Align and Surpass Human Camouflaged Perception: Visual Refocus Reinforcement Fine-Tuning

Haihua Yang; Jiangning Zhang; Kai Wu; Ruolin Shen; Xiaobin Hu; Xiaoyu Sun; Xiaozhong Ji; Yijun He

REVIEW 4 major objections 7 minor 1 cited by

Align and Surpass Human Camouflaged Perception: Visual Refocus Reinforcement Fine-Tuning

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A vision-language model trained with a staged reinforcement curriculum learns to 'refocus' on camouflaged objects and, on a deliberately hard test set, beats supervised fine-tuning and a ten-second human baseline.

desk verdict Plausible RL-for-VLM method applied to camouflage, but the headline empirical claims are staked on an underspecified test set and a tilted human comparison. read the letter →

arxiv 2505.19611 v1 pith:6ZGLAVOK submitted 2025-05-26 cs.CV cs.AI

Ruolin Shen , Xiaozhong Ji , Kai WU , Jiangning Zhang , Yijun He , HaiHua Yang , Xiaobin Hu , Xiaoyu Sun This is my paper

classification cs.CVcs.AI

keywords camouflagedobjectdetectionvision-languagemodelreinforcementfine-tuningvisualrefocuscurriculumlearningrule-basedrewardclassificationhumanperception

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 claims that current vision-language models lack a human visual skill: they cannot reliably see objects that blend into the background. It presents Visual Refocus Reinforcement Fine-Tuning, a training scheme that rewards a model first for producing correctly formatted answers about whether a concealed object exists, then for naming its category, then for drawing a tight bounding box around it, with in-context examples of stepwise 'refocus' reasoning. On camouflaged-object classification and detection benchmarks, the trained model outperforms supervised fine-tuning on the same base model, and on a hand-picked hard subset it reports higher accuracy than human participants who were given ten seconds per image. If the result holds, it suggests that a reward curriculum can elicit human-like visual search behavior in vision-language models rather than only improving their text reasoning.

What carries the argument

The machinery is a reinforcement-learning curriculum built on group-relative policy optimization, a policy-gradient update that scores each sampled answer relative to the group average rather than against an absolute reward. The curriculum adds reward components in three stages—format and presence accuracy, category correctness, and finally bounding-box intersection-over-union—so the model masters output structure and basic detection before fine localization. Exploration is widened by in-context demonstration trajectories that show multi-stage visual refocusing and by a clip-high objective that removes the KL penalty and loosens the upper clipping bound, letting the model try less likely reasoning paths. This combination is what the paper credits for the emergent refocusing behavior and for the performance gap over supervised fine-tuning.

What would settle it

Compute perceptual hashes of every image in the hard concealed test set and every image in the 14,017 training samples and look for exact or near-duplicate matches; if matches exist, or if re-running the human study without the ten-second time cap reverses the ordering on clean images, the central claims would be refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that reinforcement learning, not more supervised data, is what lets a vision-language model align with human camouflaged perception. The trained model develops three observable refocus patterns: 'focus' (zooming from a global region to a local one), 'rethink' (adjusting a candidate box after seeing a similar distractor), and 'backtracing' (jumping to a distinctive local feature like a head or wing and then zooming back out). On the hard concealed test set, the method reaches 0.923 binary presence accuracy, 0.80 category accuracy, and 0.473 detection mIoU (mean intersection-over-union), while the supervised fine-tuning baseline reaches 0.830, 0.70, and 0.365, and human participants reach 0.654 binary accuracy and 0.382 mIoU under the ten-second constraint.

Load-bearing premise

The whole comparison, including the claim of surpassing humans, rests on the assumption that the manually selected hard test images are clean of overlap with the 14,017 training images and were not cherry-picked to flatter the method; the paper states no selection protocol and does not say whether the hard set overlaps the training set.

