{"id":"0e022b03-ebd2-426f-89f4-a26ca6b36392","arxiv_id":"2412.18178","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical RNN with bidirectional minGRU scans reports higher ImageNet accuracy than DeiT backbones at lower high-resolution FLOPs.","lead":"VisionGRU is a new image classification and segmentation network built from a simplified recurrent cell, minGRU, instead of attention layers. It claims competitive accuracy against DeiT while using fewer FLOPs at high image resolutions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline '2.2% / 184%' gains are not supported by the paper's own tables: the +2.2% compares 30M VisionGRU-Ti against 22M DeiT-S, and the same-parameter DeiT-B comparison yields only +1.3%; Fig. 1's FLOPs also use mismatched models and are not lower at 512.","rationale":"Reader's CONDITIONAL verdict is appropriate, but the condition should be stricter than a missing scan-order ablation. The current text makes numerical claims contradicted by its own tables: the only same-parameter ImageNet comparison gives +1.3%, not +2.2%, and the FLOPs figure at 512 shows VisionGRU is not more efficient at all resolutions. This is an internal consistency problem, not a disagreement with field consensus. The scan-order issue raised by the reader is a legitimate missing ablation, but it is less decisive than the parameter mismatch, because the paper's reported advantage weakens or disappears once parameter counts are aligned. The paper does have some positive evidence (VisionGRU-B vs DeiT-B +1.3%, segmentation +0.9 on near-matched params), so outright rejection is not warranted; however, acceptance should require corrected claims and matched-parameter experiments.","tokens_in":10338,"tokens_out":6045,"duration_ms":51011,"concrete_test":"Use the released code to train VisionGRU-Ti with channel widths thinned to match DeiT-S's 22M parameter budget under the same DeiT recipe, and report ImageNet top-1 plus FLOPs at 224 and 512. Independently recompute Fig. 1's values. If top-1 gain over DeiT-S falls below ~1% or FLOPs are not lower at 512, the abstract's '2.2% and 184%' claims fail; if the claims persist, they are redeemed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section I states: 'with the same number of parameters, VisionGRU achieves 2.2% higher accuracy in classification and 0.9% higher mIoU in segmentation while being 184% more computationally efficient.' The only same-parameter row in Table I is VisionGRU-B (86M, 83.1%) versus DeiT-B (86M, 81.8%), which is +1.3%, not +2.2%. The +2.2% (82.0 vs 79.8) is VisionGRU-Ti at 30M versus DeiT-S at 22M, a 36% parameter advantage for VisionGRU. Figure 1 repeats this mismatch, labeling VisionGRU-Ti as 29.64M and DeiT-S as 22.37M, and at the smallest shown resolution (512) VisionGRU-Ti is 25.6 GFLOPs versus DeiT-S's 24.6 GFLOPs, contradicting the caption's 'at all resolutions.' The 184% efficiency figure (432.3/151.9 at 1248) is a single high-resolution point and does not hold at 512. Thus the central quantitative claims in the abstract and introduction are not traceable to the paper's evidence. Segmentation's +0.9 mIoU (VisionGRU-Ti 60M vs DeiT-S+MLN 58M) is closer but still not parameter-matched, and is one run without error bars.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"VisionGRU proposes a hierarchical recurrent backbone for image classification and semantic segmentation. It replaces self-attention with minGRU units arranged in a 2DGRU module that scans the 2D feature map in forward and reverse raster order, with hierarchical downsampling to obtain multi-scale features. Experiments on ImageNet-1K report 82.0% top-1 accuracy for VisionGRU-Ti and 83.1% for VisionGRU-B, and on ADE20K the UperNet+VisionGRU-Ti combination reaches 44.7% mIoU. The paper claims linear computational complexity, reduced memory, and higher efficiency than DeiT models, especially at high resolution.","tokens_in":10674,"tokens_out":4522,"duration_ms":42972,"significance":"If the empirical claims were reproducible, the paper would make a useful contribution: it would show that a linear-complexity recurrent backbone, built from minGRU and parallel prefix-scan training, can match or exceed ViTs on image classification and dense prediction. The design choice of bidirectional raster scans, inherited from SSM-based vision models, is clearly stated, and the paper explicitly identifies the 2DGRU module as the core novelty. However, the current evidence does not support the strongest headline claims: the key comparisons in the abstract and introduction are not