Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Interpretable Open-Vocabulary Referring Object Detection with Reverse Contrast Attention

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Reverse Contrast Attention, a training-free reweighting of final-layer attention, improves FitAP in 11 of 15 open-vocabulary referring object detection models, with gains up to +26.6%, and is equivalent to flooring hidden states at a…

desk verdict RCA is a neat idea, but its headline gains rest on an oracle ranking metric that uses ground-truth IoU to sort detections. read the letter →

arxiv 2507.19891 v2 pith:KHY2PZN2 submitted 2025-07-26 cs.CV cs.AI

Drandreb Earl O. Juanico , Rowel O. Atienza , Jeffrey Kenneth Go This is my paper
classification cs.CVcs.AI
keywords reversecontrastattentionopen-vocabularyreferringobjectdetectionvision-languagetransformerinterpretabilityinference-timeadaptationFitAPlocalizationreweighting
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

The paper proposes Reverse Contrast Attention (RCA), a plug-in that reshapes the final transformer layer's attention matrix at inference time, with no retraining and no change to model weights. It evaluates RCA on open-vocabulary referring object detection across 15 open-source vision-language models and reports FitAP gains in 11 of them, up to +26.6%. The paper introduces FitAP, a confidence-free average precision metric that ranks detections by the product of box area and IoU. The core motivation is that attention plasticity, previously read as evidence that attention cannot explain outputs, can be deliberately guided to improve localization. If the results hold, frozen VLMs can be made better at grounding language to image regions simply by post-processing their final attention map.

What carries the argument

The central object is the reverse-contrast reweighting of the final-layer attention matrix: each attention weight $\alpha_{ij}$ is passed through a nonmonotonic map—either inverse distance from a midpoint $m$ or a Gaussian peak around $m$—and the resulting weights are renormalized (Eq. 2). The paper's key identity is Eq. (3), which states that RCA is equivalent to flooring every final-layer hidden state at a threshold $\vartheta$. The formal workhorse is condition (6), a lower bound showing the post-RCA hidden state stays near or above $\vartheta$ whenever the renormalized attention mass assigned to subthreshold tokens is small. This penalty term links RCA's effect to attention sharpness $m$, and the paper reports a statistically significant negative correlation between $m$ and the number of subthreshold contributions for the models that improve.

What would settle it

A direct test would rerun the same 15 VLMs through the same prompts but rank detections by a real confidence signal, such as the model's logit or a post-hoc calibration score, instead of the area-IoU product, and compare pre- and post-RCA AP. If the reported gains shrink or disappear under non-oracle ranking, the central empirical claim would be undermined. The paper's own condition (6) could also be checked directly by measuring whether models with lower attention sharpness $m$ consistently fail to benefit from RCA on a new model not in the table.

Watch

Extended reading notes

Core claim

The central claim is that suppressing the most extreme attention weights and amplifying mid-level weights in the final transformer layer improves open-vocabulary referring object detection without touching any model parameters. The paper derives an equivalence: this reverse-contrast reweighting of the attention matrix acts as an element-wise flooring of the final-layer hidden states at a threshold $\vartheta$ (Eq. 3). A soft guarantee (condition 6) shows that the transformed hidden state is bounded below by $\vartheta$ plus a penalty term, proportional to the attention mass assigned to subthreshold tokens, that shrinks as attention sharpness grows. Empirically, RCA improves FitAP in 11 of 15 open-source VLMs, with the largest gains on late-fusion modular architectures; the authors argue the mechanism works by reducing noise from low-attention tokens and sharpening object localization.

Load-bearing premise

The reported FitAP gains depend on the assumption that ranking detections by the product of box area and IoU with ground truth behaves like a genuine confidence score; if that oracle-style ordering does not reflect the order a deployed model would produce, the improvements measured by FitAP may not translate to real-world localization gains.