Editorial extensions

If this is right

On the hard concealed set the gain of VRRF over supervised fine-tuning is much larger than on the easy set, so harder perception tasks benefit more from the reasoning curriculum.
Because each reward stage adds measurable improvement over the previous one, and because adding all rewards at once is much worse, the ordering of the curriculum is itself part of the result.
The model outputs intermediate bounding boxes and explanations for them, which means the refocus behavior is observable and inspectable, not just a final accuracy number.
If the user-study result generalizes, a deployed vision-language model can find concealed objects in time-limited settings better than an average human viewer, which matters for search-and-rescue and similar applications.
The same base model without reinforcement, and a much larger variant, score far below VRRF on the hard set, so the improvement is not simply a matter of model scale.

Reading between the lines

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

Editorial inference: the human comparison is speed-limited by design; a separate study with unlimited viewing time or expert observers would test whether VRRF's advantage is about quick perceptual search or about absolute recognition ability.
Editorial inference: the same staged-reward recipe could be transferred to other low-signal visual tasks such as nodule detection in medical scans or defect finding in industrial images, since the paper only gestures at these applications in its broader-impact section.
Editorial inference: an independent replication on a pre-registered hard subset drawn from the same public datasets would clarify the contribution of the method relative to the choice of test images, since the current hard set is manually selected.

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.

Desk Editor's Note

read the letter

You should read this paper if you care about whether GRPO-style RL can improve VLM performance on a hard perception task. The method is a coherent extension of known ideas: a three-stage reward curriculum (format/accuracy, then category, then IoU), a clip-high objective without KL, and in-context demonstration of refocus trajectories. Those are genuine modifications, and the ablation in Table 3 shows each component adds something. Credit where due: the paper is honest that the base machinery comes from R1-V and DeepSeek-R1, and the observation that SFT on final answers does poorly on camouflaged detection is worth making.

Now the soft spots, in order of severity. The hard test set is the load-bearing wall. The paper says only that it was "manually selected from the full datasets" and never states how many images are in it, how they were chosen, or whether they overlap with the 14,017 training samples drawn from the same public COD datasets. That is not a minor omission. If any hard-set images were in training, the Table 2 gains relative to SFT and to humans are partly memorization, not policy learning. Manual selection also raises cherry-picking risk: the set could have been curated after seeing which images are hard for current models or for naive humans. The stress-test note is right. The human comparison is also shaky: a 10-second limit with no participant count or variance makes "surpasses human perception" a headline claim built on a strawman.

The "emergence" language is overblown. The model is given in-context demonstrations of refocusing; describing the resulting behavior as emergent is a stretch. That is a conceptual point, not a load-bearing flaw. The absence of error bars and code/data release is a standard problem, and the promised public artifacts do not count as shipped.

So the central argument—that this curriculum plus in-context RL helps on genuinely hard camouflaged cases—is plausible but not yet established. The paper deserves a real referee, but only if the authors can clarify the split protocol and release the test set or confirm disjointness. I would not cite the results until that is resolved. If the split survives contact with a referee, this becomes a useful data point for the RL-in-VLM literature.

Referee Report

4 major / 7 minor

Summary. The paper proposes Visual Refocus Reinforcement Fine-Tuning (VRRF), a GRPO-based reinforcement learning pipeline that fine-tunes Qwen2.5-VL-7B for camouflaged object classification and detection. The method combines a curriculum of rule-based rewards (format, accuracy, category, IoU) with in-context demonstrations of multi-step 'refocus' reasoning trajectories, and removes the KL penalty from GRPO in favor of a clip-high objective. Experiments on four COD datasets are reported on easy and hard test sets, with the central claims that VRRF significantly outperforms an SFT baseline and surpasses human participants in a user study on hard concealed cases.

Significance. If the empirical claims hold, the paper is a useful demonstration that rule-based reinforcement learning with process-oriented rewards can improve VLM performance on a visually demanding perception task and produce interpretable step-wise localization behaviors. The ablation study in Table 3 is informative, and the promise of public code and datasets is a strength. However, the headline results are currently staked on an underspecified test-set construction and an underspecified human study, so the significance of the claims cannot be fully assessed from the manuscript as written.

major comments (4)