parameter-matched, single runs are reported without variance, and the FLOPs figure contradicts the claim of lower cost at all resolutions. With corrected reporting and additional matched-budget experiments, the contribution could become solid, but as presented the central quantitative claims need substantial revision.","major_comments":[{"comment":"The headline claim that 'with the same number of parameters, VisionGRU achieves 2.2% higher accuracy in classification and 0.9% higher mIoU in segmentation while being 184% more computationally efficient' is not supported by the paper's own data. In Table I, the only parameter-matched classification comparison is VisionGRU-B (86M, 83.1%) versus DeiT-B (86M, 81.8%), which is a gain of 1.3 points, not 2.2. The 2.2-point gain (82.0 vs 79.8) compares VisionGRU-Ti at 30M parameters with DeiT-S at 22M parameters, a 36% parameter advantage. Fig. 1 repeats this mismatched pairing, and at the smallest resolution shown (512) VisionGRU-Ti is 25.6 GFLOPs versus DeiT-S's 24.6 GFLOPs, contradicting the caption's 'at all resolutions'. The 184% efficiency figure (432.3/151.9 at 1248) is a single high-resolution point. The abstract and introduction should either be corrected to report only the matched-budget comparisons, or the authors should add properly matched experiments (same parameter count and same resolution) that justify the stated numbers.","section":"Abstract, Section I, Table I, Fig. 1"},{"comment":"No variance or statistical reliability information is reported for any experimental result. The classification differences that drive the paper's narrative are 1.3 and 2.2 accuracy points, and the segmentation gain is 0.9 mIoU; these margins may be within run-to-run variability for ImageNet and ADE20K training. The authors should report results over multiple seeds (e.g., mean and standard deviation, or individual runs) and state whether the hyperparameters were reused from DeiT for all models or tuned for VisionGRU. Without this information, the central empirical comparison cannot be evaluated.","section":"Tables I and II, Section IV"},{"comment":"The claim that the 2DGRU module 'captures both local and global context' and 'improves long-range dependency modeling' is load-bearing for the reported segmentation gains, but it is never validated by ablation. The paper assumes that two flat raster scans (forward and reverse) supply sufficient 2D global context, yet no experiment varies the scan order (e.g., row-major vs column-major), the number of scanning directions, or compares against a 2D-aware alternative. Since the mIoU improvement over DeiT is small (0.9 points) and the architecture's advantage over windowed attention is argued to come exactly from this global scanning, an ablation isolating the scanning strategy is necessary to support the design claim.","section":"Section III.C, Section IV.B"},{"comment":"The efficiency claims are not substantiated. Fig. 1 reports FLOPs for a mismatched pair (VisionGRU-Ti vs DeiT-S) and uses a single point to claim a 184% improvement; it also shows that at 512 resolution VisionGRU-Ti is not cheaper. The paper does not describe how FLOPs were computed, and it reports no wall-clock training or inference time, no GPU memory measurements, and no throughput figures, despite the abstract's claims about 'significantly reducing memory usage and computational costs'. The authors should either provide a clear FLOPs formula for their architecture, report measured runtime and memory, and restrict efficiency statements to operating points where the comparison actually holds.","section":"Section IV.C, Fig. 1"}],"minor_comments":[{"comment":"The introduction contains two nearly identical consecutive sentences describing the bidirectional scanning mechanism and its benefit for semantic segmentation; one should be removed.","section":"Section I"},{"comment":"Reference [26] appears to be misattributed: the text cites Feng et al. for minGRU and 'Were RNNs all we needed?', but the reference entry lists different authors and a different title. Please correct the citation.","section":"References"},{"comment":"References [19] and [20] both list papers titled 'Vision Mamba', which is confusing; please verify the entries and disambiguate them (e.g., with arXiv identifiers).","section":"References"},{"comment":"The formulas in Eqs. (10) and (11) have indexing that is hard to follow: for the forward path, y→,m is defined for m ≤ L_{i,j} but includes a product from k=m+1 to L_{i,j}, and the backward path uses a product from k=L_{i,j} to m−1. Please clarify the index ranges and