Editorial extensions

If this is right

  • RCA can be applied to a frozen, pretrained VLM at inference time, giving localization gains without any training or weight modification.
  • The largest FitAP improvements appear in models with late, modular vision-language fusion, so RCA prescribes a simple check: apply it to such models first.
  • The early-fusion counterexample in the paper shows that early-fusion architectures are not automatically excluded; high-capacity models with latent separation can still benefit.
  • FitAP makes it possible to benchmark detection quality in VLMs that do not emit confidence scores, preserving the precision-recall structure of AP.
  • Since RCA changes only the attention map, the same intervention offers an interpretability lens: the post-RCA attention highlights the subdued tokens that were already semantically relevant.

Reading between the lines

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

  • An extension the authors do not test: FitAP's ranking by the product of box area and IoU is an oracle-style ordering relative to ground truth; replacing it with a genuine confidence ordering such as logits is the most direct way to see whether the reported gains survive deployment conditions.
  • If RCA's mechanism is right, then models with sharper final-layer attention (higher $m$) should be the reliable beneficiaries; this yields a testable pre-screening rule for new VLMs that the paper does not formalize.
  • The same reweighting could be transferred to other prompt-based localization tasks such as open-vocabulary detection, phrase grounding, or segmentation, though the paper only evaluates referring object detection.
  • Because the threshold $\vartheta$ was fixed rather than optimized in the experiments, tuning $\vartheta$ per model or per prompt class is a plausible and untested route to larger gains.
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 / 4 minor

Summary. The paper proposes Reverse Contrast Attention (RCA), a training-free inference-time modification of the final-layer attention weights in vision-language models, and evaluates it on open-vocabulary referring object detection (OV-RefOD). RCA suppresses extreme attention values and amplifies mid-level ones, with the paper claiming an equivalence to flooring hidden states. The authors introduce FitAP, a confidence-free average precision metric that ranks detections by Abox × IoU, and report FitAP gains for 11 of 15 VLMs (Table 2), up to +26.6%, together with interpretability analyses. The central claims are that RCA improves localization and that the improvement is explained by attention sharpness and vision-language fusion timing.

Significance. If the central claim were established, the contribution would be significant: an inference-time intervention with a mechanistic explanation, broad coverage across 15 open models, and a proposed evaluation protocol for VLMs that do not emit confidence scores. The paper is honest about the softness of its theoretical bound and releases code and data. However, the quantitative results currently rest on an evaluation metric whose ranking key is computed from the same IoU used to label true positives, making the reported gains partly self-referential. Until FitAP is validated against a confidence-based or ranking-free metric, or replaced by one, the 11-of-15 result does not support the paper's central claim. The underlying attention-reweighting idea remains plausible and worth testing with standard metrics.

