Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

Omnidirectional Spatial Modeling from Correlated Panoramas

T0 review · 3 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that Pano-R1, a vision-language model fine-tuned with Group Relative Policy Optimization and a three-part reward, achieves state-of-the-art performance on CFpano, the first benchmark for question-answering across correlated

desk verdict CFpano is a genuinely new cross-frame panoramic VQA dataset, but the paper's headline claims overreach the table, and the benchmark's Kimi-generated labels need human verification before the SOTA ranking means anything. read the letter →

arxiv 2509.02164 v1 pith:DBWFVLBR submitted 2025-09-02 cs.CV

Xinshen Zhang , Tongxi Fu , Xu Zheng This is my paper
classification cs.CV
keywords omnidirectionalvisionpanoramicvisualquestionansweringcross-framereasoning360°sceneunderstandinggrouprelativepolicyoptimizationmulti-modallargelanguagemodelsreinforcementlearningbenchmarkdataset
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 sets out to close a gap in 360° scene understanding: existing panoramic datasets and models mostly reason about a single frame, while real observers — embodied agents, drivers, VR users — see a scene through many related viewpoints. To that end it introduces CFpano, which it presents as the first benchmark for cross-frame panoramic visual question answering, with 8,094 questions over more than 2,700 panorama frames built on 3D-annotated ReplicaPano scenes. It then introduces Pano-R1, a 7B vision-language model fine-tuned with Group Relative Policy Optimization under a composite reward that demands well-formatted, accurate, and internally consistent answers. The paper reports that Pano-R1 beats ten existing MLLMs across all major reasoning categories, +5.37% overall on multiple-choice accuracy, and that the reinforcement-learning phase alone adds about four points over supervised fine-tuning. If these results hold, the paper supplies a reusable evaluation instrument for multi-view panoramic reasoning and evidence that rule-based RL rewards generalize beyond math and coding into spatial multimodal understanding.

What carries the argument

The load-bearing mechanism is the composite reward function used inside GRPO: R = R_format · sqrt(clip(R_answer,0,1) × clip(R_consistency,0,1)). It couples three scoring signals so that no response earns a high reward unless it is structurally well-formed (both <think> and <answer> tags present), factually correct (exact option match for multiple choice; embedding cosine similarity for open-ended answers), and internally coherent (the reasoning must mention the chosen option or be semantically close to the final answer). Because the score feeds GRPO's advantage normalization, the policy is pushed toward responses that are simultaneously parseable, right, and self-consistent. Around this, the

What would settle it

Have several independent human annotators verify the ground-truth answers on a few hundred randomly sampled CFpano test questions, then recompute the model rankings on only the questions whose labels pass verification. If a substantial share of labels are judged wrong, or if Pano-R1's lead over the strongest baselines shrinks or reverses on the verified subset, the state-of-the-art claim falls.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is twofold. First, CFpano — 8,094 question-answer pairs, 64.74% multiple-choice — generated from ReplicaPano scenes through a two-stage pipeline: a vision model writes fine-grained captions from 3D annotations, then fills structured question templates on pairs of frames whose IDs differ by at most 20. Second, Pano-R1, built on a 7B Qwen2.5-VL base with LoRA: supervised fine-tuning on 6,475 training questions followed by GRPO reinforcement under the reward R = R_format · sqrt(clip(R_answer,0,1) × clip(R_consistency,0,1)). Format checks that the output contains <think> and <answer> tags; accuracy is exact option match for multiple choice and emb

Load-bearing premise

The benchmark's ground-truth answers were produced by the Kimi vision model from machine-written captions, with human annotators performing only random sampling checks — so if those auto-generated labels are systematically wrong or biased, every model ranking on CFpano, including the claimed state of the art, is a ranking against unreliable answers.

