Pith. sign in

REVIEW 4 major objections 5 minor 50 references

YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that adapter placement for PEFT on real-time detectors should be solved by constraint planning on the detector graph, and that the resulting plans match or exceed full fine-tuning on YOLO11s and YOLO12s while returning…

desk verdict A real engineering contribution with an honest limitation: the Refuse mechanism is a calibrated description of seven collapses, not yet a prediction for unseen architectures; the placement results are plausible and worth referee time. read the letter →

arxiv 2608.07051 v1 pith:35VWECDK submitted 2026-08-07 cs.CV

Xu Lin , WenJie Nie , Jinlong Peng , Weifu Fu , YueXiao Ma , Xiawu Zheng , Yong Liu This is my paper · ORCID
classification cs.CV MSC 68T4568T07
keywords parameter-efficientfine-tuningYOLOdetectionadapterplacementconstraintplanninglow-rankadaptationVOCRefusedecisionobject
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

Parameter-efficient fine-tuning (PEFT) methods transferred from language models select adapter targets by module name or type and can silently degrade real-time object detectors, whose graphs mix grouped and depthwise convolutions, loss-coupled DFL projections, attention blocks, text fusion, and MoE routing. The paper proposes YOLO-PEFT, which treats adapter placement as a constraint-planning problem: it parses the detector graph into operator and semantic roles, evaluates explicit operator-validity, detection-semantic, graph-interface, deployment, and budget predicates with reason codes, and either emits a budgeted target-module plan or returns Refuse before training. Under the official VOC07+12 trainval-to-VOC07 test protocol, planner-selected RS-LoRA reaches 0.7138 and 0.7307 mAP50-95 on YOLO11s and YOLO12s, versus 0.6428 and 0.6662 for Full-SFT; on RT-DETR-L, all seven evaluated LoRA-family configurations cross the catastrophic threshold $\Delta = -0.05$, so the planner refuses and falls back to Full-SFT. A controlled YOLO11 audit shows LoRA reduces peak training VRAM by 43.9 percent but takes 1.72 times longer to train.

What carries the argument

The load-bearing object is the role-aware typed graph $G=(V,E)$ of the detector: each module $i$ carries operator metadata $o_i$, semantic role $s_i$, graph-interface metadata $q_i$, and deployment metadata $h_i$. The planner evaluates predicates such as dense conv supported by the backend, grouped conv requires $G \mid r$, depthwise and normalization excluded, DFL bins and MoE routers frozen, and shape, arity, and residual-preserving substitution, then solves budget-aware rank assignment $\max_{\pi} \sum_i u(i,\pi(i))$ subject to $\sum_i c_p(i,\pi(i)) \le B$. Reliability is estimated by a linear calibration $\Delta \mathrm{mAP} \approx \beta_0 + \beta_1 \varphi_{\mathrm{attn}} + \beta_2 \varphi_{\mathrm{text}} + \beta_3 \varphi_{\mathrm{dw}} + \beta_4 \xi_p$, where the fingerprint includes attention, text-fusion, MoE, depthwise, and dense, grouped, or convolution fractions; a prediction below $\Delta = -0.05$ triggers Refuse. The runtime contract preserves train-save-merge-export invariants, with merge equivalence for the fallback grouped-convolution LoRA backend proven as $\mathrm{conv}(W_0,x) + s\, g((A_g,B_g)_g,x) = \mathrm{conv}(W_0 + s\Delta W, x)$ up to floating-point tolerance.

What would settle it

Run a full VOC fine-tuning on a held-out detector family (for example a mixture-of-transformers detector) with planner-selected RS-LoRA: if an accepted plan produces $\Delta\mathrm{mAP}$ below $-0.05$, or a refused configuration would have matched or exceeded Full-SFT, then the calibration's predictive claim fails. The paper's own absence of a held-out family makes this test the decisive one.

Watch

Extended reading notes

Core claim

The central claim, stated for a fair reader, is that the reason PEFT often fails on real-time detectors is structural misplacement, not the low-rank parameterization itself. YOLO-PEFT formulates placement as an auditable multi-constraint decision problem on a directed acyclic graph of the detector, where every module carries operator metadata, a semantic role, graph-interface metadata, and deployment metadata. The planner applies explicit constraint classes in a fixed order: operator validity, detection-semantic safety, graph-interface compatibility, architecture-conditioned policies, budget feasibility, deployment compatibility, and reliability calibration, and returns either an accepted plan with a rank assignment or Refuse with a terminal reason code. Within the evaluated VOC protocol and detector families, planner-selected RS-LoRA matches or exceeds full fine-tuning on YOLO11s and YOLO12s, while RT-DETR-L produces a calibrated Refuse because every one of the seven swept LoRA-family configurations falls below $\Delta = -0.05$; the paper explicitly scopes these results to the evaluated families and calibration coverage.