major comments (3)
  1. [Section 2.5 and Supplementary S5] FitAP is a self-referential (oracle-style) ranking. Each detection is ranked by Abox × IoU(bbox, GT), and then the same IoU value is thresholded to label that detection as true or false positive. Sorting by the label-generating quantity means the precision-recall curve is constructed in near-best-case order: any detection whose IoU is just above the threshold outranks every detection just below it, regardless of the model's actual output order at inference. The validation in S5.1 (r = 0.90–0.92 with ground-truth area) does not establish the required property, namely that the ranking approximates the model's confidence or true-positive probability; it only shows that larger objects are detected more reliably. Consequently, Table 2's +26.6% and the 11-of-15 count may reflect the metric's sensitivity to box area rather than genuine localization improvement. Please re-evaluate with a confidence-based AP for models that provide confidence, or with ranking-free metrics such as recall at fixed IoU thresholds, average IoU over matched detections, or precision at a fixed number of detections, reporting both pre- and post-RCA values.
  2. [Section 3.1 and Eq. (6)] The derivation does not establish the claimed equivalence in Eq. (3) between RCA and element-wise flooring eZ = {max(zi, ϑ)}. Equation (6) is a lower bound ezi(d) ≥ ϑ + (v− − ϑ) Σ_{j∈J↓} fαij, and the conclusion that ezi(d) ≈ ϑ requires the subthreshold mass Σ_{j∈J↓} fαij to be small. That smallness is an assumption stated in Section 3.1, not a consequence of the reweighting; the paper even acknowledges it may fail for some models. Moreover, the empirical test in Section 3.2 uses per-model ϑ values (−1.5, −13, −1.5) that are manually selected, and the correlation for Qwen2.5-VL-7B (r = −0.09) is very weak despite that model showing the largest FitAP gain. Please provide a direct test of the flooring assumption, for example by comparing the actual transformed hidden states ezi with max(zi, ϑ) on the same inputs, and report the sensitivity of the results to ϑ.
  3. [Section 2.2 and Section 3.2] The method has three free parameters — γ, m, and ϑ — and the paper does not describe how they are chosen. The statement in Section 2.2 that m is 'manually selected' and Section 3.1 that ϑ 'could be optimized' leaves open the possibility that ϑ was selected on the same COCO val 2017 split used for evaluation. If so, the reported improvements are not a fair assessment of a training-free method because the test set has been used for model selection. Please specify the exact selection procedure (validation split, grid search, fixed values) and provide an ablation over γ and m/ϑ to show that the conclusions are not artifacts of particular parameter choices.
minor comments (4)
  1. [Equation (3)] The set-builder notation eZ = {max(zi, ϑ) | zi ∈ Z, ϑ ∈ R} treats ϑ as a bound variable, but ϑ is a fixed parameter elsewhere in the paper; please write ezi = max(zi, ϑ) for each i and define the transformed sequence accordingly.
  2. [Table 2] The MiniCPM-o-2.6 row reports a +139% change, but the absolute FitAP values (0.03064 to 0.07334) are close to zero; the percentage is misleading and should be discussed in absolute terms.
  3. [Supplementary S5.2 and Figure S8] The red dashed 'envelopes' in Figure S8 are not defined; please clarify how the envelope curve is computed and whether the AP values are calculated on the envelope or on the raw precision-recall curve.
  4. [Related Work and Supplementary S6] There are several formatting and typographical errors, including 'Hyeon-Wooet al.' and 'Gausian'; these should be corrected.

Circularity Check

1 steps flagged · score 7.0 of 10

FitAP's oracle-style ranking (Abox x IoU with ground truth) makes the central 11-of-15 improvement claim partially self-referential; the RCA theory itself is not circular.

  1. self definitional [Section 2.5 and Supplementary S5 (FitAP definition and validation)]
    "FitAP ranks predicted bounding boxes according to the product of their normalized area Abox and their intersection-over-union (IoU) with ground-truth annotations. ... In the absence of a confidence score from the parsed VLM detection results, we propose to use the product of the normalized box area of detection and the IoU, Abox × IoU, for the quality ranking of detection against the ground truth data."

    Average precision requires a ranking of detections by an estimate of true-positive probability, but FitAP ranks every detection by Abox × IoU(bbox, GT), where the same IoU with the same GT is then thresholded to label that detection true or false positive. This makes the ranking an oracle ranking: detections that score high are true positives by construction of the scoring key, and detections that score low are largely false positives by the same key. The paper's validation (S5.1) only demonstrates correlation with ground-truth area (r = 0.90-0.92), not with true-positive probability. Consequently, the reported FitAP gains from RCA are not independent evidence of better localization; they reflect, in part, the metric's built-in use of ground-truth match information.