state whether the sums in Eq. (9) run over all scan positions m.","section":"Section III.C"},{"comment":"The pseudocode uses a variable 'Paths' without defining it; please make explicit that the set is {forward raster, reverse raster} and state how the aggregation is performed.","section":"Algorithm 1, Section III.C"},{"comment":"The text gives block distributions [2,2,15,2] and [2,2,8,2] for Base and Tiny, respectively, which do sum to 21 and 14, but it omits the hidden dimensions and FFN expansion factors for each stage. These details are needed to reproduce the parameter counts in Table I.","section":"Section III.D"},{"comment":"The training setup is said to use 'AdamW with a momentum of 0.9'; AdamW uses betas rather than momentum, so please correct the terminology or provide the exact beta values.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The central quantitative claims in the abstract and introduction currently do not match the tables and figures in the manuscript. This is the most pressing issue and will need to be resolved before the paper can be considered further. The reference list also contains several apparent citation errors that should be corrected during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"VisionGRU puts minGRU into a hierarchical bidirectional scanning backbone for images, and the reported ImageNet and ADE20K results are a credible engineering data point: a linear-complexity RNN can land close to small ViTs on classification and match Swin-Ti on segmentation. That part is real. The architecture is a sensible combination of prior pieces—minGRU from Feng et al., bidirectional scanning from Vim, hierarchical downsampling from Swin—and the paper describes it clearly. There is no new mathematical mechanism; Eqs. (9)-(11) are the minGRU recurrence written over raster scans, but that is fine for an applied paper.\n\nThe soft spot is the packaging, and it is load-bearing. The abstract says VisionGRU “outperforms ViTs” and the introduction says that with the same parameters it gets +2.2% accuracy, +0.9 mIoU, and 184% efficiency. Table I contradicts that: the +2.2% is VisionGRU-Ti (30M) against DeiT-S (22M), a 36% parameter advantage; the same-parameter comparison (VisionGRU-B vs DeiT-B) is +1.3%. Figure 1's caption claims lower FLOPs “at all resolutions,” but at 512×512 VisionGRU-Ti is 25.6 GFLOPs vs DeiT-S's 24.6. The 184% is a single high-resolution point at 1248×1248. The paper's headline claims are not traceable to its own evidence.\n\nOther weaknesses are standard but real. No code, no ablations, no error bars, no seeds. The segmentation gain is one run and the backbone comparison is close but not parameter-matched (60M vs 58M). The bidirectional two-scan design is the core architectural choice, yet no experiment varies scan order or compares with a single scan; the “long-range dependency” narrative rests on that untested assumption. The future-directions paragraph cites a large block of the authors' own papers, which looks like padding but is not a scientific flaw.\n\nWhen read honestly, the result is still decent: an 86M-param RNN at 83.1% ImageNet and a 60M backbone at 44.7 mIoU on ADE20K are competitive with DeiT-B and Swin-Ti. That deserves a serious referee. I would send it to review, but the revision must fix the abstract and introduction to match the tables, add ablations and error bars, and release code. Without those, the claims outrun the evidence.","headline":"A credible RNN-vision backbone whose real numbers are buried under overstated abstract and intro claims; worth reviewing but needs honest comparisons, ablations, and code.","tokens_in":11247,"tokens_out":3513,"would_cite":false,"duration_ms":28344,"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":"VisionGRU claims that a linear-complexity recurrent backbone using minGRU units and bidirectional raster scanning matches or beats vision transformers on image classification and semantic segmentation while cutting compute and memory…","keywords":["VisionGRU","minGRU","recurrent neural network","image classification","semantic segmentation","linear complexity","bidirectional scanning","vision backbone"],"falsifier":"Run VisionGRU on ADE20K with only a unidirectional raster scan or with a different scan order such as column-major or a Hilbert curve; if the mIoU does not drop appreciably when bidirectionality or raster order is removed, then the bidirectional global-context mechanism is not the source of the reported segmentation gains.","tokens_in":10092,"feed_emoji":"⚡","tokens_out":6775,"duration_ms":55178,"temperature":0.7,"pith_summary":"This