{"id":"b2b410fc-59d8-448e-b22c-12006cfbeb6a","arxiv_id":"2606.20282","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"U²Mamba proposes a two-level nested U-structure with multiscale Mamba U-blocks and hierarchical supervision to achieve competitive results on salient object detection benchmarks.","lead":"The paper introduces U²Mamba, a nested U-shaped neural network using Mamba blocks for detecting the most important objects in images. A smart generalist might read it to understand how state-space models are being adapted from language tasks into computer vision for potentially more efficient image analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly flags the need for measurable gains, but once the full manuscript and public code are taken into account, that assumption is testable rather than load-bearing for internal validity. No other technical flaw (e.g., resolution constraints, supervision scheme, or Mamba integration) appears to undermine the argument on its own terms.","tokens_in":1655,"tokens_out":311,"duration_ms":15828,"concrete_test":"Clone the GitHub repo, retrain U²Mamba on the DUTS training set using the authors' exact protocol, and evaluate on the DUTS-TE, ECSSD, and PASCAL-S test sets; if the reported MAE and max F-measure fall within 1% of the published table entries, the performance claim is corroborated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an empirical one: that the combination of multiscale Mamba U-blocks inside a two-level nested U-structure plus hierarchical supervision yields competitive SOD performance. The manuscript supplies the GitHub implementation and states that extensive experiments were run on standard benchmarks. Because the architecture is fully specified, the code is public, and the claim is framed as a performance result rather than a theoretical guarantee, there is no internal inconsistency or hidden assumption that would falsify the claim without further data. The only remaining uncertainty is whether the reported numbers hold under independent reproduction, but that is external to the argument's soundness.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes U²Mamba, a two-level nested U-structured Mamba network for salient object detection. It introduces multiscale Mamba U-blocks (MMUBs) to increase architectural depth and improve local feature extraction, embeds these within a nested U-structure to aggregate multi-scale receptive fields and long-range context without resolution constraints, and replaces standard deep supervision with hierarchical supervision that applies loss at each level. Extensive experiments on standard SOD benchmarks are reported to demonstrate highly competitive performance against state-of-the-art methods, with source code released at the provided GitHub link.","tokens_in":1757,"tokens_out":382,"duration_ms":18322,"significance":"If the empirical performance claims hold under independent verification, the architecture offers a concrete way to combine Mamba's linear-complexity long-sequence modeling with nested U-Net-style multi-scale fusion and level-wise supervision. The public implementation is a clear strength that enables direct reproducibility and further ablation studies. The work sits at the intersection of state-space models and dense prediction, potentially informing efficient alternatives to transformer-based SOD pipelines.","major_comments":[],"minor_comments":[{"comment":"Abstract: the phrase 'highly competitive performance' is used without any quantitative indicators (e.g., mean F-measure or MAE ranges); adding one or two headline numbers would strengthen the claim for readers who only see the abstract.","section":null},{"comment":"The description of hierarchical supervision versus 'traditional deep supervision' would benefit from an explicit equation or pseudocode showing how the per-level losses are weighted and aggregated, especially since this is presented as a methodological contribution.","section":null},{"comment":"Figure captions and architecture diagrams should explicitly label the two-level nesting and the placement of MMUBs so that the 'nested U-structure' can be traced without ambiguity.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive summary of our manuscript and the recommendation of minor revision. No major comments were provided in the report, so we have no specific points to address at this stage. We will carefully consider any minor suggestions during the revision process and ensure the final version maintains the claimed contributions regarding the nested U-structure, MMUBs, and hierarchical supervision for salient object detection.","responses":[],"tokens_in":1215,"tokens_out":96,"duration_ms":4856,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is the specific two-level nested U with multiscale Mamba U-blocks and a hierarchical supervision scheme instead of standard deep supervision. That combination is presented as new, and the authors supply a GitHub link, which lets others check the implementation directly.\n\nThe paper does the usual things well for this kind of work: it runs on standard SOD benchmarks, claims competitive numbers against recent methods, and avoids obvious internal contradictions in the abstract. The architecture description is concrete enough that the claim is falsifiable by reproduction.