[Sec. 4.1, Tables 1–2] The construction of the hard-concealed test set is not specified adequately: the paper states only that it was 'manually selected from the full datasets' and that the easy set was 'sampled images from the remaining data', but it does not report the number of hard or easy images, the selection criteria, or whether the hard set is disjoint from the 14,017 training samples drawn from the same four COD datasets. Because Tables 1 and 2 and the human comparison are all evaluated on these splits, any overlap or selection bias would invalidate the reported gains of VRRF over SFT and the 'surpass human' claim. Please provide the full split protocol, the split sizes, and a verified non-overlap check with the training set.
[Abstract and Table 2] The human user study is described only by a ten-second time limit; there is no participant count, no error bars or significance tests, no description of the images presented or the exact task instructions, and no justification for the time limit or for its claimed approximate equality to the model's average inference time. Since Table 2 reports 'Human perception' as a single point row (e.g., binary accuracy 0.654, detection mIoU 0.382), the headline claim that VRRF 'surpasses human camouflaged perception' is not supported as stated. Please report the full protocol, sample sizes, inter-participant variance, and a statistical comparison.
[Sec. 4.1, Tables 1–2] The SFT baseline is not described: the paper does not specify the SFT training data, number of epochs, learning rate, prompt format, or whether it uses the same 14,017 samples and the same base model. Because the central claim is 'significantly superior performance compared to Supervised Fine-Tuning (SFT) baselines', the comparison is not interpretable without this information. Please provide the SFT training details and confirm that the baseline was trained under matched conditions.
[Sec. 3.2, Fig. 3, Sec. 4.4] Calling the refocus behavior 'emergence' is an overstatement, because the in-context demonstrations in the prompt explicitly contain multi-stage refocus trajectories (focus, rethink, backtracing). The policy is therefore imitating provided demonstrations rather than spontaneously discovering this behavior. If the paper wants to claim emergence, it should either train without in-context demonstrations and show that the behavior still appears, or restrict the claim to the specific dynamic adjustments (e.g., backtracing patterns) that are not present in the demonstrations.

minor comments (7)

[Sec. 3.2, Sec. 4.1] The values of the clipping parameters epsilon and delta in Eq. (3) are not reported, even though clip-high is a key modification of the GRPO objective; please report them.
[Sec. 4.1] The easy set is described as 'sampled images from the remaining data', but the sampling method and the resulting sample sizes are not given; please clarify.
[Sec. 4.4] The manuscript uses informal expressions such as '4-th row' and '5-th row' when referring to Figure 5; please use 'fourth row' and 'fifth row' or equivalent formal phrasing.
[References and Table 2] Reference [15] is cited for GPT4.1 but points to the GPT-4 technical report; please clarify which model version was actually used and cite the corresponding source.
[Figure 1] The text inside Figure 1 appears compressed and difficult to read in the current rendering; please provide a cleaner version.
[Table 3] The row 'All-at-once (No Curriculum)' should explicitly state which reward components and whether in-context demonstrations are included, since the surrounding text says 'adding all components simultaneously' but the row label is ambiguous.
[Data availability] The paper states that implementation code and datasets will be made public but does not give a release plan or license information; a concrete data and code availability statement would be helpful.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: VRRF is a standard empirical RL pipeline whose gains are measured against held-out benchmarks; the main hard-test-set overlap concern is a validity risk, not a circularity.

full rationale

The paper's central claims are empirical: a Qwen2.5-VL-7B model is trained with a curriculum of rule-based rewards (format, accuracy, category, IoU; Eqs. 4-8) and evaluated on easy/hard subsets of four public COD datasets. No predicted quantity is defined in terms of a fitted parameter, and no equation reduces a reported result to its training input. The reward functions are standard label-based rewards, not fitted parameters renamed as predictions. The 'visual refocus' behavior is explicitly seeded by in-context demonstrations (Fig. 3 and Section 3.2), so describing it as 'emergent' is an overstatement, but this is a conceptual framing issue rather than circular reasoning because the model still must generalize beyond the demonstrations. The self-citation [29] is a co-author's prior COD network cited only in related work and is not load-bearing for the VRRF results. The most serious concern is the underspecified hard-concealed test set ('manually selected from the full datasets' with no stated protocol for disjointness from the 14,017 training samples), which threatens the validity of the SFT and human comparisons; however, potential data contamination or selection bias is a benchmark-integrity issue, not a circularity of the derivation. Overall, no step in the paper's argument reduces by construction to its own inputs, so circularity is minimal.

