{"id":"95684f94-01c8-4b53-be92-b7dbc2ac9af6","arxiv_id":"2606.08866","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Replacing context heads in six CNN segmentation models with geometry-guided Mamba yields consistent mIoU gains on Cityscapes at moderate extra computational cost.","lead":"This paper tests replacing standard context heads in CNN semantic segmentation models with a geometry-guided Mamba module that uses boundary and flow cues. Smart generalists might read it because it explores whether state-space models can serve as lighter, more accurate alternatives for gathering long-range image context in practical vision systems.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"The claim that geometry-guided cues integrate cleanly into Mamba's selective scan as a true plug-and-play replacement rests on an unverified assumption about scan-order modulation and cue fusion.","rationale":"The reader's weakest_assumption directly identifies the integration step as load-bearing; the abstract-only limitation is now superseded by the full-text placeholder, but the same integration detail remains the least-secured precondition for the headline claim. No other internal inconsistency is visible from the given material.","tokens_in":1703,"tokens_out":343,"duration_ms":12361,"concrete_test":"Extract the exact equations or pseudocode for the modified selective scan (likely in §3 or the method figure) and verify whether the geometric cues enter only as an additive bias to the input-dependent parameters (B, C, Δ) or whether they alter the scan order or discretization; if the latter, recompute GFLOPs and re-train one model (e.g., PSPNet) with the original scan to measure mIoU drop.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical result (consistent mIoU gains on Cityscapes after swapping context heads in six architectures) requires that boundary and centripetal-flow signals modulate the selective scan without altering its core recurrence, introducing new hyperparameters, or creating architecture-specific failure modes. The abstract states only that guidance is 'injected into the selective scan process'; if this injection changes the scan direction selection, adds per-layer normalization, or requires retuning the SSM state dimension for each backbone, the 'plug-and-play' and 'moderate extra GFLOPs' assertions become conditional on details not visible in the provided text.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript claims that Geometry-Guided Mamba (G-Mamba) can be generalized as a plug-and-play context module for CNN semantic segmentation by injecting boundary and centripetal-flow geometric cues into the selective scan process. Replacing the context heads of six models (DeepLabV3+, DANet, CCNet, PSPNet, PSANet, OCRNet) while keeping a ResNet-101 backbone fixed yields consistent mIoU gains on Cityscapes at 1024×1024 resolution with only moderate extra GFLOPs.","tokens_in":1818,"tokens_out":435,"duration_ms":16422,"significance":"If the reported gains hold under the stated conditions, the work would indicate that geometry-guided SSM modules offer a computationally lighter alternative to ASPP, PPM, or attention-based context heads while preserving or improving boundary accuracy. The evaluation across six distinct architectures strengthens the generality claim relative to single-backbone studies.","major_comments":[{"comment":"The description of cue injection (abstract and method) states only that guidance is 'injected into the selective scan process' without specifying whether this modulates scan direction selection, adds per-layer normalization, or requires retuning the SSM state dimension. This detail is load-bearing for the 'plug-and-play' and 'no model-specific retuning' assertions; if any of these occur, the moderate-GFLOPs claim becomes conditional.","section":"Abstract / Method"},{"comment":"The central empirical claim of 'consistent mIoU gains' is presented without numerical deltas, standard deviations across runs, or explicit controls for post-hoc hyperparameter choices in the provided text. This prevents assessment of whether the gains exceed what could arise from re-optimizing the original context heads.","section":"Abstract / Experiments"}],"minor_comments":[{"comment":"The citation to the original DGM-Net work should explicitly state which components are reused versus newly generalized for the plug-and-play setting.