full rationale

The main quantitative claim — RCA improves OV-RefOD in 11 of 15 VLMs — rests on FitAP, a metric introduced in this paper. By construction, FitAP sorts detections by Abox × IoU with the ground-truth box and then labels each detection true/false positive using an IoU threshold on that same match. The ranking therefore contains the label information, making the resulting precision-recall curve highly self-referential: high-IoU detections are preferentially ranked first, guaranteeing a favorable AP shape independent of any notion of confidence. The paper's own words state that the critical step is to ensure the metric 'reliably correlates with the probability of detection being true positive,' yet the provided evidence is only a correlation with GT area, not with true-positive probability. Thus the 11-of-15 improvement and the +26.6% headline gain are stronger statements than the metric can support. The theoretical RCA-to-flooring argument (Eqs. 3–6) is not circular: it is a soft heuristic bound explicitly conditioned on the subthreshold mass being small, and it does not reduce to the data it claims to explain. No load-bearing self-citation chain appears; the cited prior works are contextual and independent. The empirical evaluation, however, is partially circular because the central result is generated by an oracle-ranked metric whose ranking key overlaps with its positive/negative labeling criterion.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that attention is a usable localization signal, on the FitAP ranking assumption that ground-truth overlap can stand in for confidence, and on a per-model floor threshold with no explicit selection protocol. The method itself introduces no new physical entities; its tunable components are gamma, m, and theta.

free parameters (3)
  • gamma (contrast sharpness) = 1 (fixed by choice)
    Section 2.2 says gamma is free but possibly optimizable, and is set to 1 without loss of generality; it controls how sharply deviations from the midpoint m are penalized.
  • m (central attention value) = manual/derived from attention as mean column-wise maximum
    Section 2.2 states m is manually selected; Section 3 defines it as the mean of column-wise max attention across heads. It determines which activations are suppressed versus amplified, and no protocol is given for selecting it per model.
  • theta (floor threshold in Eq. 3) = -1.5 (Qwen2.5-VL), -13 (DeepSeek-VL2), -1.5 (WeThink)
    Theta is a free parameter linked to m and gamma, manually selected per model (Section 2.2, Figure 5 captions). It directly controls the claimed flooring and is likely chosen after seeing results; the paper admits no systematic optimization.
assumptions (4)
  • domain assumption Final-layer attention weights carry usable localization signal that can be reweighted to improve detection.
    This is the core premise of RCA, invoked throughout Sections 1 and 2; the cited Jain and Wallace debate [11] questions whether attention explains outputs, so this is an assumption, not a given.
  • domain assumption Abox x IoU is a reliable substitute for confidence scores when ranking detections for AP.
    FitAP depends on this. Supplementary S5 supports it only with correlations to ground-truth area, not by validation against confidence-based AP.
  • ad hoc to paper The renormalized attention mass assigned to subthreshold contributors, sum over J_down of f-alpha, is small for improved models.
    The 'soft guarantee' (Inequality 6) needs this term near zero; Section 3 explicitly says the guarantee relies on it and that this may explain why some models do not improve.
  • domain assumption Meaningful vision-language alignment emerges in middle-to-late transformer layers, so final-layer intervention is the right place.
    The paper adopts this from Venhoff et al. [23] to justify operating on the final layer; it is a prior result, not proven here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Open-Vocabulary Referring Object Detection with Reverse Contrast Attention." pith.science (2026). https://pith.science/paper/KHY2PZN2

@misc{pith2026250719891,
  author       = {Pith},
  title        = {Pith review of: Interpretable Open-Vocabulary Referring Object Detection with Reverse Contrast Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHY2PZN2}},
  note         = {Machine review of arXiv:2507.19891}
}
abstract

We propose Reverse Contrast Attention (RCA), a plug-in method that enhances object localization in vision-language transformers without retraining. RCA reweights final-layer attention by suppressing extremes and amplifying mid-level activations to let semantically relevant but subdued tokens guide predictions. We evaluate it on Open Vocabulary Referring Object Detection (OV-RefOD), introducing FitAP, a confidence-free average precision metric based on IoU and box area. RCA improves FitAP in 11 out of 15 open-source VLMs, with gains up to $+26.6\%$. Effectiveness aligns with attention sharpness and fusion timing; while late-fusion models benefit consistently, models like $\texttt{DeepSeek-VL2}$ also improve, pointing to capacity and disentanglement as key factors. RCA offers both interpretability and performance gains for multimodal transformers. Codes and dataset are available from https://github.com/earl-juanico/rca

Figures

Figures reproduced from arXiv: 2507.19891 by the authors.