Load-bearing premise

The load-bearing premise is that the reliability calibration, with coefficients fitted on the same five detector families used for evaluation, can make pre-training Refuse decisions meaningful within claimed calibrated coverage; because no family is held out, the RT-DETR-L refusal is validated on data that already determined that refusal.

Editorial extensions

If this is right

  • Planner-selected PEFT can replace full fine-tuning for YOLO11s and YOLO12s on the VOC protocol, gaining +7.1 and +6.5 mAP50-95 over Full-SFT while freezing most of the detector weights.
  • Refuse is a legitimate planner outcome: on RT-DETR-L it avoids seven training runs that would each collapse below $\Delta = -0.05$, and Full-SFT becomes the fallback rather than a failure.
  • Adapter training trades memory for time: on the controlled YOLO11 audit, LoRA cuts peak VRAM by 43.9 percent but runs 1.72 times longer, so efficiency claims must separate VRAM from wall-clock.
  • Accepted plans survive the full deployment contract: adapters can be saved, reloaded, merged, and exported through ONNX and TensorRT, with merged models recovering base operator cost.

Reading between the lines

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

  • If the planner is applied to a detector family outside the five calibrated ones, the Refuse decision should be re-derived from that family's own collapses; the paper's own admission that held-out-architecture validation remains open suggests the core value is the auditable decision log rather than the refusal threshold itself.
  • The same graph-role filters could block unsafe placements for cross-modal and mixture-of-transformers detectors before any training, because the predicates are structural rather than task-specific.
  • A decisive test of the calibration's predictive power would hold out an entire detector family at calibration time and measure false-refusal and false-acceptance rates on that held-out family.
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

4 major / 5 minor

Summary. The paper proposes YOLO-PEFT, a structure-aware constraint-planning framework for adapter placement on real-time detectors. Given a detector graph, a PEFT request, and a resource budget, the framework parses operator and semantic roles, applies hard constraints (operator validity, detection-head semantics, graph-interface, deployment), allocates ranks under a budget, and either emits a placement plan or returns a Refuse decision with a Full-SFT fallback. The main empirical claim is that under the official VOC07+12 trainval-to-VOC07 test protocol, planner-selected RS-LoRA reaches 0.7138 and 0.7307 mAP50-95 on YOLO11s and YOLO12s, versus 0.6428 and 0.6662 for Full-SFT, while RT-DETR-L returns Refuse because all seven evaluated LoRA-family configurations cross a predefined catastrophe threshold of Δ = −0.05. The paper also reports a controlled efficiency audit (LoRA reduces peak VRAM by 43.9% but training takes 1.72× longer), a merge-equivalence proof for a fallback grouped-convolution LoRA backend, and an scoped MoE stress test. The authors repeatedly and explicitly state that the Refuse rule is calibrated only within the evaluated families and that refusal on unseen detector architectures remains an open validation problem.

Significance. If the results hold, the paper makes a useful contribution by recasting PEFT placement as an auditable constraint-satisfaction problem rather than module-name matching, and by demonstrating on two YOLO variants that graph-aware placement can substantially beat full fine-tuning on a mid-scale detection benchmark. The paper is unusually transparent about its limitations: it discloses test-set involvement in checkpoint selection, reports a controlled memory/time trade-off rather than a universal speedup, and scopes the Refuse mechanism to calibrated coverage. It also ships a deployment contract with ONNX/TensorRT checks and a merge-equivalence proof for the fallback backend. The main weakness is that the Refuse-before-training capability is not yet validated as a predictive tool for unseen architectures; current evidence is a within-sample classification of observed collapses on the evaluated families.