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. The comments highlight areas where additional clarity is needed regarding the method details and empirical reporting. We address each major comment below.","responses":[{"response":"We agree that the current description of cue injection lacks the necessary specificity. In the revised manuscript, we will expand Section 3 (Method) to explicitly state that the boundary and centripetal-flow cues are used solely to modulate scan direction selection and the selective parameters within the existing SSM formulation. No per-layer normalization is added, the SSM state dimension remains unchanged, and no model-specific retuning of hyperparameters is performed. The geometric modulation is computed from fixed, pre-derived cues and applied uniformly, which we will illustrate with pseudocode. This clarification will reinforce that the module is truly plug-and-play while keeping the reported GFLOPs increase moderate and unconditional.","revision_made":"yes","referee_comment":"[Abstract / Method] The description of cue injection (abstract and method) states only that guidance is 'injected into the selective scan process' without specifying whether this modulates scan direction selection, adds per-layer normalization, or requires retuning the SSM state dimension. This detail is load-bearing for the 'plug-and-play' and 'no model-specific retuning' assertions; if any of these occur, the moderate-GFLOPs claim becomes conditional."},{"response":"The experiments section contains a table reporting per-model mIoU values before and after replacement, from which the consistent gains can be computed. However, we acknowledge that the abstract and main text do not include explicit deltas, standard deviations, or a dedicated statement on hyperparameter controls. In revision, we will update the abstract to include representative numerical deltas and add a paragraph in the experiments section stating that all replacements used the identical training protocols, schedules, and hyperparameters as the original models (no post-hoc search or retuning). We will also report standard deviations from three runs with different seeds where available, or note fixed-seed single-run results for reproducibility.","revision_made":"yes","referee_comment":"[Abstract / Experiments] The central empirical claim of 'consistent mIoU gains' is presented without numerical deltas, standard deviations across runs, or explicit controls for post-hoc hyperparameter choices in the provided text. This prevents assessment of whether the gains exceed what could arise from re-optimizing the original context heads."}],"tokens_in":1342,"tokens_out":517,"duration_ms":21468,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main point is that geometry-guided Mamba can replace the usual context heads (ASPP, PPM, attention) in DeepLabV3+, DANet, CCNet, PSPNet, PSANet, and OCRNet while keeping the ResNet-101 backbone fixed. They claim this produces steady mIoU lifts on Cityscapes at 1024x1024 with only moderate added GFLOPs.\n\nWhat is new is the plug-and-play replacement study across those six models. Prior DGM-Net work introduced the directional geometric cues and Mamba scan, but this version tests whether the same module works as a drop-in upgrade rather than requiring a full new network. That systematic swap is the concrete addition and gives a practical sense of how well the idea travels.\n\nThe approach does one thing cleanly: it keeps the backbone and training setup constant so the comparison isolates the context module change. Reporting GFLOPs alongside accuracy is also useful for anyone who actually deploys these models.\n\nThe soft spots sit in the experimental reporting and the fusion mechanics. The abstract gives no deltas, no variance across runs, and no ablation on how boundary and centripetal-flow signals are injected into the selective scan. If that injection requires per-model retuning of state dimension, scan order, or extra normalization, the \"plug-and-play\" and \"moderate extra GFLOPs\" claims become conditional. The stress-test concern about scan-order modulation is worth verifying in the full text; if the paper shows the recurrence itself stays untouched and no new hyperparameters appear, the worry drops. Otherwise the gains could partly reflect hidden fitting.\n\nThis is for people building or tuning CNN segmentation pipelines who want a lighter alternative to attention-based context. It is not paradigm-shifting but could be a useful reference if the numbers and controls hold. I would send it to peer review because the multi-model test is a reasonable way to check generalizability, even if the paper will likely need clearer protocol and ablation sections.","headline":"This extends DGM-Net by swapping G-Mamba into six existing CNN segmentation heads and reports consistent Cityscapes mIoU gains with moderate extra GFLOPs, but the integration details and numerical evidence need checking.","tokens_in":2338,"tokens_out":490,"would_cite":false,"duration_ms":19022,"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":"Geometry-guided