\n\nThe soft spots are the usual ones for incremental architecture papers. The novelty reduces to swapping in Mamba blocks and adding one extra nesting level plus a different loss schedule; there is no first-principles reason given why this should outperform simpler Mamba-U-Net baselines by a meaningful margin. Without seeing the ablation tables or error bars in the full text it is hard to judge whether the gains are robust or mainly from tuning. The significance claim stays modest because SOD is already a crowded subfield.\n\nThis is for readers who track every Mamba variant in vision and want another data point with code. A serious referee should look at it because the work is fully specified, empirical, and reproducible rather than hand-wavy. I would send it to review but would not expect it to change the field.","headline":"U²Mamba is a routine nesting of Mamba blocks inside a two-level U-structure plus hierarchical supervision for SOD, with public code but no obvious leap over prior variants.","tokens_in":2225,"tokens_out":347,"would_cite":false,"duration_ms":9284,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"U²Mamba uses a nested U-structure with multiscale Mamba blocks and level-wise supervision to match top salient object detection performance.","keywords":["salient object detection","Mamba","nested U-structure","multiscale blocks","hierarchical supervision","contextual features","U-Net architecture"],"falsifier":"Evaluating U²Mamba on standard SOD benchmarks such as DUTS or ECSSD and finding that its F-measure, MAE, or S-measure fall below those of current leading methods by a clear margin.","tokens_in":2566,"feed_emoji":"🔍","tokens_out":623,"duration_ms":25377,"temperature":0.7,"pith_summary":"The paper presents U²Mamba, a U-structured network for salient object detection that places multiscale Mamba U-blocks inside a two-level nested U-architecture. This design increases network depth for stronger local feature extraction while letting shallow and deep layers contribute different receptive fields to gather more context and long-range information without resolution limits. It replaces conventional deep supervision with a hierarchical scheme that applies loss at every level during training. The resulting model reaches highly competitive accuracy on standard benchmarks against existing methods.","feed_headline":"Nested U-Mamba model matches top results in salient object detection","feed_subtitle":"Multiscale blocks inside the nested structure plus level-wise loss let the network gather broader context across scales.","key_machinery":"The two-level nested U-structure with multiscale Mamba U-blocks (MMUBs), which increases depth and fuses receptive fields across layers to collect contextual information without resolution constraints.","core_discovery":"U²Mamba builds a two-level nested U-structure that embeds multiscale Mamba U-blocks to deepen the network and integrate multi-scale receptive fields, combined with hierarchical supervision that computes loss at each level, producing richer contextual and longer-range features for salient object detection.","pith_inferences":["The nested design could transfer to other dense prediction tasks such as semantic segmentation or edge detection where multi-scale context matters.","Replacing attention mechanisms with Mamba blocks inside the nested U may lower memory use for high-resolution inputs compared with transformer alternatives.","Adding a third nesting level or varying the block scale schedule offers a direct route to test further gains in feature hierarchy."],"forward_implications":["The network integrates receptive fields from both shallow and deep layers to collect richer contextual and longer-range data.","Increased model depth from the multiscale Mamba U-blocks improves local feature extraction.","Hierarchical supervision during training replaces single top-level loss and applies loss at each level.","The full architecture achieves highly competitive results against state-of-the-art methods without dataset-specific adjustments."],"fun_headline_variants":["Two-level nested U-Mamba with multiscale blocks for salient detection","U2Mamba nests Mamba U-blocks to integrate multi-scale receptive fields","Level-wise loss in nested U-structure improves contextual features for SOD","Mamba U-blocks deepen U2Mamba for longer-range data in object detection"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The assumption that the nested U-structure, multiscale Mamba blocks, and hierarchical supervision together deliver measurable performance gains over prior Mamba and U-Net baselines on salient object detection tasks.","fun_headline_variants_meta":{"raw":{"variants":["Two-level nested U-Mamba with multiscale blocks for salient detection","U2Mamba nests Mamba U-blocks to integrate multi-scale receptive fields","Level-wise loss in nested U-structure improves contextual features for SOD","Mamba U-blocks deepen U2Mamba for longer-range data in object detection"]},"model":"grok-4.3","cost_usd":0.006166,"raw_usage":{"total_tokens":2879,"prompt_tokens":610,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":61662000,"prompt_tokens_details":{"text_tokens":610,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2192,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":610,"tokens_out":77,"duration_ms":20537,"temperature":1.0,"reasoning_tokens":2192,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T17:44:10.793650+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Evaluating U²Mamba on standard SOD benchmarks such as DUTS or ECSSD and finding that its F-measure, MAE, or S-measure fall below those of current leading methods by a clear margin.","supporting_citations":[],"review_version":1}