paper aims to show that a recurrent neural network redesigned for 2D image data can match or beat vision transformers on classification and segmentation while using far less compute and memory. The proposed architecture, VisionGRU, uses a simplified gated recurrent unit (minGRU) to process image features in linear time, with a bidirectional raster-scanning module intended to capture global context. On ImageNet-1K, VisionGRU-Ti reaches 82.0% top-1 accuracy versus 72.2% for DeiT-Ti, and VisionGRU-B reaches 83.1% versus 81.8% for DeiT-B; on ADE20K segmentation it reports 44.7% mIoU versus 43.8% for a DeiT-based baseline. If these results hold, RNN-based backbones become a viable alternative to attention models, especially at high input resolutions where self-attention's quadratic cost becomes prohibitive.","feed_headline":"Linear-complexity RNN outperforms ViTs on ImageNet, ADE20K","feed_subtitle":"VisionGRU reports 2.2% higher accuracy, 0.9% higher mIoU with far fewer FLOPs at high resolution.","key_machinery":"The load-bearing component is the 2DGRU module, built from minGRU units. minGRU is a simplified Gated Recurrent Unit that removes the reset gate and the dependence of the update gate and candidate state on previous hidden states, so the recurrence can be computed in parallel using a prefix-scan algorithm without backpropagation through time. In 2DGRU, a feature map is flattened along a raster scan and processed in both forward and reverse directions by two minGRU passes, then the outputs are aggregated position-wise, followed by a residual connection and a feed-forward network. The claimed effect is that every spatial location integrates information from all preceding and succeeding locations in the scan, which the paper argues provides long-range 2D context. Hierarchical downsampling layers between 2DGRU stages progressively shrink spatial dimensions and expand channels, mimicking the multi-scale structure of CNN and Swin backbones while keeping total compute linear in image size.","core_discovery":"The central discovery claimed is that a hierarchical recurrent backbone built from minGRU units can outperform vision transformers on both image classification and semantic segmentation without sacrificing efficiency. The key reported numbers are 82.0% top-1 accuracy for VisionGRU-Ti (30M params) against 72.2% for DeiT-Ti (6M) and 83.1% for VisionGRU-B (86M) against 81.8% for DeiT-B (86M), plus 44.7% mIoU on ADE20K with a UperNet head. The paper attributes these gains to the 2DGRU module, which scans feature maps in forward and reverse raster order with bidirectional minGRU layers, letting each spatial position accumulate context from the whole image. Combined with downsampling stages that reduce sequence length while widening channels, the model retains multi-scale features with computational complexity that grows linearly with image area, a direct contrast with self-attention's quadratic scaling. The paper frames this as evidence that RNNs, when simplified for parallel training and adapted to 2D structure, can close the gap with attention-based architectures that had displaced them.","pith_inferences":["The paper does not test whether the bidirectional raster scan is essential; a controlled comparison with unidirectional scanning, column-major scanning, or a serpentine order would isolate whether the claimed global-context benefit comes from bidirectionality or simply from scan depth.","The reported FLOPs advantage might not translate linearly to wall-clock speed on modern hardware, because repeated sequential dependencies and memory-access patterns in raster scans could introduce overhead; benchmarking throughput and latency at multiple resolutions would clarify the practical gain.","If the scan-order assumption holds, extending VisionGRU to video by scanning the spatial-temporal volume could give a linear-complexity alternative to 3D attention, though the paper does not address this.","The model's parameter efficiency relative to DeiT-Ti is not apples-to-apples (30M vs 6M), so a fairer comparison controlling for parameter count at the same budget would strengthen the claim that RNNs close the gap with attention."],"forward_implications":["If VisionGRU's results replicate, recurrent backbones become a practical choice for high-resolution vision tasks where self-attention's quadratic cost is a bottleneck.","The 2DGRU's bidirectional scanning could be transferred to other dense prediction tasks, such as detection or depth estimation, where global context matters.","The