear in related work where they belong, not as load-bearing premises. No circularity. The math is straightforward; no hidden fitting of constants.\n\nWho this is for: people working on BEV perception and HD map construction. It deserves a serious referee: the architecture is transferable and the ablations are informative. My recommendation: send it to review with a request for corrected gain statements, error bars or code, and a clearer efficiency comparison. I would not cite it as SOTA until those are addressed.","headline":"A plausible, well-ablated SSM temporal-fusion system for online HD map construction, with a real SOTA story on nuScenes but an internally inconsistent gain claim and no variance estimates yet.","tokens_in":12858,"tokens_out":2607,"would_cite":false,"duration_ms":26843,"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":"MambaMap shows that fusing past BEV features and instance queries through a gated state space model improves online vectorized HD map construction, outperforming state-of-the-art methods on nuScenes and Argoverse2.","keywords":["online HD map construction","vectorized map","state space model","temporal fusion","memory bank","BEV features","instance queries","autonomous driving"],"falsifier":"A decisive test: on a sequence with closely spaced or crossing lanes, compare MambaMap's Hungarian assignments with ground-truth instance correspondences over frames. If L2 query matching often swaps identities, the fused queries should show visible errors at crossings, and replacing the L2 cost with the ground-truth assignment should raise mAP; if instead the model stays accurate and the permutation matches truth, the assumption holds.","tokens_in":11831,"feed_emoji":"🛣️","tokens_out":5168,"duration_ms":61428,"temperature":0.7,"pith_summary":"This paper tries to establish that online vectorized HD map construction benefits from long-range temporal fusion carried out by a state space model instead of a single recurrent hidden state or stacked attention over many frames. The proposed framework, MambaMap, keeps a small memory bank of historical BEV features and instance queries, warps them to the current frame, and fuses them through a gated state space block with multi-directional scanning. It also reorders historical instance queries to match the current frame before scanning them in instance-first and temporal-first orders. The paper reports that this consistently beats previous state-of-the-art methods on nuScenes and Argoverse2, including geographically disjoint splits, at both 60x30m and 100x50m perception ranges.","feed_headline":"State-space temporal fusion tops HD map benchmarks","feed_subtitle":"MambaMap beats StreamMapNet by up to 3.9 mAP on nuScenes and Argoverse2, at small and large ranges.","key_machinery":"The carrying device is the Gated State Space block: each 1D sequence is projected into a lower-dimensional state-space path (using the DSS diagonal state-space layer) and a higher-dimensional gating signal, then the gated output is element-wise multiplied and added to the input as a residual. Applied to four scanning directions over the BEV feature map and to the two query-order sequences, it gives linear-time receptive fields over the warped history. The other load-bearing piece is the memory bank of size N=4 with Hungarian query-order alignment, which supplies the temporal context without storing long redundant sequences.","core_discovery":"The central claim is that selective temporal fusion in state space is the right mechanism for streaming HD map construction. Concretely, MambaMap stores the last N refined BEV features and N refined instance query sets, aligns historical BEV features to the current frame with a 4x4 ego-motion warp, fuses them via a Gated State Space block (a diagonal state-space layer with GELU gating and residual connection), and scans the fused BEV map in left, right, up, and down directions before averaging. At the instance level, historical queries are reordered by Hungarian matching on L2 distance and scanned both instance-first (all instances of a frame together) and temporal-first (same instance slot across frames) to produce fused queries. The author's claim is that this design captures long-range temporal dependencies with linear complexity and yields state-of-the-art mAP: 67.3/66.5 on nuScenes small/large range, 64.9/60.7 on Argoverse2, with a 3.9 mAP gain over the StreamMapNet baseline on the small nuScenes range.","pith_inferences":["One testable extension is to make the instance matching cost part of training (e.g., via a learned matching head) and see whether the 3.9 mAP gain grows or shrinks; the paper's ablations compare costs but do not analyze when L2 matching fails.","The same memory-bank-plus-gated-SSM recipe could be transferred to