Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Putting Registers to Work: Task Registers for Token Pruning in Vision Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Token-pruning choices differ across vision tasks, and a single evolving task register can adapt them, keeping dense predictions close to the full model at 1.3x encoder speed.

desk verdict A genuinely new task-register pruning mechanism with strong dense-task results, but the cross-task probe story is exaggerated and the code is not actually released. read the letter →

arxiv 2608.10989 v1 pith:JFPYANVF submitted 2026-08-11 cs.CV cs.AI

classification cs.CVcs.AI
keywords tokenpruningvisiontransformertaskregistersemanticsegmentationobjectdetectionimageclassificationadaptivecomputationefficientinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Token pruning in Vision Transformers is usually designed as a single policy, but pretrained backbones are reused across tasks with very different spatial demands. This paper asks which pruning decisions transfer across image classification, semantic segmentation, and object detection, and answers with controlled probes: the same parameter-free criteria rank differently per task, classification is especially hurt by attention-based pruning in the earliest layers, and the two dense tasks prefer opposite ways to reconstruct removed positions. On the strength of those findings it proposes Task-Adaptive Pruning (TAP), which gives each task its own learned register token whose evolving state selects tokens, distributes an exact removal budget over depth, and sets a recovery scale for dense features. At a keep rate of half the tokens, the jointly fine-tuned TAP-J drops 0.2 mIoU on ADE20K and 0.3 box AP on COCO while accelerating the encoder about 1.3x, with ImageNet-1K accuracy competitive.

What carries the argument

The central object is the task register: one learned $d$-dimensional vector per task that is the only task-specific token allowed into the shared backbone. It evolves through ordinary transformer blocks, and at each pruning layer its normalized state is turned into a selection score $f_i^{(l)}$ by dotting the register query against each patch key using the block's existing projections, a per-layer removal count through a sigmoid readout $w^\top r_t^{(l)} + b$ that provably spends the exact global budget, and a recovery scale $\alpha_t^{(l)} = \sigma(u^\top r_t^{(l)} + v)$ used only at dense readouts. Removed tokens are matched to the retained token with the highest key-space cosine similarity, and their stored feature offset $\delta_i$ is added back scaled by $\alpha_t^{(l)}$, keeping reconstructed features out of later encoder blocks. The register is what lets one sparse backbone carry task-dependent computation without per-task selectors.

What would settle it

Run the same controlled probe on segmentation and detection at a matched input resolution with the same eligible pruning layers; if the rank correlation between criterion drops is no longer negative, the paper's motivating task-conflict claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that token reduction cannot be treated as a single transferable rule across vision tasks. Controlled probes freeze each no-pruning pipeline and change one pruning choice at one eligible layer at a time; they show that segmentation and detection rank eleven criteria almost oppositely (Spearman rank correlation -0.62), that attention-based selection causes larger drops than spatial coverage in the earliest classification layers, and that segmentation benefits from restoring the full stored feature offset at removed positions while detection prefers the surviving stand-in alone. TAP turns these observations into a mechanism: one learned task register per task is activated per forward pass, and its state drives three decisions: register-to-patch scoring for token selection (reusing the block's query and key projections, so no extra projection parameters), an exact integer allocation of the global removal budget across pruning layers, and a per-layer recovery scale $\alpha_t^{(l)}$ that interpolates between the stand-in endpoint and the full offset. The reported consequence is that at $\rho=0.5$ TAP-J reaches 47.0 mIoU at $1.30\times$ encoder throughput on ADE20K and 53.7 box AP at $1.32\times$ on COCO while remaining competitive on ImageNet-1K.

Load-bearing premise

The load-bearing premise is that the differences in criterion ranking and recovery preference come from the tasks themselves, not from the fact that the segmentation and detection pipelines differ in input resolution, head, and eligible pruning layers.

