{"id":"4522cb9a-7737-488a-87ff-2666d05f3111","arxiv_id":"2508.20376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A bidirectional, multi-scale Mamba scan for cross-task interaction improves multi-task dense prediction accuracy on NYUD-V2 and PASCAL-Context over prior state-of-the-art methods.","lead":"BIM, a Mamba-based decoder, scans image features in two directions and at multiple scales so several dense-prediction tasks can share information without quadratic compute. It reports state-of-the-art numbers on NYUD-V2 and PASCAL-Context and releases code, making it a useful reference for multi-task perception models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA-superiority claim rests on single-run comparisons without variance; several PASCAL deltas are tiny or tied and NYUD hyperparameters are test-set-selected.","rationale":"The reader's weakest-assumption analysis correctly identifies the absence of multiple seeds and significance testing as the most load-bearing weakness. The paper's own numbers contain both large favorable deltas (NYUD-V2 Semseg, Depth) and near-zero or tied deltas (PASCAL Saliency, Normal), which is exactly the pattern expected when single-run measurements are compared without variance. The conditional verdict is appropriate because the architectural description is coherent, the linear-complexity claim is empirically supported by Table 10's constant per-task FLOP increment, and the ablations are directionally consistent. However, the headline 'superiority' claim needs controlled multi-seed comparisons before it can be accepted. I did not find an internal inconsistency in the BI-Scan or MS-Scan derivations that would require rejection; the main risk is empirical, not architectural. My concrete test would directly settle whether the reported gains exceed run-to-run noise.","tokens_in":14952,"tokens_out":4893,"duration_ms":50598,"concrete_test":"Reimplement or obtain the released code for MTMamba and BIM, then train both under an identical pipeline (same augmentations, optimizer, learning rate, iterations, batch size, and loss weights) on NYUD-V2 and PASCAL-Context with at least three random seeds. Report mean and standard deviation for every metric, plus paired deltas with 95% confidence intervals. Before evaluating on the test sets, fix all hyperparameters using a validation split. If the NYUD-V2 Semseg mIoU delta falls below roughly 1 point or the PASCAL parsing/boundary/semseg deltas include zero in their confidence intervals, the SOTA-superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that BIM outperforms state-of-the-art methods on NYUD-V2 and PASCAL-Context—is supported only by a single evaluation per configuration. Tables 1, 2, 4–8 and the ablation figures report no seeds, standard deviations, or significance tests. On PASCAL-Context the paper asserts 'superior performance on all tasks compared to MTMamba,' but Table 2 shows saliency exactly ties at 84.14 and normal estimation improves by only 0.01 mErr; boundary, semseg, and parsing deltas are 0.25, 0.14, and 0.53 mIoU, respectively. These deltas are plausibly within run-to-run variation. On NYUD-V2 the headline deltas are larger (Semseg +1.58 mIoU, Depth −0.0333 RMSE), but they are also single runs against published MTMamba numbers rather than a controlled reimplementation under identical training conditions. Any difference in augmentation, iteration count, learning-rate schedule, or seed is therefore confounded with the architectural contribution. In addition, the scan-scale and scan-number configurations in Tables 6–7 appear to be selected using test-set performance, which risks selection overfitting. Because the claimed superiority is exactly the quantity that would collapse if seed-level variance approaches these magnitudes, the paper's central empirical conclusion is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BIM, a Mamba-based decoder for multi-task dense prediction. The decoder replaces MTMamba's cross-task scan with BI-Scan, which concatenates task features into long sequences in task-first and position-first modes and processes them bidirectionally with an SSM, and with MS-Scan, which tokenizes at multiple scales before SSM processing. Experiments on NYUD-V2 and PASCAL-Context report improvements over MTMamba and other baselines (e.g., NYUD-V2 semantic segmentation mIoU 57.40 vs. 55.82; PASCAL-Context parsing mIoU 73.15 vs. 72.62), and ablations attribute gains to both scan mechanisms. The paper's central claim is that BIM achieves state-of-the-art multi-task dense prediction while retaining linear interaction complexity.","tokens_in":15190,"tokens_out":7611,"duration_ms":71110,"significance":"The