other streaming BEV tasks, such as 3D object detection or motion prediction, where per-instance temporal identity is also important.","The saturation at N=4 suggests the model learns which historical frames matter; an explicit learned importance weighting over the memory bank might push the operating point past N=4, a possibility the paper does not explore.","Because the method effectively tracks instances through query matching, it could produce stable instance IDs over time from the same machinery, which would be useful for downstream planning."],"forward_implications":["On the evidence reported, SSM-based fusion is a direct alternative to recurrent hidden-state streaming for vectorized map construction, with linear complexity in the number of historical frames.","A memory bank of four frames plus gated selection is enough to recover most of the temporal benefit; growing the bank to six frames slightly hurts accuracy, suggesting redundancy and noise start to dominate.","Because the gains hold on geographically disjoint splits, the temporal mechanism appears to generalize to unseen city layouts rather than memorizing training locations.","Instance matching by plain L2 distance between query vectors is sufficient for temporal consistency, avoiding extra supervision or ground-truth correspondences."],"supporting_citations":[{"why":"StreamMapNet is the primary streaming baseline and the source of the two-stage training schedule and comparison results MambaMap seeks to beat.","marker":"[10]"},{"why":"DSS supplies the diagonal state space layer used inside the Gated State Space block for sequence modeling.","marker":"[29]"},{"why":"Gated state spaces provide the gating mechanism that the paper adapts for selective information integration.","marker":"[30]"},{"why":"BEVFormer serves as the BEV feature encoder that produces the initial BEV representation from multi-view images.","marker":"[22]"},{"why":"SQD-MapNet is a competing query-denoising temporal method compared on both the original and new splits.","marker":"[15]"},{"why":"MemFusionMap is a competing memory-based temporal fusion method evaluated on the geographically disjoint split.","marker":"[17]"},{"why":"Mamba is the conceptual inspiration for applying selective state space models to temporal fusion, though the final implementation adopts DSS.","marker":"[21]"}],"fun_headline_variants":["MambaMap: State-space temporal fusion for online HD maps","State-space model fuses temporal data for HD maps","MambaMap: Linear-complexity temporal fusion for HD maps","State-space scanning boosts HD map accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Euclidean distance between two learned instance query vectors reliably identifies whether they represent the same road element across frames; if that fails, the ordering step can mix up different map elements and corrupt the temporal fusion.","fun_headline_variants_meta":{"raw":{"variants":["MambaMap: State-space temporal fusion for online HD maps","State-space model fuses temporal data for HD maps","MambaMap: Linear-complexity temporal fusion for HD maps","State-space scanning boosts HD map accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000931,"raw_usage":{"total_tokens":4012,"prompt_tokens":1001,"completion_tokens":3011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":2946}},"tokens_in":617,"tokens_out":3011,"duration_ms":28860,"temperature":1.0,"reasoning_tokens":2946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:41:40.377763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: on a sequence with closely spaced or crossing lanes, compare MambaMap's Hungarian assignments with ground-truth instance correspondences over frames. If L2 query matching often swaps identities, the fused queries should show visible errors at crossings, and replacing the L2 cost with the ground-truth assignment should raise mAP; if instead the model stays accurate and the permutation matches truth, the assumption holds.","supporting_citations":[{"cited_title":"Streammapnet: Streaming mapping network for vectorized online hd map construction","cited_arxiv_id":null,"evidence_quote":"StreamMapNet is the primary streaming baseline and the source of the two-stage training schedule and comparison results MambaMap seeks to beat."},{"cited_title":"Bevformer: learning bird’s-eye- view representation from lidar-camera via spatiotemporal transformers","cited_arxiv_id":null,"evidence_quote":"BEVFormer serves as the BEV feature encoder that produces the initial BEV representation from multi-view images."},{"cited_title":"Stream query denoising for vectorized hd-map construction","cited_arxiv_id":null,"evidence_quote":"SQD-MapNet is a competing query-denoising temporal method compared on both the original and new splits."}],"review_version":1}