major comments (4)
  1. [Sec. 4.2, Eq. (5)] The reliability calibration in Eq. (5) is fitted and assessed on the same seven RT-DETR-L runs, and the paper itself states that this 'is not a held-out-architecture test.' The variant-level LOVO (86.7% accuracy, F1 = 0.850) holds out variants, not detector families, so it does not demonstrate that the Refuse rule can predict a collapse for an unseen architecture-adapter combination. Since the framework's advertised contribution is 'Refuse before training,' the authors should either add a held-out-family validation (e.g., calibrate on four of the five families and test the refusal decision on the fifth) or explicitly reframe the contribution as a within-coverage post-hoc risk assessment and remove any implication of pre-training prediction for new architectures.
  2. [Sec. 4.1 and Appendix H] The primary mAP comparisons are obtained by selecting the best checkpoint on the VOC2007 test set for every method. The paper discloses this and calls the results 'controlled relative comparisons,' but model selection on the test set can inflate absolute scores and potentially alter rankings in ways not captured by the shared protocol. Because the central claim is that planner-selected PEFT outperforms Full-SFT, the authors should report at least one validation-based selection (for example, on a held-out subset of the training set or on VOC2012 val) to confirm that the superiority holds when the test set is not used for checkpoint selection.
  3. [Sec. 4.3, Table 4] In the YOLO-Master-EsMoE-S stress test, the Full-SFT anchor is taken from a run whose W&B state is logged as 'crashed.' Using metrics from a crashed run as a baseline anchor is not reliable, even if the row is secondary and marked with a dagger. This row should be re-run to completion or omitted from the stress-test table; as presented, the 'HRA beats Full-SFT' observation in that table rests on a possibly corrupted anchor.
  4. [Sec. 3.2, Eq. (5)] The catastrophe threshold Δ = −0.05 is described as 'predefined' but no justification or sensitivity analysis is provided. Since the RT-DETR-L Refuse decision is driven entirely by this threshold, a short analysis showing that the decision is robust to reasonable variations (for example, Δ = −0.03 or −0.08) would materially strengthen the claim that the rule is not an artifact of a single manually chosen constant.
minor comments (5)
  1. [Sec. 3.2, Eq. (4)] The utility function u(i, π(i); p, ψ) is not fully specified in the main text; the rule-only decomposition u = u_op + u_sem + u_range + u_rank − λc_p is given only in words. Please provide explicit definitions or a more detailed supplementary reference in the main text.
  2. [Supplementary Tab. S2] The heatmap in Fig. 2 clips colors below −0.30, but several entries are −0.600 (the clipped failure value). The caption and table should state this clipping explicitly so readers do not interpret the clipped cells as equal to −0.30.
  3. [Sec. 4.1] The explanation that the W&B field 'val2007' is an alias for the VOC2007 test set is important and easily overlooked; consider moving it from Appendix H into the main experimental setup for greater prominence.
  4. [Fig. 1] In the top pipeline, the 'Feasible? No' branch leads to 'REFUSE' and then to 'Full-SFT fallback'; consider labeling this fallback as a valid, auditable outcome in the figure, matching the paper's framing that Refuse is a first-class planning result.
  5. [Sec. 4.2] The LOVO accuracy of 86.7% and F1 of 0.850 are reported without a description of the exact folds (e.g., number of variants, class balance, or variance across folds). A one-sentence fold description would improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

RT-DETR-L Refuse is calibrated and counted on the same seven collapses, so the refusal is an in-sample fit rather than an independent prediction.

  1. fitted input called prediction [Sec. 3.2 'Reliability calibration' (Eq. 5); Sec. 4.2 'Refusal within Evaluated Coverage'; Tab. 1 footnote]
    "calibrated mode estimates ∆mAP≈β0 +β1ϕattn +β2ϕtext +β3ϕdw +β4ξp, (5) where ξp is a variant coefficient fitted on the canonical matrix. ... The RT-DETR-L rule is calibrated on seven observed collapses, not an unseen architecture. ... Treating ∆<−0.05 as unsafe, these seven attempted configurations have a 7/7 (100%) unsafe-placement rate."

    The coefficients in Eq. (5) are fitted on the same seven RT-DETR-L LoRA-family runs that are then reported as all crossing the catastrophe threshold, so the 7/7 refusal is a re-statement of the calibration data rather than a before-training prediction for a new architecture-adapter combination. The paper explicitly disclaims held-out-architecture status ('this is not a held-out-architecture test'), but the abstract's 'calibrated Refuse-to-Full-SFT decision' and the claimed pre-training risk-estimation contribution still rely on an in-sample classification. The Refuse outcome is forced by the fitted model plus the manually chosen threshold, and it cannot by itself demonstrate that the planner detects high-risk combinations before training.