Editorial extensions

If this is right

  • A single fixed pruning policy leaves measurable accuracy on the table on dense tasks; the same backbone can support task-specific pruning paths with only a few thousand extra parameters.
  • At half the tokens, dense predictions stay within 0.2 to 0.3 points of the unpruned model while the encoder runs about 1.3 times faster.
  • One mechanism can coordinate which tokens survive, where the budget is spent, and how removed positions are rebuilt, and it can be trained exactly to a target token count without auxiliary rate losses.
  • A frozen shared backbone with per-task registers and low-rank updates is a workable multi-task operating point, at the cost of a larger accuracy drop than per-task fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The probe comparison bundles task semantics with pipeline differences such as input resolution, head architecture, and eligible layers, so a matched-pipeline study would be needed to attribute the ranking conflict purely to the task.
  • If the opposite recovery endpoints generalize, other dense tasks such as depth estimation, video segmentation, or panoptic segmentation should learn their own reconstruction rule rather than reusing one endpoint.
  • Because the task register is a tiny interface and scoring reuses existing projections, the same recipe could be probed for adding a new task to a frozen backbone without retraining the pruning controller.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper studies whether token-pruning policies transfer across image classification, semantic segmentation, and object detection with a ViT backbone. It first runs controlled probes on frozen no-pruning checkpoints, applying parameter-free reduction criteria one layer at a time, and reports three task-level differences: dense tasks rank criteria differently, classification is especially sensitive to attention-based pruning in early layers, and the dense pipelines prefer opposite recovery endpoints. These findings motivate Task-Adaptive Pruning (TAP), in which a learned per-task register token propagates through the backbone and conditions token scoring, layerwise removal allocation, and dense-feature recovery. Two regimes are evaluated: TAP-J (per-task full fine-tuning) and TAP-F (frozen shared backbone with LoRA and task registers). At a final keep rate of 0.5, TAP-J reaches 47.0 mIoU at 1.30x encoder throughput on ADE20K and 53.7 box AP at 1.32x encoder throughput on COCO, with 83.2 top-1 accuracy at 1.26x on ImageNet-1K; TAP-F reports larger drops. The supplement contains an exact-budget proof, unit tests, verification tables, and extensive ablations.

Significance. If the results hold, the paper provides a practical and unusually lightweight token-pruning controller. The exact-budget guarantee is proved in Proposition 1 and supported by unit tests, the main operating points are verified in Table S11, and the empirical package is strong: multiple seeds with standard deviations, calibration-split baselines, provenance tables for baselines, and a claim-to-evidence map in Table S23. The dense-task results are competitive with fully fine-tuned methods at about 1.3x encoder speedups. However, the central interpretive claim that pruning behavior differs by task semantics is not established by the evidence as presented, because the probes compare complete pipelines with different resolutions, heads, and eligible-layer schedules, and the headline TAP-J result does not demonstrate cross-task sharing because it trains a separate backbone per task. The paper's value lies more in the controlled comparison and the TAP controller than in the causal task-semantics narrative.

major comments (3)
  1. [Abstract; Figure 1; Controlled study; Supplement J] The probe evidence in Figure 1 and Tables S9 and S10 is presented as revealing differences between tasks ('segmentation and detection rank the criteria differently', 'the dense tasks prefer opposite recovery endpoints'), but the probes simultaneously vary input resolution (224 vs 512 vs 1024), task head (linear, SegFormer MLP, Cascade Mask R-CNN), and eligible pruning layers (blocks 3/6/9 for classification and segmentation; only global-attention blocks 3/6/9/12 for detection). The Spearman correlation of -0.62 and the recovery-endpoint reversal in Table S10 therefore measure complete-pipeline differences, not task semantics. The main controlled-study paragraph does contain the caveat 'compare complete pipelines rather than task semantics alone,' but the abstract, the contribution list, and Figure 1 do not carry that caveat and instead use causal task language. Because these probe findings are the paper's stated motivation and first contribution, the authors should either add a matched-pipeline control (for example, fixing resolution and pruning-layer schedule across tasks while varying the head/objective) or consistently reframe all probe claims as pipeline-level transfer results rather than task-level ones.
  2. [Experiments, Main comparison; Tables 1-3; Figure 4] The sentence 'The largest gains occur on the dense tasks, linking the conflicts identified in Figure 1(b-d) to TAP's task-conditioned decisions' is not supported by the comparisons in Tables 1-3. TAP-J is fully fine-tuned per task, whereas the static rule is parameter-free and is not fine-tuned, so the comparison conflates task-conditioning with the benefit of fine-tuning. The ablations in Table 4 are conducted in the frozen-base TAP-F regime and show relative gains, but TAP-F itself loses 1.0 top-1, 1.4 mIoU, and 1.5 box AP against no pruning and is worse than the static rule on classification and detection. To attribute the observed gains to the task-conditioned decisions, the paper needs fully fine-tuned controls, such as a fine-tuned attention top-k policy, a fine-tuned static layer schedule, and a TAP-J variant whose register readouts are replaced by fixed or task-mean policies.
  3. [Introduction; Method; Tables 1-3 and S12] The paper's headline model TAP-J is not a shared-backbone model, since it trains a separate backbone, register, and readouts for each task. The only shared-backbone model is TAP-F, whose published operating points are substantially weaker (Table S12: TAP-F loses 1.0 top-1, 1.4 mIoU, and 1.5 box AP at rho=0.5). This weakens the framing in the abstract and introduction that pretrained ViTs are 'reused across tasks' and that TAP is a mechanism for unified vision systems. The authors should either present TAP-J explicitly as a per-task upper bound and TAP-F as the cross-task result, or add a cross-task training variant that yields competitive numbers while using a genuinely shared backbone.
minor comments (5)
  1. [Figure 1 caption and panel (b)] The labels 'seg needs coverage' and 'det needs attention' in Figure 1(b) are causal and task-level; add a qualifier such as 'under these task pipelines' so the caption does not contradict the caveat later in the Controlled-study paragraph.
  2. [Supplement, Table S20] Please clarify the batch size used for the per-image component latencies in Table S20; as written, the segmentation end-to-end value of 5.58 ms per image appears inconsistent with the 280 im/s encoder throughput reported in Table 1, and the table should state whether it uses the throughput protocol's batch size or a single-image latency protocol.
  3. [Supplement, Table S9] The Spearman correlation of -0.62 is computed over only 11 criteria whose scores share the same underlying data; report a permutation-based confidence interval or p-value to indicate the stability of the rank reversal.
  4. [Method, Eq. (3) and Supplement A.5] The supplement correctly notes that adding an early-layer offset to a later endpoint 'is an approximation, since the intervening blocks do not provide an explicit transport map'; the main text should at least point to this limitation where Eq. (3) is introduced, since it is a core assumption of the recovery mechanism.
  5. [Abstract and Method] The name TAP-J ('jointly adapted') is confusing because the model trains a separate backbone per task; state in the abstract or method that 'joint' refers to jointly fine-tuning the backbone, register, and readouts within each task, not to joint training across tasks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's empirical probes, learned policies, and ablations are independent of each other and of the claimed conclusions.