Figure 1
Figure 1. Illustration of the RCA mechanism. In a standard transformer layer, the hidden state zi is a vector that represents the token i as a superposition of the value vectors vj ∈ V of the text prompt and the image regions according to the attention distribution A = {αij} ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Selected examples suggestive of RCA’s positive impact on [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Selected examples from PaliGemma2-3B-mix-448 and valley2 dpo, qualitatively illustrating the observed improve￾ments in detection after RCA is applied: solid green boxes (ground truth); dashed red boxes (detections) tion, while it amplifies contributions closer to the midpoint. Thus, Ae is a more equalized attention profile, implying more stable and bounded hidden states from the final trans￾former layer. As extreme … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualizing the link between RCA and OV-RefOD in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Correlations between the number |S| of subthreshold contributions and mean cross-head max m of attention weights evaluated on 2064 (I, Q) pairs with Pearson correlations and p-value: (a) r = −0.09, p = 0.00004, (b) r = −0.73, p < 0.00001, (c) r = −0.02, p = 0.32. At th…
Figure 6
Figure 6. Figure 6: Modular vs. early fusion strategies in VLM [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    Generic attention- model explainability for interpreting bi-modal and encoder- decoder transformers

    Hila Chefer, Shir Gur, and Lior Wolf. Generic attention- model explainability for interpreting bi-modal and encoder- decoder transformers. In ICCV, pages 397–406, 2021. 1

    work page 2021

  2. [2]

    Transformer inter- pretability beyond attention visualization

    Hila Chefer, Shir Gur, and Lior Wolf. Transformer inter- pretability beyond attention visualization. In CVPR, pages 782–791, 2021. 2

    work page 2021

  3. [3]

    OvarNet: Towards open- vocabulary object attribute recognition

    Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. OvarNet: Towards open- vocabulary object attribute recognition. In CVPR, pages 23518–23527, 2023. 2

    work page 2023

  4. [4]

    Re- verse attention for salient object detection

    Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Re- verse attention for salient object detection. In ECCV, pages 234–250, 2018. 2

    work page 2018

  5. [5]

    Spatial- rgpt: Grounded spatial reasoning in vision-language models

    An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Rui- han Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial- rgpt: Grounded spatial reasoning in vision-language models. NeurIPS, 37:135062–135093, 2025. 2

    work page 2025

  6. [6]

    The Pascal Visual Ob- ject Classes (VOC) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Ob- ject Classes (VOC) challenge. IJCV, 88:303–338, 2010. 4

    work page 2010

  7. [7]

    Open-vocabulary object detection via vision and language knowledge distillation

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 4

    work page 2022

  8. [8]

    Attention-based multimodal fusion for video descrip- tion

    Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video descrip- tion. In ICCV, pages 4193–4202, 2017. 7

    work page 2017

Show all 36 references
  1. [9]

    Semantic segmentation with reverse attention

    Qin Huang, Chunyang Xia, Chihao Wu, Siyang Li, Ye Wang, Yuhang Song, and C-C Jay Kuo. Semantic segmentation with reverse attention. In BMVC, pages 18.1–18.13, 2017. 2

    2017

  2. [10]

    Scratching visual transformer’s back with uniform attention

    Nam Hyeon-Woo, Kim Yu-Ji, Byeongho Heo, Dongyoon Han, Seong Joon Oh, and Tae-Hyun Oh. Scratching visual transformer’s back with uniform attention. In ICCV, pages 5807–5818, 2023. 2

    2023

  3. [11]

    Attention is not expla- nation

    Sarthak Jain and Byron C Wallace. Attention is not expla- nation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543–3556, 2019. 1

    2019

  4. [12]