full rationale

The central empirical claims are not circular: the YOLO11s/YOLO12s mAP comparisons use the official VOC07+12 trainval-to-VOC07 test protocol, and the memory/time audit is a controlled measurement with disclosed scope. The constraint-planning rules, budget allocation, and merge-equivalence proof are deterministic and independent of the reliability model. The one load-bearing circular step is the RT-DETR-L Refuse mechanism: Eq. (5) is fitted on the canonical matrix that includes RT-DETR-L, the catastrophe threshold is applied to those same seven runs, and the resulting 7/7 refusal is presented as supporting a calibrated Refuse decision. The paper is unusually transparent about this limitation, which lowers the severity, but the 'prediction' still reduces to the fit by construction. Score 6: partial circularity in the refusal/safety contribution while the placement and efficiency contributions retain independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim depends on hand-coded safety predicates, a fitted reliability calibration with hand-picked threshold, and a test-set-based selection protocol. The 10-D fingerprint is an internal abstraction without external validation. No new physical entities are introduced.

free parameters (4)
  • Reliability calibration coefficients beta_0..beta_4 = Not reported numerically; fitted on canonical matrix (Tab. 1)
    Eq. (5) approximates delta mAP as beta_0 + beta_1*phi_attn + beta_2*phi_text + beta_3*phi_dw + beta_4*xi_p. These coefficients are fitted on the same evaluated matrix and drive the RT-DETR-L Refuse decision.
  • Catastrophe threshold delta = -0.05 mAP50-95
    Predefined threshold for defining a catastrophic placement and triggering Refuse; chosen by hand rather than derived.
  • Default rank and alpha = r=16, alpha=32
    Default PEFT configuration used for the headline planner-selected RS-LoRA runs; the rank sweep suggests r=16 as a Pareto choice, but it is a chosen hyperparameter.
  • Training-prior hyperparameters = lambda_ortho=0.5, rho=0.85, dropout 0 to 0.15
    Hand-chosen settings in Sec. 3.4 (orthogonal regularization strength, layer-wise LR decay, dynamic dropout) that are part of the reported recipe but not central to the placement claim.
assumptions (4)
  • domain assumption Manual operator-validity and semantic-safety predicates are correct (depthwise unsafe, DFL projections unsafe, MoE routers unsafe, geometry-sensitive regression paths unsafe).
    Sec. 3.2 lists these as mandatory filters; the ablation attributes the largest mAP gain to head-semantic exclusions. If these rules are wrong, the planner's gains could come from unrelated effects.
  • domain assumption Family-scoped architecture policies calibrated on evaluated families apply to new configurations within the same families.
    Sec. 3.2 states these are empirical safeguards, not unseen-family guarantees. The RT-DETR-L refusal transfers diagnostic-matrix collapses (300 epochs, 320px) to core-protocol decisions without held-out family validation.
  • domain assumption VOC2007 test-set-based best-checkpoint selection yields controlled relative comparisons.
    Appendix H discloses that VOC2007 test participates in the common checkpoint-selection protocol; the paper treats all methods as fairly compared relative to each other. This assumes selection bias affects methods roughly equally.
  • standard math Standard linear algebra and im2col convolution-to-matrix equivalence hold for the merge-equivalence proof.
    Appendix B proves merge equivalence informally for the fallback Conv2d LoRA backend, relying on standard convolution-as-matrix-multiplication facts rather than machine-checked proof.
invented entities (1)
  • 10-dimensional architecture fingerprint phi(G)
    purpose: Characterizes detector structure (attention ratio, text fusion, MoE, depthwise, dense/grouped conv ratios, plus five scale statistics) for reliability calibration and refusal decisions.
    Introduced in Sec. 3.1. The core five dimensions drive stability analysis and variant-level LOVO, but the paper explicitly states the fingerprint is not validated on unseen detector families, so it has no independent falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family." pith.science (2026). https://pith.science/paper/35VWECDK

@misc{pith2026260807051,
  author       = {Pith},
  title        = {Pith review of: YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/35VWECDK}},
  note         = {Machine review of arXiv:2608.07051}
}
read the original abstract

Generic parameter-efficient fine-tuning (PEFT) methods transferred from language models can fail silently on real-time detectors, whose heterogeneous operators and detection-specific components impose placement constraints absent from regular Transformer stacks. We propose YOLO-PEFT, a structure-aware framework that formulates adapter placement as an auditable constraint-planning problem. Given a detector graph, a PEFT request, and a resource budget, YOLO-PEFT assigns operator and semantic roles, evaluates explicit operator-validity, detector-semantic, graph-interface, and deployment predicates, records a reason code for each excluded module, and either emits a budgeted target-module plan or returns Refuse before training. Under the official VOC07+12 trainval-to-VOC07 test protocol, planner-selected RS-LoRA reaches 0.7138 and 0.7307 mAP50-95 on YOLO11s and YOLO12s, respectively, compared with 0.6428 and 0.6662 for Full-SFT. On RT-DETR-L, all seven evaluated LoRA-family configurations cross the predefined catastrophic threshold, supporting a calibrated Refuse-to-Full-SFT decision within the evaluated coverage. A controlled YOLO11 audit further shows that LoRA reduces peak training memory by 43.9 percent, although training takes 1.72 times longer. Within the evaluated detector families, placement policies, and calibration coverage, YOLO-PEFT replaces manual target-module trial and error with explicit, inspectable planning while preserving verified train-save-merge-export paths; refusal on unseen detector architectures remains an open validation problem. Project Page: github.com/Tencent/YOLO-Master

Figures

Figures reproduced from arXiv: 2608.07051 by the authors.