Editorial extensions

If this is right

  • CFpano gives the field a reusable testbed: any MLLM can be scored on multi-view 360° reasoning, with category-level breakdowns (spatial relations, occlusion, counting, perspective) that localize where models fail.
  • GRPO with rule-based rewards transfers from math and coding to spatial multimodal reasoning: the RL phase alone raised MCQ accuracy from 52.6% to 56.8% over supervised fine-tuning, with the largest jumps in Image Characteristics and Perspective Question Design.
  • Scale matters for selection-style questions: the 7B model beats the 3B variant by about 7 MCQ points overall, while open-ended answer quality is closer, suggesting bigger models buy accuracy more than fluency.
  • All ten zero-shot baselines land below 52% MCQ, so cross-frame panoramic reasoning remains largely unsolved for off-the-shelf MLLMs.
  • The consistency reward's chunked embedding similarity offers a reusable, model-agnostic way to reward reasoning that matches the answer, applicable to other multi-frame or long-context VQA settings.

Reading between the lines

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

  • Editorial inference: because the same vision model that wrote the captions also generated the ground-truth answers, with only random human spot-checks, the benchmark may inherit that model's perceptual blind spots; a fully human-verified subset would tell whether Pano-R1's lead is real or partly an artifact of label bias.
  • Editorial inference: the +5.37% gain may concentrate in template families whose answers follow straightforwardly from the 3D metadata; cross-referencing the appendix's sub-category tables against template predictability would reveal whether the model learned spatial reasoning or template heuristics.
  • Editorial inference: a direct stress test would be to present the same questions with the frame pair order swapped, or with one frame replaced by a frame from a different scene; if accuracy does not drop correspondingly, the model is relying on caption or template priors rather than on the actual image pair.
  • Editorial inference: the same reward triplet (format, accuracy, consistency) should transfer to other correlated-view settings such as multi-camera driving footage or video VQA without modification.
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 / 7 minor

Summary. The paper introduces CFpano, a VQA benchmark for cross-frame correlated 360-degree panoramas, built from the ReplicaPano dataset and containing 8,094 question-answer pairs (64.74% multiple-choice, 35.26% open-ended) across five reasoning categories. It then proposes Pano-R1, a Qwen2.5-VL model fine-tuned with GRPO and three reward functions: format reward, accuracy reward, and consistency reward. The authors report state-of-the-art performance on CFpano, with an overall MCQ accuracy of 56.78% and QA similarity of 0.8316, and claim to outperform strong baselines on all major reasoning categories. Ablations compare SFT-only versus SFT+RL training and 3B versus 7B model sizes.

Significance. If the claims are established, the paper would provide a useful new benchmark for cross-frame panoramic VQA and evidence that GRPO-style RL with tailored rewards can improve multi-view 360-degree reasoning. The use of ReplicaPano 3D annotations for grounded question/caption generation is a constructive design choice, and the category/sub-category taxonomy enables fine-grained evaluation. However, the central SOTA claim is currently not supported by the paper's own table, the open-ended QA metric is partly circular with the training reward, and the ground-truth labels are AI-generated with only cursory human checks. These are not mere presentation issues; they affect the validity of the headline results. With correction of the claims and additional verification, the underlying benchmark and method could be valuable to the community.