Mamba replaces conventional context heads in CNN segmentation networks and yields consistent mIoU gains on Cityscapes at moderate extra cost.","keywords":["semantic segmentation","Mamba","plug-and-play module","geometric guidance","context aggregation","Cityscapes","state space model","CNN enhancement"],"falsifier":"Inserting the module into the six CNN models and measuring either no mIoU improvement or a large GFLOPs increase on the Cityscapes validation set at 1024 by 1024 resolution would disprove the claim of practical enhancement.","tokens_in":2607,"feed_emoji":"🖼️","tokens_out":693,"duration_ms":21407,"temperature":0.7,"pith_summary":"The paper shows that Directional Geometric Mamba can function as a modular context aggregation unit inside existing CNN segmentation models rather than requiring an entirely new architecture. Geometric signals from object boundaries and centripetal flow are injected directly into the selective scan to steer long-range feature propagation. The module is swapped into the context heads of six representative networks while the ResNet-101 backbone stays fixed. Experiments on Cityscapes at 1024 by 1024 resolution record steady accuracy lifts accompanied by only moderate added GFLOPs. This positions geometry-guided state-space modules as drop-in alternatives to atrous pyramid pooling, pyramid pooling, or attention-based heads.","feed_headline":"Geometry-guided Mamba lifts CNN segmentation accuracy","feed_subtitle":"Plug-in selective scans using boundary and centripetal cues replace standard heads and deliver consistent mIoU gains with moderate added GFL","key_machinery":"The directional geometric Mamba selective scan that modulates long-range feature propagation with explicit boundary and centripetal-flow cues.","core_discovery":"The central claim is that geometry-guided Mamba, when used as a plug-and-play context module by injecting boundary and centripetal-flow cues into the selective scan, improves segmentation performance across multiple CNN architectures on Cityscapes while adding only moderate computational overhead.","pith_inferences":["The approach may transfer to other dense-prediction tasks such as instance segmentation or monocular depth estimation that also benefit from boundary-aware long-range context.","If the geometric cues remain stable across domains, the module could allow lighter backbones by reducing the need for heavy attention or pyramid modules.","Combining the fixed geometric guidance with additional learnable parameters inside the scan could produce further accuracy or efficiency trade-offs."],"forward_implications":["The same geometry-guided module produces consistent mIoU gains when substituted for the context heads of DeepLabV3+, DANet, CCNet, PSPNet, PSANet, and OCRNet.","Extra computation stays moderate in GFLOPs when the models process 1024 by 1024 images.","Long-range context can be aggregated under explicit geometric constraints instead of relying solely on learned attention or fixed pooling.","State-space-model context modules can serve as practical alternatives or complements to standard CNN context heads without altering the backbone."],"fun_headline_variants":["G-Mamba as plug-and-play context for CNN segmentation","mIoU gains with geometry-guided Mamba in CNN models","Selective Mamba scans using boundary cues for segmentation","G-Mamba module tested in DeepLabV3+ and similar networks"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The boundary and centripetal-flow signals integrate cleanly into the Mamba selective scan without model-specific retuning or new failure modes when inserted into the six tested CNN architectures.","fun_headline_variants_meta":{"raw":{"variants":["G-Mamba as plug-and-play context for CNN segmentation","mIoU gains with geometry-guided Mamba in CNN models","Selective Mamba scans using boundary cues for segmentation","G-Mamba module tested in DeepLabV3+ and similar networks"]},"model":"grok-4.3","cost_usd":0.007539,"raw_usage":{"total_tokens":3428,"prompt_tokens":610,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":75387000,"prompt_tokens_details":{"text_tokens":610,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2751,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":610,"tokens_out":67,"duration_ms":12998,"temperature":1.0,"reasoning_tokens":2751,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T18:20:57.819633+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Inserting the module into the six CNN models and measuring either no mIoU improvement or a large GFLOPs increase on the Cityscapes validation set at 1024 by 1024 resolution would disprove the claim of practical enhancement.","supporting_citations":[],"review_version":1}