Figure 1
Figure 1. System overview. The pipeline (1) parses a detector from a supported, evaluated family into a role-aware typed graph, (2) filters unsafe targets, selects the PEFT method, and resolves an adapter placement π via budget-aware rank assignment P i ci(ri)≤B, or, within calibrated coverage, returns Refuse with a Full-SFT fallback, (3) trains sidecar adapters while freezing the base weights (W=W0+∆W(π)) and saves an adapte… view at source ↗
Figure 2
Figure 2. Controlled PEFT efficiency audit. On the audited YOLO11 configuration, adapter training reduces peak VRAM but increases wall-clock time. The figure reports only the controlled measurements used in this paper. of magnitude. Methods like DoRA amplify this mismatch via their magnitude decomposition, leading to rapid loss divergence within 15 epochs and yielding a negative delta of ∆ = −0.292. Furthermore, adapting regr… view at source ↗
Figure 3
Figure 3. Architecture-conditioned PEFT behavior. ∆mAP versus Full-SFT for 14 variants on five heterogeneous detectors. Red cells are catastrophic (∆ < −0.05); gray denotes not swept and colors clip below −0.30. The catastrophe rate rises with ϕattn, and no common ranking survives across the five evaluated families. pressure rather than inevitable collapse. Conversely, the depthwise predicate admits no additional rank-4 targe… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Paper-measured controlled YOLO11 systems audit. Relative to Full-SFT, LoRA lowers peak VRAM from 28.57 to 16.03 GB (43.9%) but increases training time from 118.2 to 203.8 s (1.72×); all displayed PEFT variants are slower than Full-SFT. Tab. 6. MPS deployment audit on c…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 38 canonical work pages

  1. [1]

    Conv-Adapter: Exploring parameter efficient transfer learning for ConvNets

    Hao Chen, Ran Tao, Han Zhang, Yidong Wang, Xiang Li, Wei Ye, Jindong Wang, Guosheng Hu, and Marios Savvides. Conv-Adapter: Exploring parameter efficient transfer learning for ConvNets. InCVPR Workshops, pages 1551–1561, 2024. doi: 10.1109/CVPRW63382.2024.00162. URLhttps://arxiv.org/abs/2208.07463

    arXiv 2024

  2. [2]

    AdaptFormer: Adapting vision transformers for scalable visual recognition

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. AdaptFormer: Adapting vision transformers for scalable visual recognition. InNeurIPS, 2022

    work page 2022

  3. [3]

    VFM-Adapter: Adapting visual foundation models for dense prediction with dynamic hybrid operation mapping

    Zheng Chen, Yu Zeng, Zehui Chen, Hongzhi Gao, Lin Chen, Jiaming Liu, and Feng Zhao. VFM-Adapter: Adapting visual foundation models for dense prediction with dynamic hybrid operation mapping. InAAAI, volume 39, pages 2385–2393, 2025. doi: 10.1609/aaai.v39i3.32239. URLhttps://ojs.aaai.org/index.php/AAAI/article/ view/32239

    work page doi:10.1609/aaai.v39i3.32239 2025

  4. [4]

    YOLO-World: Real-time open-vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. YOLO-World: Real-time open-vocabulary object detection. InCVPR, 2024

    work page 2024

  5. [5]

    Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL visual object classes challenge: A retrospective.International Journal of Computer Vision, 111:98–136, 2015

    work page 2015

  6. [6]

    Lightweight modular parameter-efficient tuning for open- vocabulary object detection, 2024

    Bilal Faye, Hanane Azzag, and Mustapha Lebbah. Lightweight modular parameter-efficient tuning for open- vocabulary object detection, 2024. URLhttps://arxiv.org/abs/2408.10787

    work page arXiv 2024

  7. [7]