design is well specified and the component ablations are internally consistent, which makes the proposal credible as a method. The BI-Scan idea of injecting task-order and position-order serializations into a linear-complexity SSM is a useful and clearly explained contribution, and MS-Scan is a sensible adaptation of Mamba to multi-scale visual structure. If the numbers are reproducible, the NYUD-V2 improvements in semantic segmentation and depth are nontrivial. The availability of code is also a strength. However, the empirical evidence for the 'state-of-the-art' claim is currently thin: every configuration is evaluated once, several PASCAL-Context deltas are tiny or zero, and hyperparameters are selected on the test set. These issues are fixable but need to be addressed before the central claim can be accepted.","major_comments":[{"comment":"The central SOTA claim is based on a single evaluation per configuration with no seeds, standard deviations, or significance tests. On PASCAL-Context, Table 2 shows saliency exactly tying MTMamba at 84.14, normal estimation improving by only 0.01 mErr, and boundary/semseg/parsing changing by 0.25, 0.14, and 0.53 mIoU, respectively; these deltas are within the range of run-to-run variation typical for dense prediction benchmarks. The NYUD-V2 headline deltas are larger, but they are single runs compared against published MTMamba numbers rather than a controlled reimplementation. Please report mean and standard deviation over at least three seeds (or an appropriate significance test) and clarify whether the MTMamba baseline was retrained under exactly the same data pipeline.","section":"Section 4.2, Tables 1-2"},{"comment":"The final configuration is selected using NYUD-V2 test-set numbers. Table 5 chooses task order S-D-N-B, Table 6 chooses scan scale {1,4}, and Table 7 chooses Type 2, all based on the test metrics reported in those tables; the final Table 1 numbers are then produced by the selected configuration. Selection on the test set can inflate the reported gains and makes the headline improvements optimistic. Please move configuration selection to a held-out validation split, or report the test-set results of all configurations and disclose the selection protocol.","section":"Section 4.3, Tables 5-7"},{"comment":"The loss weights lambda_t in Eq. (10) are never specified in the implementation details of Section 4.1. Multi-task performance is sensitive to loss weighting, so the reader cannot reproduce the reported numbers without this information. Please state the values used for all tasks on both datasets.","section":"Section 3.4, Eq. (10)"},{"comment":"The comparisons in Tables 1 and 2 mix numbers taken from prior publications with numbers produced by the authors' own pipeline. For the central comparison against MTMamba, the difference between published numbers and a controlled reimplementation is confounded with the architectural contribution. Please either retrain all relevant baselines with the same augmentation, iteration count, learning-rate schedule, and seeds, or clearly restrict the claim to 'our implementation configuration.'","section":"Section 4.2, Table 1"}],"minor_comments":[{"comment":"There is a typo in the first contribution bullet: 'Bidricational' should be 'Bidirectional.'","section":"Section 1, Contributions"},{"comment":"The labels 'Intercation Complexity' should read 'Interaction Complexity.'","section":"Figure 1"},{"comment":"The label 'Flap & Concat' appears to be a typo; it should likely read 'Flatten & Concat.'","section":"Figure 3"},{"comment":"The sentence introducing F_{t,N}^s contains a duplicated 'is': 'F_{t,N}^s is is derived' should read 'F_{t,N}^s is derived.'","section":"Section 3.3, Eq. (3)"},{"comment":"The text says multi-scale scanning is introduced 'without increasing computational complexity,' but Table 3 reports BIM at 547 GFLOPs versus MTMamba's 541 GFLOPs. Please either provide a precise FLOP calculation or soften the statement to 'with small additional cost.'","section":"Section 3.3, MS-Scan"},{"comment":"The description of window tokenization assumes that the channel dimension C is divisible by the number of branches N; please state this assumption explicitly.","section":"Section 3.3, MS-Scan"},{"comment":"DBIM is included in Table 3 but is only described in the supplementary material; either describe it in the main text or remove it from the main table.","section":"Table 3"},{"comment":"The bar charts in Figure 6 do not show numerical values on the y-axis, making the claimed differences hard to verify; please add value labels or a table.","section":"Figure 6"},{"comment":"The phrase 'preliminary