GTR: Gated Token Recurrence for Efficient Dense Prediction
Zhe Feng, Longfei Liu, Wei Liu, Kai Chen, Jiangang Kong, Wei Zhou, Yifeng Qian, Dexiong Chen, et al.
cs.CV · cs.LG
Quick verdict
Linear-attention backbone GTR claims softmax-level accuracy on six dense tasks; results are extensive and well-ablated, but single-run with narrow baselines.
Summary
GTR introduces a softmax-free vision backbone for dense prediction built on gated linear attention (GLA): twelve single-scale stride-16 blocks, each performing one directional scan that cycles through four directions across depth, combined with a Spatial SwiGLU block that embeds a 3x3 depthwise convolution in the value branch. The backbone is trained by distillation from a frozen detection-specialized DINOv3 teacher using only a linear projection and a squared L2 loss on final patch tokens — no masking or intermediate-layer supervision. With Objects365 pre-training, GTR-S/M/L/X reach 53.6/57.3/58.9/59.4 COCO AP, and GTR-L runs at 1.908 ms median batch-one FP16 latency on an RTX 4090. The same design transfers to instance segmentation (45.0/49.8 mask AP at S/X), pose estimation (70.1/75.5 keypoint AP), DOTA oriented detection (80.1/81.3 AP50), Cityscapes segmentation (81.5/83.6 mIoU), and ground-truth-aligned NYU depth (delta1 0.946/0.954). A custom chunkwise CUDA kernel is reported 4.0x faster than FLA v0.5.0 at 1.6K tokens, and TensorRT deployment reaches 2.282-8.769 ms on DRIVE AGX Thor. Limitations: single-run results without variance estimates, comparisons largely restricted to YOLO26/EdgeCrafter-family baselines, a non-standard test-time-augmented depth protocol, and no ImageNet classification results for the backbone itself.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A homogeneous GLA backbone with four-directional scanning and Spatial SwiGLU achieves the best AP in all four scale groups on COCO (53.6/57.3/58.9/59.4 with Objects365 pre-training), with GTR-L fastest among L-scale baselines (1.908 ms).
- Single squared-L2 distillation of final patch tokens from a frozen DINOv3 teacher outperforms adapted multi-loss/multi-stage methods (50.7 vs 50.0/48.2 AP, Table 3).
- Linear recurrence translates into practical efficiency: 2.06x latency growth for 4x tokens (GTR-S, Table 6) and a custom chunkwise kernel 4.0x faster than FLA v0.5.0 at 1.6K tokens.
- Replacing GLA with softmax attention in three of twelve blocks yields no meaningful accuracy change (within 0.1 AP, Table 5), supporting the pure-GLA design.
- GTR deploys on automotive edge hardware at 2.282-8.769 ms median batch-one latency on DRIVE AGX Thor with cosine similarity >= 0.9989 to the FP32 reference (Tables 25-26).
Strengths
Breadth with one backbone design
The same 12-block GLA design transfers across six tasks and four scales (Tables 1-2, 7-12): 53.6/57.3/58.9/59.4 COCO AP with Objects365, 45.0/49.8 mask AP and 70.1/75.5 keypoint AP at S/X, 80.1/81.3 DOTA AP50, 81.5/83.6 Cityscapes mIoU, 0.946/0.954 aligned NYU delta1 — each under a task-specific head with per-task recipes disclosed in Appendix D.
The distillation control is genuinely informative
Table 3 holds teacher, student, and COCO schedule fixed: final-output L2 gets 50.7 AP vs adapted ViT-Linearizer 50.0 (mask ratio swept 0.2-0.9) and adapted ViT-AdaLA 48.2, vs 31.1 from scratch — real evidence that single-loss final-token alignment suffices in this setting.
Hybrid-attention null result cleanly justifies the design
Resolution scaling substantiates the linear-complexity claim
Deployment rigor exceeds typical backbone papers
Weaknesses
No statistical uncertainty for accuracy claims
Depth protocol is heavily assisted and narrow
Baseline provenance is mixed and latency asymmetries remain
Efficiency wins are scale- and task-dependent, not universal
No classification benchmark for the backbone itself
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GTR (arXiv 2609.26590v2) is a softmax-free vision backbone for dense prediction. It stacks twelve single-scale stride-16 blocks of key-only gated linear attention (GLA; Yang et al., 2024b), where each block performs a single directional scan and the direction cycles left-to-right, right-to-left, top-to-bottom, bottom-to-top across depth (Sec. 2.1). Local spatial structure is injected by Spatial SwiGLU (Eq. 1), a 3x3 depthwise convolution in the value branch that replaces positional embeddings. The backbone is trained by final-output distillation: a frozen, detection-adapted DINOv3 teacher and the student see the same full image, and the entire objective is a squared L2 loss between a linearly projected student output and the teacher's final patch tokens (Eq. 2). Claims: (i) best-in-group COCO accuracy at four scales (53.6/57.3/58.9/59.4 AP with Objects365); (ii) linear-complexity efficiency that survives implementation (2.06x latency growth for 4x tokens); (iii) transfer to five additional tasks; (iv) production edge deployment (2.282-8.769 ms on DRIVE AGX Thor, Tables 25-26).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GTR in one sentence?
GTR is a linear-complexity vision backbone that replaces softmax attention with gated token recurrence (gated linear attention) plus a local depthwise-convolution block (Spatial SwiGLU), trained by simple final-layer distillation from DINOv3 and aimed at real-time dense prediction such as detection, segmentation, pose, and depth.
How accurate is GTR on COCO object detection?
With Objects365 pre-training, GTR-S/M/L/X reach 53.6/57.3/58.9/59.4 box AP on COCO val2017 at 640x640 input; GTR-L does so at 1.908 ms median latency on an RTX 4090 (FP16, batch size one, compiled execution).
Is GTR code available?
The paper links a project page and a GitHub repository (intellindust-ai-lab.github.io/projects/GTR), but the extracted text contains no explicit release statement, so availability of code, weights, and the custom CUDA operator should be checked on the project page.
How much faster is the custom GLA kernel than existing libraries?
In an isolated benchmark on an RTX 4090 (FP16 inputs, CUDA Graph execution), the authors' inference-specialized chunkwise CUDA operator is reported 2.6x, 4.0x, and 6.4x faster than the general-purpose FLA v0.5.0 library at sequence lengths of 256, 1.6K, and 16.4K tokens respectively.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated parts of Table 1, some references, and sections of Appendices D.1.4/E.3; GTR-S/M rows missing from main Table 1 were cross-checked against Appendix Table 7.
- Figures 1-19 are not analyzable from text; all qualitative visualization claims rest on the authors' descriptions.
- The supplied metadata author list omits Xuanlong Yu and corresponding author Xi Shen, who appear in the paper text.
- GitHub/project-page links appear as labels only; code availability could not be verified from the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.