    Pet-dino: Unifying visual cues into grounding dino with prompt-enriched training

    Weifu Fu, Jinyang Li, Bin-Bin Gao, Jialin Li, Yuhuan Lin, Hanqiu Deng, Wenbing Tao, Yong Liu, and Chengjie Wang. Pet-dino: Unifying visual cues into grounding dino with prompt-enriched training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13039–13048, June 2026

    work page 2026

  8. [8]

    Multi-point positional insertion tuning for small object detection

    Kanoko Goto, Takumi Karasawa, Takumi Hirose, Rei Kawakami, and Nakamasa Inoue. Multi-point positional insertion tuning for small object detection. InICASSP, pages 1–5, 2025. doi: 10.1109/ICASSP49660.2025.10888925. URLhttps://arxiv.org/abs/2412.18090. 19

    arXiv 2025

Show all 50 references
  1. [9]

    Parameter-efficient fine-tuning for large models: A comprehensive survey, 2024

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey, 2024

    2024

  2. [10]

    Sensitivity-aware visual parameter- efficient fine-tuning

    Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter- efficient fine-tuning. InICCV, pages 11791–11801, 2023. doi: 10.1109/ICCV51070.2023.01086. URL https: //arxiv.org/abs/2303.08566

    2023

  3. [11]

    Towards a unified view of parameter-efficient transfer learning

    Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. InICLR, 2022

    2022

  4. [12]

    Parameter-efficient model adaptation for vision transformers

    Xuehai He, Chunyuan Li, Pengchuan Zhang, Jianwei Yang, and Xin Eric Wang. Parameter-efficient model adaptation for vision transformers. InAAAI, volume 37, pages 817–825, 2023. doi: 10.1609/aaai.v37i1.25160. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/25160

    2023 doi

  5. [13]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InICLR, 2022

    2022

  6. [14]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. InECCV, 2022

    2022

  7. [15]

    Convolutional bypasses are better vision transformer adapters

    Shibo Jie, Zhi-Hong Deng, Shixuan Chen, and Zhijuan Jin. Convolutional bypasses are better vision transformer adapters. InECAI, pages 202–209, 2024. doi: 10.3233/FAIA240489. URLhttps://arxiv.org/abs/2207.07039

    2024 arXiv

  8. [16]

    YOLOv8 by Ultralytics.https://github.com/ultralytics/ult ralytics, 2023

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLOv8 by Ultralytics.https://github.com/ultralytics/ult ralytics, 2023. GitHub repository, Accessed: 2026-05-10

    2023

  9. [17]

    YOLO11 by Ultralytics.https://github.com/ultralytics/ult ralytics, 2024

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO11 by Ultralytics.https://github.com/ultralytics/ult ralytics, 2024. GitHub repository and documentation, Accessed: 2026-05-10

    2024

  10. [18]

    DA-Ada: Learning domain-aware adapter for domain adaptive object detection

    Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, Xiaqing Li, Yongwei Zhao, Ling Li, and Yunji Chen. DA-Ada: Learning domain-aware adapter for domain adaptive object detection. InNeurIPS, volume 37, pages 103574–103598, 2024. doi: 10.52202/079017-3289. URLh...

    2024 arXiv

  11. [19]

    Lors: Low-rank residual structure for parameter-efficient network stacking

    Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, and Chengjie Wang. Lors: Low-rank residual structure for parameter-efficient network stacking. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15866–15876, 2024

    2024

  12. [20]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection

    Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. InNeurIPS, 2020

    2020

  13. [21]

    Scaling & shifting your features: A new baseline for efficient model tuning

    Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. InNeurIPS, 2022

    2022

  14. [22]

    YOLO-Master: MOE-accelerated real-time detection,

    Xu Lin, Jinlong Peng, Zhenye Gan, Jiawen Zhu, and Jun Liu. YOLO-Master: MOE-accelerated real-time detection,

  15. [23]

    Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning

    Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. InNeurIPS, 2022

    2022

  16. [24]

    DoRA: Weight-decomposed low-rank adaptation

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. InICML, 2024

    2024

  17. [25]

    RT-DETR: DETRs beat YOLOs on real-time object detection

    Wenyu Lv, Yi Zhao, Shangliang Xu, Jinman Wei, Guanzhong Wang, Cheng Cui, Yuning Du, Qingqing Dang, and Yi Liu. RT-DETR: DETRs beat YOLOs on real-time object detection. InCVPR, 2024

    2024

  18. [26]

    PEFT: State-of-the-art parameter-efficient fine-tuning methods.https://github.com/huggingface/peft, 2022

    Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, and Sayak Paul. PEFT: State-of-the-art parameter-efficient fine-tuning methods.https://github.com/huggingface/peft, 2022. Accessed: 2026-05-10