decoder has an output channel number of 768' is unclear; presumably 'initial decoder channel count' is meant.","section":"Section 4.1, Implementation Details"},{"comment":"The caption describes a 'quantitative visual analysis,' but the figure is qualitative; please rephrase.","section":"Supplementary, Figure 8"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe thing to know: the actual architecture contribution is real and clearly presented; the empirical case for 'state of the art' is not yet. BIM's BI-Scan (task-first/position-first bidirectional scanning across tasks) and MS-Scan (multi-scale tokenization) are genuinely different from MTMamba, VMamba, and InvPT. The equations are concrete, the complexity argument is linear and plausible, and the ablations in Tables 4–7 consistently point the same way. Credit where due: this is a coherent module design, and the authors released code.\n\nSoft spots are all on the measurement side, and they are not minor. Every number in the paper is a single run. No seeds, no error bars, no significance tests. On PASCAL-Context the 'superior on all tasks' claim sits on a tie in saliency (84.14 vs 84.14) and a 0.01 improvement in normal error. Those deltas are well inside plausible run-to-run variation. On NYUD-V2 the headline deltas are bigger (semseg +1.58 mIoU, depth −0.033 RMSE), but those are single runs against published MTMamba numbers, not a controlled reimplementation under identical training conditions. Augmentation, iteration count, LR schedule, seed—any of those differences is confounded with the architectural change. On top of that, the scan scale and task order are chosen using test-set performance (Tables 5–6), which risks selection overfitting. The paper's own Table 5 shows several task orders within a hair of each other; the margin between the chosen order and the next is larger than the variance you'd want to see.\n\nThe structure and background are fine. Citations cover the relevant Mamba and multi-task literature. I don't see circular reasoning or invented entities. The claim is simply stronger than the evidence.\n\nWho this is for: people working on Mamba-based decoders and multi-task dense prediction. It's a solid architecture paper that needs more rigorous evaluation. I would send it to peer review rather than desk-reject, but request multi-seed runs, validation-based selection of hyperparameters, and either a reimplementation of MTMamba or a clear caveat that the comparison is against published numbers. If the deltas survive that, the paper is a real contribution.\n\nRecommendation: engage with it, but insist on the variance analysis.","headline":"A genuinely new Mamba decoder module with clear ablations, but the SOTA claim rests on single-run, test-set-selected numbers.","tokens_in":15735,"tokens_out":2540,"would_cite":false,"duration_ms":23048,"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":"Bidirectional Interaction Mamba (BIM) claims to preserve complete cross-task information in dense prediction at linear interaction complexity, and reports state-of-the-art numbers on NYUD-V2 and PASCAL-Context.","keywords":["multi-task dense prediction","Mamba","Bidirectional Interaction Scan","cross-task interaction","state space model","multi-scale scan","semantic segmentation","depth estimation"],"falsifier":"Re-run BIM and MTMamba on NYUD-V2 and PASCAL-Context under the paper's exact training protocol with, say, five random seeds each, and compare the BIM-minus-MTMamba deltas (for example +1.58 semseg mIoU and $-0.033$ depth RMSE on NYUD-V2) against the seed-level standard deviation. If those deltas do not exceed the seed noise, the claimed superiority over the previous best method is not established.","tokens_in":14747,"feed_emoji":"🖼️","tokens_out":10274,"duration_ms":82019,"temperature":0.7,"pith_summary":"This paper tries to settle the trade-off between interaction completeness and computational cost in multi-task dense prediction. It proposes the Bidirectional Interaction Mamba (BIM), a decoder built on Mamba's state-space model, whose Bidirectional Interaction Scan (BI-Scan) serializes task features in both task-first and position-first orders so that every task's full spatial information participates in cross-task exchange while interaction complexity stays linear in the number of tasks. A second mechanism, Multi-Scale Scan (MS-Scan), splits features across channels and scans them at several window scales to capture scene structure at multiple granularities. The paper reports state-of-the-art results on NYUD-V2 and PASCAL-Context, including semantic segmentation mIoU of 57.40 on NYUD-V2 versus 55.82 for the prior best method, at 