minGRU-based design implies training can be parallelized without BPTT, so scaling to larger images does not automatically increase memory as steeply as attention models.","The reported efficiency gains (e.g., 151.9 GFLOPs at 1248x1248 vs 432.3 for DeiT-S) suggest mobile or real-time deployments could benefit from replacing transformer backbones with this RNN.","The architecture's success would validate the broader direction of recurrence- and state-space-based vision models that aim to bridge the performance gap with attention."],"supporting_citations":[{"why":"Supplies the minGRU unit, the simplified GRU that removes reset gates and hidden-state dependencies for gates, enabling parallel prefix-scan training without BPTT.","marker":"[11]"},{"why":"Establishes the Vision Transformer baseline that treats images as patch sequences, setting the global-context standard VisionGRU claims to match or beat.","marker":"[5]"},{"why":"Provides the DeiT training recipe and the DeiT-Ti/S/B baselines that VisionGRU compares against for accuracy and efficiency.","marker":"[28]"},{"why":"Inspires the hierarchical downsampling design with patch merging that VisionGRU adopts to reduce sequence length while increasing channels.","marker":"[15]"},{"why":"Introduces bidirectional scanning in a state-space vision model, the strategy that 2DGRU adapts for its global context modeling.","marker":"[19]"},{"why":"Provides the UperNet segmentation head used to evaluate VisionGRU's dense prediction performance on ADE20K.","marker":"[33]"},{"why":"Supplies the ImageNet-1K dataset, the primary benchmark for the reported classification accuracy comparisons.","marker":"[27]"}],"fun_headline_variants":["Linear-complexity RNN beats ViTs on ImageNet and ADE20K","VisionGRU: RNN model outruns ViTs with linear scaling","GRU-based vision net tops ViT accuracy, cuts memory and FLOPs","Hierarchical RNN achieves 82% top-1, surpassing DeiT on ImageNet","RNN with linear cost outperforms transformer vision models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that flattening a 2D feature map into two flat raster scans (forward and reverse) gives each pixel enough global 2D context for the claimed accuracy gains; no experiment varies the scan order or compares against a spatially-aware alternative.","fun_headline_variants_meta":{"raw":{"variants":["Linear-complexity RNN beats ViTs on ImageNet and ADE20K","VisionGRU: RNN model outruns ViTs with linear scaling","GRU-based vision net tops ViT accuracy, cuts memory and FLOPs","Hierarchical RNN achieves 82% top-1, surpassing DeiT on ImageNet","RNN with linear cost outperforms transformer vision models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001325,"raw_usage":{"total_tokens":5440,"prompt_tokens":1042,"completion_tokens":4398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":4296}},"tokens_in":658,"tokens_out":4398,"duration_ms":28498,"temperature":1.0,"reasoning_tokens":4296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:57:44.143548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VisionGRU on ADE20K with only a unidirectional raster scan or with a different scan order such as column-major or a Hilbert curve; if the mIoU does not drop appreciably when bidirectionality or raster order is removed, then the bidirectional global-context mechanism is not the source of the reported segmentation gains.","supporting_citations":[{"cited_title":"Were rnns all we needed?,","cited_arxiv_id":null,"evidence_quote":"Supplies the minGRU unit, the simplified GRU that removes reset gates and hidden-state dependencies for gates, enabling parallel prefix-scan training without BPTT."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Inspires the hierarchical downsampling design with patch merging that VisionGRU adopts to reduce sequence length while increasing channels."},{"cited_title":"Vision mamba: Efficient visual representation learning with bidirectional state space model,","cited_arxiv_id":null,"evidence_quote":"Introduces bidirectional scanning in a state-space vision model, the strategy that 2DGRU adapts for its global context modeling."},{"cited_title":"Upernet: Unified perceptual parsing for scene understanding,","cited_arxiv_id":null,"evidence_quote":"Provides the UperNet segmentation head used to evaluate VisionGRU's dense prediction performance on ADE20K."},{"cited_title":"Imagenet: A large-scale hierarchical image database,","cited_arxiv_id":null,"evidence_quote":"Supplies the ImageNet-1K dataset, the primary benchmark for the reported classification accuracy comparisons."}],"review_version":1}