    2022

  19. [27]

    Munish Monga, Vishal Chudasama, Pankaj Wasnik, and C. V. Jawahar. EW-DETR: Evolving world object detection via incremental low-rank detection transformer, 2026. URLhttps://arxiv.org/abs/2602.20985

    2026

  20. [28]

    Pro-Tuning: Unified prompt tuning for vision tasks.IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4653–4667, 2024

    Xing Nie, Bolin Ni, Jianlong Chang, Gaofeng Meng, Chunlei Huo, Zhaoxiang Zhang, Shiming Xiang, Qi Tian, and Chunhong Pan. Pro-Tuning: Unified prompt tuning for vision tasks.IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4653–4667, 2024. doi: 10.1109/TCSV...

    2024

  21. [29]

    Xinyang Pu and Feng Xu. Low-rank adaption on transformer-based oriented object detector for satellite onboard processing of remote sensing images.IEEE Transactions on Geoscience and Remote Sensing, 63:1–13, 2025. doi: 10.1109/TGRS.2024.3524578. URLhttps://arxiv.org/abs/2406.02385

    2025

  22. [30]

    Learning multiple visual domains with residual adapters

    Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. InNeurIPS, 2017

    2017

  23. [31]

    VL-Adapter: Parameter-efficient transfer learning for vision- and-language tasks

    Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. VL-Adapter: Parameter-efficient transfer learning for vision- and-language tasks. InCVPR, pages 5217–5227, 2022. doi: 10.1109/CVPR52688.2022.00516. URL https: //arxiv.org/abs/2112.06825

    2022

  24. [32]

    Analyzing the impact of low-rank adaptation for cross-domain few-shot object detection in aerial images,

    Hicham Talaoubrid, Anissa Mokraoui, Ismail Ben Ayed, Axel Prouvost, Sonimith Hang, Monit Korn, and Rémi Harvey. Analyzing the impact of low-rank adaptation for cross-domain few-shot object detection in aerial images,

  25. [33]

    YOLOv12: Attention-centric real-time object detectors

    Yunjie Tian, Qixiang Ye, and David Doermann. YOLOv12: Attention-centric real-time object detectors. In NeurIPS, 2025

    2025

  26. [34]

    Source-free domain adaptation for YOLO object detection

    Simon Varailhon, Masih Aminbeidokhti, Marco Pedersoli, and Eric Granger. Source-free domain adaptation for YOLO object detection. InComputer Vision – ECCV 2024 Workshops, pages 218–235. Springer, 2025. doi: 10.1007/978-3-031-91672-4_14

    2024 doi

  27. [35]

    SIA-OVD: Shape-invariant adapter for bridging the image-region gap in open-vocabulary detection

    Zishuo Wang, Wenhao Zhou, Jinglin Xu, and Yuxin Peng. SIA-OVD: Shape-invariant adapter for bridging the image-region gap in open-vocabulary detection. InACMMM, pages 4986–4994, 2024. doi: 10.1145/3664647.3680642. URLhttps://arxiv.org/abs/2410.05650

    2024

  28. [36]

    CoPEFT: Fast adaptation framework for multi-agent collaborative perception with parameter-efficient fine-tuning

    Quanmin Wei, Penglin Dai, Wei Li, Bingyi Liu, and Xiao Wu. CoPEFT: Fast adaptation framework for multi-agent collaborative perception with parameter-efficient fine-tuning. InAAAI, volume 39, pages 23351–23359, 2025. doi: 10.1609/aaai.v39i22.34502. URLhttps://ojs.aaai.org/index...

    2025 doi

  29. [37]

    DroneFINE: Domain-aware parameter-efficient fine-tuning of vision-language detectors for drone images, 2026

    Ke Wu, Yanan Zhang, Yingjie Gao, Wenhao Li, Chenyu Zhou, Xinzhu Ma, Jiaxin Chen, and Di Huang. DroneFINE: Domain-aware parameter-efficient fine-tuning of vision-language detectors for drone images, 2026. URLhttps://arxiv.org/abs/2607.00338

    2026 arXiv

  30. [38]

    VMT-Adapter: Parameter-efficient transfer learning for multi-task dense scene understanding

    Yi Xin, Junlong Du, Qiang Wang, Zhiwen Lin, and Ke Yan. VMT-Adapter: Parameter-efficient transfer learning for multi-task dense scene understanding. InAAAI, volume 38, pages 16085–16093, 2024. doi: 10.1609/aaai.v38i1 4.29541. URLhttps://arxiv.org/abs/2312.08733