    MDETR- modulated detection for end-to-end multi-modal understand- ing

    Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR- modulated detection for end-to-end multi-modal understand- ing. In ICCV, pages 1780–1790, 2021. 4

    2021

  5. [13]

    Shallow and reverse attention network for colon polyp segmentation

    Go-Eun Lee, Jungchan Cho, and Sang-II Choi. Shallow and reverse attention network for colon polyp segmentation. Sci- entific Reports, 13(1):15243, 2023. 2

    2023

  6. [14]

    Grounded language-image pre-training

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jian- wei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965–10975,

  7. [15]

    Distilled re- verse attention network for open-world compositional zero- shot learning

    Yun Li, Zhe Liu, Saurav Jha, and Lina Yao. Distilled re- verse attention network for open-world compositional zero- shot learning. In ICCV, pages 1782–1791, 2023. 2

    2023

  8. [16]

    Rta-former: Reverse transformer attention for polyp segmen- tation

    Zhikai Li, Murong Yi, Ali Uneri, Sihan Niu, and Craig Jones. Rta-former: Reverse transformer attention for polyp segmen- tation. In 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages 1–5. IEEE, 2024. 2

    2024

  9. [17]

    Attention is not enough: Mitigating the distribution discrepancy in asynchronous multimodal sequence fusion

    Tao Liang, Guosheng Lin, Lei Feng, Yan Zhang, and Feng- mao Lv. Attention is not enough: Mitigating the distribution discrepancy in asynchronous multimodal sequence fusion. In ICCV, pages 8148–8156, 2021. 7

    2021

  10. [18]

    Faster R-CNN: Towards real-time object detection with re- gion proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with re- gion proposal networks. NeurIPS, 28, 2015. 4

    2015

  11. [19]

    Self- attention with relative position representations

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self- attention with relative position representations. In Proceed- ings of the 2018 Conference of the North American Chap- ter of the Association for Computational Linguistics: Hu- man Language Technologies, Volume 2 (Short Pap...

    2018

  12. [20]

    GroundVLP: Harnessing zero-shot visual grounding from vision-language pre-training and open-vocabulary ob- ject detection

    Haozhan Shen, Tiancheng Zhao, Mingwei Zhu, and Jianwei Yin. GroundVLP: Harnessing zero-shot visual grounding from vision-language pre-training and open-vocabulary ob- ject detection. In AAAI, pages 4766–4775, 2024. 2

    2024

  13. [21]

    Zero-shot open-vocabulary OOD object detection and grounding using vision language mod- els

    Poulami Sinhamahapatra, Shirsha Bose, Karsten Roscher, and Stephan G ¨unnemann. Zero-shot open-vocabulary OOD object detection and grounding using vision language mod- els. In Northern Lights Deep Learning Conference , pages 230–238. PMLR, 2025. 2

    2025

  14. [22]

    Reverse and boundary attention net- work for road segmentation

    Jee-Young Sun, Seung-Wook Kim, Sang-Won Lee, Ye-Won Kim, and Sung-Jea Ko. Reverse and boundary attention net- work for road segmentation. In ICCV, pages 0–0, 2019. 2

    2019

  15. [23]

    How visual representations map to language feature space in multimodal llms

    Constantin Venhoff, Ashkan Khakzar, Sonia Joseph, Philip Torr, and Neel Nanda. How visual representations map to language feature space in multimodal llms. arXiv preprint arXiv:2506.11976, 2025. 2, 7

    2025 arXiv

  16. [24]

    OV-VG: A benchmark for open-vocabulary visual grounding

    Chunlei Wang, Wenquan Feng, Xiangtai Li, Guangliang Cheng, Shuchang Lyu, Binghao Liu, Lijiang Chen, and Qi Zhao. OV-VG: A benchmark for open-vocabulary visual grounding. Neurocomputing, 591:127738, 2024. 2

    2024

  17. [25]