547 GFLOPs. If these results hold, BIM shows that a dense-prediction decoder can exchange complete task information across all tasks at linear cost, removing the main obstacle to scaling cross-task interaction to many tasks.","feed_headline":"Bidirectional Mamba scan tops multi-task dense-prediction benchmarks","feed_subtitle":"Pairing task-first and position-first scans, BIM beats prior state of the art on NYUD-V2 and PASCAL-Context.","key_machinery":"The two load-bearing mechanisms are BI-Scan and MS-Scan. BI-Scan is a bidirectional sequence construction for cross-task interaction: task-first mode serializes each task's tokens by a spatial scan pattern and concatenates the resulting subsequences along the task order; position-first mode regroups tokens that share a spatial position across tasks into length-$T$ subsequences and concatenates those along the scan pattern. Both directions are passed through the selective state-space model of Mamba, and the backward pass is reversed along the task dimension before concatenation. MS-Scan partitions the feature channels into branches, tokenizes each branch at a different window scale (for example $\\{1,4\\}$), runs SS2D four-way scanning in each branch, and fuses the branches back by inverse windowing and channel concatenation. The paper's argument is that this combination preserves every task's full spatial information during interaction (no information is compressed through a bridge representation) while keeping interaction complexity $O(n)$, and supplies multi-scale scene structure that different tasks can weight through gating.","core_discovery":"The central claim is that cross-task interaction in dense prediction does not require choosing between information integrity and linear complexity. BIM's BI-Scan takes each task's normalized feature map, splits it along channels, and builds two complementary sequences: task-first, where each task's full spatial scan is concatenated along the task order into one long sequence of length $T \\times (H \\times W)$, and position-first, where spatially aligned tokens from all tasks are gathered into $H \\times W$ subsequences of length $T$. Both are processed by the selective state-space model from Mamba, giving each task access to the complete information of the other tasks while the cost grows only linearly with $T$. MS-Scan adds multi-granularity scene modeling by partitioning channels and scanning non-overlapping windows at multiple scales, with SS2D-style four-way scans per branch. On NYUD-V2 the paper reports Semseg mIoU 57.40, Depth RMSE 0.4733, Normal mErr 18.55, Boundary odsF 78.72, and on PASCAL-Context Semseg mIoU 81.25, Parsing mIoU 73.15, Saliency maxF 84.14, Normal mErr 14.13, Boundary odsF 79.05, with 547 GFLOPs; the ablation tables attribute the gains to the bidirectional scan and to the multi-scale scan.","pith_inferences":["Going beyond the paper: because the two scan orderings are defined purely by serialization, the same BI-Scan structure could be paired with any linear sequence encoder, not just Mamba's selective state-space model; a testable extension is whether the gains persist with a different linear sequence layer, which would separate the scan design from the state-space backbone.","Going beyond the paper: the reported per-task cost that stays constant as tasks are added suggests BIM is a plausible decoder for many-task settings (well beyond four tasks) where pairwise or transformer interactions become prohibitive; the paper does not run that regime, but it is a direct quantitative prediction of the architecture.","Going beyond the paper: the ablation shows task-first and position-first scans are complementary, so a natural next experiment is applying BI-Scan at multiple decoder depths or interleaving it with MS-Scan more than once per MFR block to see whether the current single application is the saturation point."],"forward_implications":["If BIM is correct, the completeness-efficiency trade-off in cross-task interaction is not inherent: a decoder can give every task access to every other task's full features at linear interaction cost.","The two scan modes are complementary: combining task-first and position-first modes outperforms either alone, with negligible added computation, so future multi-task decoders can treat the two orderings as paired modules.","Multi-scale scanning improves Mamba-based image modeling: the $\\{1,4\\}$ scan-scale setting yields the largest multi-task gain, indicating that granularity diversity, not just longer sequences, matters for dense prediction.","The design scales predictably with task count: adding a fourth task on NYUD-V2 costs the same +86 GFLOPs and +42M parameters as adding the third task, consistent with the claimed linear complexity.","Task-order robustness means the method does not need per-dataset task-ordering search: several fixed orders perform similarly, and only random order degrades somewhat."],"supporting_citations":[{"why":"Supplies the selective state-space model whose linear-time sequence modeling is the computational core of both scans.","marker":"[9]"},{"why":"MTMamba is the previous best mamba-based multi-task decoder; it is the main comparison baseline and the ablation baseline.","marker":"[16]"},{"why":"Provides the SS2D four-way scan and the scan patterns that BI-Scan and MS-Scan adapt.","marker":"[17]"},{"why":"The pretrained Swin Transformer encoder on which all experiments and ablations are built.","marker":"[18]"},{"why":"NYUD-V2 is one of the two benchmark datasets that supply the main results.","marker":"[25]"},{"why":"PASCAL-Context is the other benchmark dataset that supplies the main results.","marker":"[3]"},{"why":"InvPT is a transformer-based baseline whose numbers anchor the state-of-the-art comparison.","marker":"[35]"},{"why":"TaskPrompter is a transformer-based baseline used in the state-of-the-art comparison.","marker":"[36]"},{"why":"InvPT++ is the extended inverted-pyramid baseline compared on both datasets.","marker":"[37]"},{"why":"MQTransformer is a transformer-based baseline compared on both datasets.","marker":"[32]"}],"fun_headline_variants":["Bidirectional Mamba scan wins dense-prediction benchmarks at linear cost","Linear-complexity Mamba scan boosts dense-prediction accuracy","Bidirectional Mamba scan beats state of the art on NYUD-V2 and PASCAL-Context","Bidirectional Mamba scan: full cross-task interaction in linear time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported accuracy differences between BIM and the previous best method are larger than run-to-run training variation, since every configuration in the comparison tables is evaluated once, with no seeds, standard deviations, or significance tests reported.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional Mamba scan wins dense-prediction benchmarks at linear cost","Linear-complexity Mamba scan boosts dense-prediction accuracy","Bidirectional Mamba scan beats state of the art on NYUD-V2 and PASCAL-Context","Bidirectional Mamba scan: full cross-task interaction in linear time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00172,"raw_usage":{"total_tokens":6853,"prompt_tokens":1043,"completion_tokens":5810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":5726}},"tokens_in":659,"tokens_out":5810,"duration_ms":35902,"temperature":1.0,"reasoning_tokens":5726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:45:52.434518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run BIM and MTMamba on NYUD-V2 and PASCAL-Context under the paper's exact training protocol with, say, five random seeds each, and compare the BIM-minus-MTMamba deltas (for example +1.58 semseg mIoU and $-0.033$ depth RMSE on NYUD-V2) against the seed-level standard deviation. If those deltas do not exceed the seed noise, the claimed superiority over the previous best method is not established.","supporting_citations":[{"cited_title":"Mamba: Linear-time sequence mod- eling with selective state spaces, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the selective state-space model whose linear-time sequence modeling is the computational core of both scans."},{"cited_title":"MTMamba: Enhancing multi-task dense scene understanding by mamba-based de- coders","cited_arxiv_id":null,"evidence_quote":"MTMamba is the previous best mamba-based multi-task decoder; it is the main comparison baseline and the ablation baseline."},{"cited_title":"Inverted pyramid multi-task trans- former for dense scene understanding","cited_arxiv_id":null,"evidence_quote":"InvPT is a transformer-based baseline whose numbers anchor the state-of-the-art comparison."},{"cited_title":"Taskprompter: Spatial-channel multi-task prompting for dense scene understanding","cited_arxiv_id":null,"evidence_quote":"TaskPrompter is a transformer-based baseline used in the state-of-the-art comparison."},{"cited_title":"Invpt++: Inverted pyramid multi- task transformer for visual scene understanding.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024","cited_arxiv_id":null,"evidence_quote":"InvPT++ is the extended inverted-pyramid baseline compared on both datasets."},{"cited_title":"Multi-task learning with multi-query trans- former for dense prediction","cited_arxiv_id":null,"evidence_quote":"MQTransformer is a transformer-based baseline compared on both datasets."}],"review_version":2}