    2024 arXiv

  31. [39]

    Pre-train, adapt and detect: Multi-task adapter tuning for camouflaged object detection, 2023

    Yinghui Xing, Dexuan Kong, Shizhou Zhang, Geng Chen, Lingyan Ran, Peng Wang, and Yanning Zhang. Pre-train, adapt and detect: Multi-task adapter tuning for camouflaged object detection, 2023. URLhttps: //arxiv.org/abs/2307.10685. Preprint

    2023 arXiv

  32. [40]

    Component-coordinated and uncertainty-enhanced LoRA for few-shot source-free domain adaptive object detection.Neurocomputing, 650:130787, 2025

    Xudong Yao, Hao Liu, and Xiaoshan Yang. Component-coordinated and uncertainty-enhanced LoRA for few-shot source-free domain adaptive object detection.Neurocomputing, 650:130787, 2025. doi: 10.1016/j.neucom.2025.13 0787

    2025 doi

  33. [41]

    SpotPatch: Parameter-efficient transfer learning for mobile object detection

    Keren Ye, Adriana Kovashka, Mark Sandler, Menglong Zhu, Andrew Howard, and Marco Fornoni. SpotPatch: Parameter-efficient transfer learning for mobile object detection. InComputer Vision – ACCV 2020, pages 239–256. Springer, 2021. doi: 10.1007/978-3-030-69544-6_15. URLhttps://a...

    2020 arXiv

  34. [42]

    1% vs 100%: Parameter- efficient low rank adapter for dense predictions

    Dongshuo Yin, Yiran Yang, Zhechao Wang, Hongfeng Yu, Kaiwen Wei, and Xian Sun. 1% vs 100%: Parameter- efficient low rank adapter for dense predictions. InCVPR, pages 20116–20126, 2023. doi: 10.1109/CVPR52729.20 23.01926. URL https://openaccess.thecvf.com/content/CVPR2023/html/...

    2023 doi

  35. [43]

    Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions

    Dongshuo Yin, Xueting Han, Bin Li, Hao Feng, and Jing Bai. Parameter-efficient is not sufficient: Exploring parameter, memory, and time efficient adapter tuning for dense predictions. InACMMM, pages 1398–1406, 2024. doi: 10.1145/3664647.3680940. URLhttps://arxiv.org/abs/2306.09729

    2024

  36. [44]

    Bridging the gap between low-rank and orthogonal adaptation via householder reflection adaptation

    Shen Yuan, Haotian Liu, and Hongteng Xu. Bridging the gap between low-rank and orthogonal adaptation via householder reflection adaptation. InNeurIPS, 2024

    2024

  37. [45]

    REAL-OW: Rehearsal-free open world object detection with low-rank adaptation and dual-stage objectness modeling, 2026

    Huazhong Zhang, Xiaowen Fu, Yang Zhang, Linlin Shen, and Jinbao Wang. REAL-OW: Rehearsal-free open world object detection with low-rank adaptation and dual-stage objectness modeling, 2026. URL https: //arxiv.org/abs/2607.03004. Preprint. 21

    2026 arXiv

  38. [46]

    AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. InICLR, 2023

    2023

  39. [47]

    YOLO- IOD: Towards real time incremental object detection

    Shizhou Zhang, Xueqiang Lv, Yinghui Xing, Qirui Wu, Di Xu, Chen Zhao, and Yanning Zhang. YOLO- IOD: Towards real time incremental object detection. InAAAI, volume 40, pages 12744–12752, 2026. doi: 10.1609/aaai.v40i15.38271. URLhttps://arxiv.org/abs/2512.22973

    2026

  40. [48]

    SCT: A simple baseline for parameter-efficient fine-tuning via salient channels.International Journal of Computer Vision, 132(3): 731–749, 2024

    Henry Hengyuan Zhao, Pichao Wang, Yuyang Zhao, Hao Luo, Fan Wang, and Mike Zheng Shou. SCT: A simple baseline for parameter-efficient fine-tuning via salient channels.International Journal of Computer Vision, 132(3): 731–749, 2024. doi: 10.1007/s11263-023-01918-3. URLhttps://a...

    2024 arXiv

  41. [49]

    AutoPEFT: Automatic configuration search for parameter-efficient fine-tuning

    Han Zhou, Xingchen Wan, Ivan Vulić, and Anna Korhonen. AutoPEFT: Automatic configuration search for parameter-efficient fine-tuning. InTACL, 2024. 22

    2024

  42. [2025]

    Preprint

    URLhttps://arxiv.org/abs/2504.06330. Preprint

    arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.