full rationale

This is an empirical method paper rather than a derivation, and no load-bearing step reduces by construction to its own inputs. The controlled probes freeze a no-pruning checkpoint and evaluate parameter-free criteria at one eligible layer without retraining, so their rankings are independent of the later learned TAP policy. The learned recovery scales and allocations are trained with task losses and only subsequently compared with the probe endpoints, for example in Figure 5 versus Table S10; the probe values are not used to fit or select alpha, and the ablations in Table 4 vary one component at a time. The exact-budget allocation is proven directly with an invariant in Proposition 1, not imported from prior work. The only substantive weakness is an interpretation caveat: the probes compare complete pipelines that differ in resolution, head, and eligible layers, and the paper itself states this in the Controlled-study paragraph ('the probes compare complete pipelines rather than task semantics alone') and in Supplement J. That confound concerns attribution of the findings to task semantics, not circularity: it does not make any result equivalent to its inputs. There is no self-citation chain, imported uniqueness theorem, or renamed known result. A score of 0 is therefore appropriate.

Assumptions & free parameters 7 free parameters · 7 assumptions · 1 invented entities

This is an empirical machine-learning method, so the ledger records trained parameters, standard deep-learning assumptions, and one new architectural entity. No parameter-free prediction is made; performance claims are the outcome of training. The main load-bearing assumptions are the transferability of frozen-probe rankings to trained policies and the attribution of probe differences to task identity, the latter being only partly controlled.