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

The central claim rests on the GRPO training framework and on the validity of the evaluation splits. The free parameters listed above are hand-chosen hyperparameters that the paper does not report numerically. The most consequential assumption is the fairness of the manually selected hard test set and the time-limited human baseline, both of which are ad hoc to this paper.

free parameters (3)

epsilon (GRPO lower clip bound) = not reported
Used in the clip-high objective (Eq. 3); chosen by hand but not specified numerically.
delta (GRPO upper clip bound) = not reported
Set larger than epsilon to encourage exploration; the paper states delta > epsilon but gives no value.
number of in-context examples per prompt = not reported
The prompt in Figure 3 shows multiple examples, but the exact count n is not stated, and the authors do not ablate it.

assumptions (4)

standard math GRPO objective as defined in Eq. (1) is a valid optimization target for this task.
The method builds directly on Group Relative Policy Optimization from DeepSeekMath, cited as [33].
domain assumption The four COD benchmark datasets provide reliable ground truth for presence, category, and bounding boxes.
The rewards depend on these labels; any labeling error propagates into the trained policy.
ad hoc to paper Manual selection of the hard test set produces a fair, representative difficulty split.
The paper does not describe the selection protocol or rule out overlap with the 14,017 training samples, yet this split drives the main performance claims.
ad hoc to paper A 10-second time limit in the user study is a fair basis for the 'surpass human' claim.
The time limit is chosen to approximate the model's inference time, but it may systematically disadvantage human participants, making the comparison favorable to the model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Align and Surpass Human Camouflaged Perception: Visual Refocus Reinforcement Fine-Tuning." pith.science (2026). https://pith.science/paper/6ZGLAVOK

@misc{pith2026250519611,
  author       = {Pith},
  title        = {Pith review of: Align and Surpass Human Camouflaged Perception: Visual Refocus Reinforcement Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZGLAVOK}},
  note         = {Machine review of arXiv:2505.19611}
}

read the original abstract

Current multi-modal models exhibit a notable misalignment with the human visual system when identifying objects that are visually assimilated into the background. Our observations reveal that these multi-modal models cannot distinguish concealed objects, demonstrating an inability to emulate human cognitive processes which effectively utilize foreground-background similarity principles for visual analysis. To analyze this hidden human-model visual thinking discrepancy, we build a visual system that mimicks human visual camouflaged perception to progressively and iteratively `refocus' visual concealed content. The refocus is a progressive guidance mechanism enabling models to logically localize objects in visual images through stepwise reasoning. The localization process of concealed objects requires hierarchical attention shifting with dynamic adjustment and refinement of prior cognitive knowledge. In this paper, we propose a visual refocus reinforcement framework via the policy optimization algorithm to encourage multi-modal models to think and refocus more before answering, and achieve excellent reasoning abilities to align and even surpass human camouflaged perception systems. Our extensive experiments on camouflaged perception successfully demonstrate the emergence of refocus visual phenomena, characterized by multiple reasoning tokens and dynamic adjustment of the detection box. Besides, experimental results on both camouflaged object classification and detection tasks exhibit significantly superior performance compared to Supervised Fine-Tuning (SFT) baselines.

Figures

Figures reproduced from arXiv: 2505.19611 by the authors.

**Figure 1.** Intriguing discovery of SOTA multi-modal models on limitation: these models struggle to replicate human cognitive processes in leveraging foreground-background similarity relationships for visual analysis. Mimicking human visual camouflaged reasoning perception, our Visual Refocus Reinforcement Fine-Tuning visual system progressively and logically ‘refocus’ visual concealed content. Abstract Current multi-modal mode… view at source ↗

**Figure 2.** Overview of Visual Refocus Reinforcement Fine-Tuning. • We collect the camouflaged dataset and build the evaluation systems to analyze the camouflaged perception ability. Through comprehensive experiments, we show that our approach to camouflaged object analysis achieves significantly better results than standard supervised fine-tuning (SFT) methods, especially on extremely challenging test set. Furthermore, empiri… view at source ↗