    Ra-net: reverse attention for generalizing residual learning

    Zhenyuan Wang, Xuemei Xie, Jianxiu Yang, and Xiaodan Song. Ra-net: reverse attention for generalizing residual learning. Scientific Reports, 14(1):12771, 2024. 2

    2024

  18. [26]

    Attention is not not expla- nation

    Sarah Wiegreffe and Yuval Pinter. Attention is not not expla- nation. In 2019 Conference on Empirical Methods in Nat- ural Language Processing and 9th International Joint Con- ference on Natural Language Processing, EMNLP-IJCNLP 2019, pages 11–20. Association for Computational...

    2019

  19. [27]

    Valley2: Exploring multimodal models with scalable vision- language design

    Ziheng Wu, Zhenghao Chen, Ruipu Luo, Can Zhang, Yuan Gao, Zhentao He, Xian Wang, Haoran Lin, and Minghui Qiu. Valley2: Exploring multimodal models with scalable vision- language design. arXiv preprint arXiv:2501.05901, 2025. 8

    2025 arXiv

  20. [28]

    Im- age inpainting with learnable bidirectional attention maps

    Chaohao Xie, Shaohui Liu, Chao Li, Ming-Ming Cheng, Wangmeng Zuo, Xiao Liu, Shilei Wen, and Errui Ding. Im- age inpainting with learnable bidirectional attention maps. In ICCV, pages 8858–8867, 2019. 2

    2019

  21. [29]

    WeThink: To- ward general-purpose vision-language reasoning via rein- forcement learning

    Jie Yang, Feipeng Ma, Zitian Wang, Dacheng Yin, Kang Rong, Fengyun Rao, and Ruimao Zhang. WeThink: To- ward general-purpose vision-language reasoning via rein- forcement learning. arXiv preprint arXiv:2506.07905, 2025. 4, 8

    2025 arXiv

  22. [30]

    RLAIF-V: Open-source AI feedback leads to su- per GPT-4V trustworthiness

    Tianyu Yu, Haoye Zhang, Qiming Li, Qixin Xu, Yuan Yao, Da Chen, Xiaoman Lu, Ganqu Cui, Yunkai Dang, Taiwen He, et al. RLAIF-V: Open-source AI feedback leads to su- per GPT-4V trustworthiness. In CVPR, pages 19985–19995,

  23. [31]

    Open-vocabulary object detection using captions

    Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih- Fu Chang. Open-vocabulary object detection using captions. In CVPR, pages 14393–14402, 2021. 4

    2021

  24. [32]

    LED: LLM enhanced open-vocabulary object detection without human curated data generation

    Yang Zhou, Shiyu Zhao, Yuxiao Chen, Zhenting Wang, and Dimitris N Metaxas. LED: LLM enhanced open-vocabulary object detection without human curated data generation. arXiv preprint arXiv:2503.13794, 2025. 2 Interpretable Open-V ocabulary Referring Object Detection with Reverse ...

    2025

  25. [33]

    The tendency of VLM’s predictions to maintain propor- tional sizing with the actual object in the image, and

  26. [34]

    Thus, we examine the correlation between the area of ground truths (normalized to image size) and those of de- tection Abox

    influence of the actual object’s size on the detection ac- curacy. Thus, we examine the correlation between the area of ground truths (normalized to image size) and those of de- tection Abox. Our results (Figure S7) confirm that this cor- relation is strong (Pearson r = 0.90),...

  27. [35]

    m ↑ =⇒ fαij assigns less weight to j ∈ J↓

  28. [36]

    In the first case of inverse-distance reweighting: α′ ij = 1 1 + γ|αij − m| , which peaks at αij = m and decreases as αij deviates from m

    This assertion holds for both RCA schemes: • Inverse-distance from m • Gausian peaking around m Therefore, the penalty term in condition (6) shrinks with increasing m, which increases ezi(d), decreasing the sub- threshold count. In the first case of inverse-distance reweightin...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.