free parameters (7)
  • Task register initial vector r_t (per task) = Learned, 768-dimensional per task; exact trained values not reported
    Trained controller state that conditions token selection, budget allocation, and recovery scale; all central performance numbers depend on it.
  • Allocation readout (w, b) = Learned, shared, dimension 768+1
    Maps the active register state to the fraction of the remaining removal budget proposed at each pruning layer.
  • Recovery readout (u, v) = Learned, shared, dimension 768+1
    Maps the active register state to the recovery scale alpha used for dense-feature reconstruction.
  • LoRA low-rank updates (TAP-F only) = Rank 8, 1,179,648 parameters total
    Task-specific feature adaptation in the frozen-base regime; the TAP-F results depend on these.
  • Keep rate rho = 0.5 main operating point; 0.3 and 0.7 also evaluated
    Chosen by hand as the balanced accuracy-throughput operating point, not learned.
  • Pruning layer set L = {3, 6, 9} for all tasks
    Chosen by hand; the method does not learn where to prune.
  • Gumbel temperature schedule = Cosine from 1.0 to 0.1
    Annealing schedule for the straight-through estimator, chosen by hand.
assumptions (7)
  • domain assumption MAE-pretrained ViT-B features transfer to classification, segmentation, and detection.
    All experiments start from the same MAE-pretrained ViT-B family (He et al. 2022); no evidence is given for other backbones or pretraining schemes.
  • domain assumption Probing a frozen no-pruning checkpoint with parameter-free criteria predicts how a trained pruning policy will behave.
    The controlled probes remove tokens without retraining and the paper uses their rankings to motivate TAP; the transfer of frozen-state rankings to end-to-end trained policies is assumed (Section C).
  • domain assumption Observed differences across the three pipelines are attributable to task identity rather than pipeline confounds.
    The pipelines differ in resolution, head, and eligible pruning layers; Supplementary Section J admits the study measures complete pipelines, not task semantics in isolation.
  • ad hoc to paper Early feature offsets can be added to later surviving endpoints at dense readout points without an explicit transport map.
    Recovery Equation (3)/S11 assumes residual-stream coordinates remain comparable across depth; the paper calls this an approximation.
  • domain assumption Key-space cosine similarity is a useful rule for matching removed tokens to stand-ins.
    Validated by ablations (Table S15) but chosen by design, without theoretical justification.
  • domain assumption Straight-through estimation with a cardinality-constrained soft mask gives useful gradients for hard top-k selection.
    Standard practice cited to Bengio et al. 2013 and Jang et al. 2017; the paper validates it empirically with estimator controls (Table S17).
  • domain assumption The global keep rate is fixed and per-image compute budgets are not learned.
    Acknowledged limitation in Section J; the register only distributes a fixed total removal budget over depth.
invented entities (1)
  • Task register (one learned initial vector per task, propagated through the backbone)
    purpose: Acts as an active controller that determines token selection, depth-wise budget allocation, and dense recovery scaling for the current task.
    The task register is a new architectural component. Ablations (shared register loses 1.3 mIoU and 1.1 box AP) give within-paper evidence, but no external falsifiable handle is provided beyond the reported experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Putting Registers to Work: Task Registers for Token Pruning in Vision Transformers." pith.science (2026). https://pith.science/paper/JFPYANVF

@misc{pith2026260810989,
  author       = {Pith},
  title        = {Pith review of: Putting Registers to Work: Task Registers for Token Pruning in Vision Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFPYANVF}},
  note         = {Machine review of arXiv:2608.10989}
}
abstract

Token-pruning policies are usually designed for a single recognition pipeline, but pretrained Vision Transformers are reused across tasks with different spatial demands. We ask which parts of a pruning policy transfer across image classification, semantic segmentation, and object detection. For each pipeline, controlled probes freeze the no-pruning checkpoint and apply a series of parameter-free reduction criteria at one eligible layer at a time without retraining. The probes reveal three differences: segmentation and detection rank the criteria differently, classification is especially sensitive to attention-based pruning in the earliest layers, and the dense tasks prefer opposite recovery endpoints. These findings motivate Task-Adaptive Pruning (TAP). Existing register tokens serve as task-agnostic storage for feature artifacts. TAP instead introduces one task register per task and activates only the current one. Its evolving state ranks tokens, distributes an exact removal budget over depth, and sets the recovery scale for dense features. At a final keep rate of $\rho=0.5$, our jointly adapted model, TAP-J, reaches $47.0$ mIoU at $1.30\times$ encoder throughput on ADE20K and $53.7$ box AP at $1.32\times$ encoder throughput on COCO while remaining competitive on ImageNet-1K.

Figures

Figures reproduced from arXiv: 2608.10989 by the authors.

Figure 1
Figure 1. Controlled probes apply pruning to one layer of the corresponding fixed no-pruning checkpoint and vary one design [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of TAP. Each task has its own learned initial register, and only the register associated with the current [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Tokens that reach the last block for one image at [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Accuracy drop against encoder speedup with the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Register-conditioned validation summaries. (a) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Attention Is All You Need , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  2. [2]

    International Conference on Learning Representations (ICLR) , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations (ICLR) , year=

  3. [3]

    International Conference on Learning Representations (ICLR) , year=

    Not All Patches Are What You Need: Expediting Vision Transformers via Token Reorganizations , author=. International Conference on Learning Representations (ICLR) , year=