**Figure 3.** Prompt example used for in-context reinforcement learning. The [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Examples from our hard-concealed object set. Can you find them? Best viewed in color and zoomed-in. Datasets and Evaluation Metrics. We evaluate our proposed VRRF on four public benchmark datasets for Camouflaged Object Detection (COD): COD10K [35], NC4K [36], CAMO […

**Figure 5.** Figure 5: Illustration of “Visual Refocus” representation pattern. The first three rows show ‘focus’ in the form of global to local zoom-in, 4-th row denotes ‘backtracing’ from local to global extension retracing after perceiving the discriminate head and wing part, and 5-th row…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

OracleFusion: Assisting the Decipherment of Oracle Bone Script with Structurally Constrained Semantic Typography
cs.CV 2025-06 conditional novelty 6.0 of 10

OracleFusion adds structural constraints and MLLM-driven semantic analysis to vector typography, generating meaning-preserving illustrations of oracle bone script that outperform prior baselines.

Reference graph

Works this paper leans on

33 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Openai o1 system card

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

arXiv 2024
[3]

Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

Yifan Du, Zikang Liu, Yifan Li, Wayne Xin Zhao, Yuqi Huo, Bingning Wang, Weipeng Chen, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

arXiv 2025
[5]

Seg-zero: Reasoning- chain guided segmentation via cognitive reinforcement

Yuqi Liu, Bohao Peng, Zhisheng Zhong, Zihao Yue, Fanbin Lu, Bei Yu, and Jiaya Jia. Seg-zero: Reasoning- chain guided segmentation via cognitive reinforcement. arXiv preprint arXiv:2503.06520, 2025

arXiv 2025
[6]

Vision-r1: Evolving human-free alignment in large vision-language models via vision-guided reinforcement learning

Yufei Zhan, Yousong Zhu, Shurong Zheng, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Vision-r1: Evolving human-free alignment in large vision-language models via vision-guided reinforcement learning. arXiv preprint arXiv:2503.18013, 2025

arXiv 2025
[7]

Boosting the generaliza- tion and reasoning of vision language models with curriculum reinforcement learning

Huilin Deng, Ding Zou, Rui Ma, Hongchen Luo, Yang Cao, and Yu Kang. Boosting the generaliza- tion and reasoning of vision language models with curriculum reinforcement learning. arXiv preprint arXiv:2503.07065, 2025

arXiv 2025
[8]

Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl

Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536, 2025

arXiv 2025
[9]

Visual-rft: Visual reinforcement fine-tuning

Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025

arXiv 2025
[11]

R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization

Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization. arXiv preprint arXiv:2503.12937, 2025

arXiv 2025

Show all 33 references

[12]

Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-improvement

Yihe Deng, Hritik Bansal, Fan Yin, Nanyun Peng, Wei Wang, and Kai-Wei Chang. Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-improvement. arXiv preprint arXiv:2503.17352, 2025

2025 arXiv
[13]

Flamingo: a visual language model for few-shot learning

Jean-Baptiste Alayrac et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, volume 35, pages 23716–23736, 2022

2022
[14]

Visual instruction tuning

Haotian Liu et al. Visual instruction tuning. In Advances in Neural Information Processing Systems , volume 36, 2024

2024
[15]

Gpt-4 technical report

Josh Achiam et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

2023 arXiv
[16]

Instructblip: Towards general-purpose vision-language models with instruction tuning

Wenliang Dai et al. Instructblip: Towards general-purpose vision-language models with instruction tuning. In Proceedings of the Conference on Neural Information Processing Systems, 2023

2023
[17]

Dragonfly: Multi-resolution zoom supercharges large visual-language model

Kezhen Chen et al. Dragonfly: Multi-resolution zoom supercharges large visual-language model. arXiv preprint arXiv:2406.00977, 2024

2024 arXiv
[18]

How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites

Zhe Chen et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 10

2024 arXiv
[19]

Llava-onevision: Easy visual task transfer

Bo Li et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024

2024 arXiv
[21]

Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

Peng Wang et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024

2024 arXiv
[22]

Expanding performance boundaries of open-source multimodal models with model, data, and testtime scaling

Zhe Chen et al. Expanding performance boundaries of open-source multimodal models with model, data, and testtime scaling. arXiv preprint arXiv:2412.05271, 2024

2024 arXiv
[23]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

Daya Guo et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

2025 arXiv
[24]

R1-onevision: Advancing generalized multimodal reasoning through cross-modal formaliza- tion

Yi Yang et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formaliza- tion. arXiv preprint arXiv:2503.10615, 2025

2025 arXiv
[25]

R1-v: Reinforcing super generalization ability in visionlanguage models with less than $3

Liang Chen et al. R1-v: Reinforcing super generalization ability in visionlanguage models with less than $3. https://github.com/Deep-Agent/R1-V , 2025. Accessed: 2025-02-02

2025
[27]

R1-zero’s” aha moment” in visual reasoning on a 2b non-sft model

Hengguang Zhou et al. R1-zero’s” aha moment” in visual reasoning on a 2b non-sft model. arXiv preprint arXiv:2503.05132, 2025

2025 arXiv
[28]

Concealed object detection

Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE TPAMI, 2021

2021
[29]

High-resolution iterative feedback network for camouflaged object detection

Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 881–889, 2023

2023
[30]

Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection

Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8805–8814, 2024

2024
[31]

Camouflage

IC Cuthill. Camouflage. JOZ, 308(2):75–92, 2019

2019
[32]

Animal camouflage: current issues and new perspectives

Martin Stevens and Sami Merilaita. Animal camouflage: current issues and new perspectives. PTR:BS, 364(1516):423–427, 2009

2009
[33]

Deepseekmath: Pushing the limits of mathematical reasoning in open language models

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

2024 arXiv
[34]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

2025 arXiv
[35]

Camouflaged object detection

Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In CVPR, 2020

2020
[36]

Simultane- ously localize, segment and rank the camouflaged objects

Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultane- ously localize, segment and rank the camouflaged objects. In CVPR, 2021

2021
[37]

Anabranch network for camouflaged object segmentation

Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. CVIU, 184:45–56, 2019

2019
[38]

Animal camouflage analysis: Chameleon database

P Skurowski, H Abdulameer, J Błaszczyk, T Depta, A Kornacki, and P Kozieł. Animal camouflage analysis: Chameleon database. Unpublished Manuscript, 2018. 11

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Openai o1 system card

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

arXiv 2024

[2] [3]

Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

Yifan Du, Zikang Liu, Yifan Li, Wayne Xin Zhao, Yuqi Huo, Bingning Wang, Weipeng Chen, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

arXiv 2025

[3] [5]

Seg-zero: Reasoning- chain guided segmentation via cognitive reinforcement

Yuqi Liu, Bohao Peng, Zhisheng Zhong, Zihao Yue, Fanbin Lu, Bei Yu, and Jiaya Jia. Seg-zero: Reasoning- chain guided segmentation via cognitive reinforcement. arXiv preprint arXiv:2503.06520, 2025

arXiv 2025

[4] [6]

Vision-r1: Evolving human-free alignment in large vision-language models via vision-guided reinforcement learning

Yufei Zhan, Yousong Zhu, Shurong Zheng, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Vision-r1: Evolving human-free alignment in large vision-language models via vision-guided reinforcement learning. arXiv preprint arXiv:2503.18013, 2025

arXiv 2025

[5] [7]

Boosting the generaliza- tion and reasoning of vision language models with curriculum reinforcement learning

Huilin Deng, Ding Zou, Rui Ma, Hongchen Luo, Yang Cao, and Yu Kang. Boosting the generaliza- tion and reasoning of vision language models with curriculum reinforcement learning. arXiv preprint arXiv:2503.07065, 2025

arXiv 2025

[6] [8]

Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl

Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536, 2025

arXiv 2025

[7] [9]

Visual-rft: Visual reinforcement fine-tuning

Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025

arXiv 2025

[8] [11]

R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization

Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization. arXiv preprint arXiv:2503.12937, 2025

arXiv 2025