  4. [4]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  5. [5]

    AAAI Conference on Artificial Intelligence (AAAI) , year=

    Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer , author=. AAAI Conference on Artificial Intelligence (AAAI) , year=

  6. [6]

    European Conference on Computer Vision (ECCV) , year=

    Adaptive Token Sampling for Efficient Vision Transformers , author=. European Conference on Computer Vision (ECCV) , year=

  7. [7]

    International Conference on Learning Representations (ICLR) , year=

    Token Merging: Your ViT but Faster , author=. International Conference on Learning Representations (ICLR) , year=

  8. [8]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    DiffRate: Differentiable Compression Rate for Efficient Vision Transformers , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

Show all 41 references
  1. [9]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Beyond Attentive Tokens: Incorporating Token Importance and Diversity for Efficient Vision Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  2. [10]

    arXiv preprint arXiv:2506.10967 , year=

    Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs , author=. arXiv preprint arXiv:2506.10967 , year=

  3. [11]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Token Cropr: Faster ViTs for Quite a Few Tasks , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  4. [12]

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

    VLTP: Vision-Language Guided Token Pruning for Task-Oriented Segmentation , author=. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

  5. [13]

    International Conference on Learning Representations (ICLR) , year=

    Vision Transformers Need Registers , author=. International Conference on Learning Representations (ICLR) , year=

  6. [14]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Vision Transformers Need More Than Registers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  7. [15]

    IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , year=

    Which Tokens to Use? Investigating Token Reduction in Vision Transformers , author=. IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , year=

  8. [16]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Dynamic Token Pruning in Plain Vision Transformers for Semantic Segmentation , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  9. [17]

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

    Revisiting Token Pruning for Object Detection and Instance Segmentation , author=. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

  10. [18]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    A-ViT: Adaptive Tokens for Efficient Vision Transformer , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  11. [19]

    European Conference on Computer Vision (ECCV) , year=

    SPViT: Enabling Faster Vision Transformers via Latency-Aware Soft Token Pruning , author=. European Conference on Computer Vision (ECCV) , year=

  12. [20]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  13. [21]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  14. [22]

    International Conference on Machine Learning (ICML) , year=

    Training Data-Efficient Image Transformers and Distillation through Attention , author=. International Conference on Machine Learning (ICML) , year=

  15. [23]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Masked Autoencoders Are Scalable Vision Learners , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  16. [24]

    International Conference on Learning Representations (ICLR) , year=

    Categorical Reparameterization with Gumbel-Softmax , author=. International Conference on Learning Representations (ICLR) , year=

  17. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  18. [26]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    ImageNet: A Large-Scale Hierarchical Image Database , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  19. [27]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Scene Parsing Through ADE20K Dataset , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  20. [28]

    European Conference on Computer Vision (ECCV) , year=

    Microsoft COCO: Common Objects in Context , author=. European Conference on Computer Vision (ECCV) , year=

  21. [29]

    European Conference on Computer Vision (ECCV) , year=

    Unified Perceptual Parsing for Scene Understanding , author=. European Conference on Computer Vision (ECCV) , year=

  22. [30]

    European Conference on Computer Vision (ECCV) , year=

    Exploring Plain Vision Transformer Backbones for Object Detection , author=. European Conference on Computer Vision (ECCV) , year=

  23. [31]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Cascade R-CNN: Delving into High Quality Object Detection , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  24. [32]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  25. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  26. [34]

    International Conference on Learning Representations (ICLR) , year=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. International Conference on Learning Representations (ICLR) , year=

  27. [35]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  28. [36]

    International Conference on Learning Representations (ICLR) , year=

    Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity , author=. International Conference on Learning Representations (ICLR) , year=

  29. [37]

    arXiv preprint arXiv:1308.3432 , year=

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation , author=. arXiv preprint arXiv:1308.3432 , year=

  30. [38]

    arXiv preprint arXiv:2504.13181 , year=

    Perception Encoder: The best visual embeddings are not at the output of the network , author=. arXiv preprint arXiv:2504.13181 , year=

  31. [39]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  32. [40]

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

    Token Pooling in Vision Transformers for Image Classification , author=. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

  33. [41]

    IEEE International Conference on Multimedia Information Processing and Retrieval (MIPR) , year=

    SegFormer++: Efficient Token-Merging Strategies for High-Resolution Semantic Segmentation , author=. IEEE International Conference on Multimedia Information Processing and Retrieval (MIPR) , year=

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.