major comments (3)
  1. [Abstract and §5.2, Table 1] The abstract and conclusion claim that Pano-R1 'outperforms strong baselines on all major reasoning categories' and achieves '+5.37% in overall performance.' Table 1 contradicts this: on Basic Understanding MCQ, Pano-R1 scores 42.13% versus 55.56% for InternVL3-8B, and on Quantitative Reasoning MCQ, Pano-R1 scores 42.80% versus 47.03% for Qwen2.5-Omni-7B. The main text itself acknowledges (last paragraph of §5.2) that the model does not lead in every metric. The global SOTA claim is therefore overstated and must be revised to a per-category, per-metric statement.
  2. [§4.2, Eqs. (6), (8), and §5.1] The open-ended QA evaluation is not independent of the training reward. The QA accuracy reward (Eq. 6) and QA consistency reward (Eq. 8) both use cosine similarity with the all-MiniLM-L6-v2 SentenceTransformer embedding, and §5.1 states that evaluation computes semantic similarity with the same all-MiniLM-L6-v2 model. Consequently, the reported QA improvements (e.g., 0.8316 versus 0.7967 for Qwen2.5-VL-3B) reflect, at least in part, direct optimization of the evaluation metric. The claim that Pano-R1 achieves state-of-the-art open-ended VQA ability is therefore not established. The authors should evaluate with a different embedding model, a third-party judge, or human evaluation.
  3. [§3.3] The ground-truth answers in CFpano are generated by the closed Kimi vision model (moonshot-v1-8k-vision-preview) from automatically written captions, with quality control described only as 'human annotators perform random sampling checks.' No sample size, number of annotators, inter-annotator agreement, or correction outcome is reported. Given that the benchmark rankings, including the claimed +5.37% advantage, are computed against these labels, the correctness of the labels is load-bearing. The authors must provide a concrete verification protocol: either full human verification of all 8,094 answers or a statistically justified sample with detailed agreement and correction statistics, plus release of the annotation instructions.
minor comments (7)
  1. [Abstract] The phrase 'the first benchmark dataset dedicated to cross-frame correlated panoramas' should be qualified as 'the first cross-frame correlated panoramic VQA benchmark,' since earlier panoramic VQA datasets exist (VQA360, Pano-AVQA, OmniVQA, OSR-Bench).
  2. [§5.3.1] Typo: 'our ptraining pipeline' should be 'our training pipeline.'
  3. [Throughout] Inconsistent naming: CFpano vs. CFPano, ReplicaPano vs. Replicapano, Qwen2.5VL vs. Qwen2.5-VL. Please standardize.
  4. [Table 20 and §3.2] The 'Flex' question type is used in Table 20 but is not defined in the text. Clarify whether this means the template can be instantiated as either MCQ or QA.
  5. [§5.1] GRPO hyperparameters are not fully specified: the sampling group size N, KL coefficient beta, and clip range epsilon appear in Eqs. (2)-(3) but are never given. LoRA rank/alpha are provided, but the target modules are not. These details are needed for reproducibility.
  6. [Figure 1] The radar chart normalizes QA similarity from 0.5-1.0 to 0-100%, but the normalization is not described in the figure or main text in a way that permits interpretation. Add a note on the original scale and the normalization formula.
  7. [References] References [33] and [34] are the same paper (arXiv:2505.14197). This should be consolidated or renumbered.

Circularity Check

1 steps flagged · score 6.0 of 10

Open-ended QA results are partially circular: the RL reward and the evaluation metric are the same all-MiniLM-L6-v2 cosine similarity.

  1. fitted input called prediction [Eq. 6 and Eq. 8 (§4.2) vs. §5.1 Experimental Setup]
    "For QA, the reward uses cosine similarity of embeddings... R^QA_answer = cosine_sim(f(â), f(a*)) ... For QA, it computes the semantic similarity between the reasoning text (from <think>) and the answer (from <answer>) using embeddings ... R_consistency = max_{c in Chunks(r)} cosine_sim(f(c), f(â)) ... [§5.1] For QA, we compute semantic similarity using Sentence-BERT (all-MiniLM-L6-v2)."

    The open-ended QA evaluation metric in §5.1 is the same cosine similarity over the same all-MiniLM-L6-v2 embedding that defines the RL accuracy reward (Eq. 6) and consistency reward (Eq. 8). Pano-R1 is trained with GRPO to maximize exactly this function (combined in Eq. 9), so the reported QA scores (e.g., 0.8316 in Table 1) are not an independent measurement of open-ended VQA ability; they are, by construction, an optimization of the scoring function itself. This is a fitted-input-called-prediction pattern: the model is fit to the evaluation metric and the 'prediction' is the value of that same metric.

full rationale

The paper's MCQ results and the dataset resource itself are not circular: MCQ accuracy is exact match, independent of the embedding choice, and CFpano is constructed from ReplicaPano 3D annotations plus template-based questions. However, the open-ended QA results are partially circular. Eq. 6 (accuracy reward) and Eq. 8 (consistency reward) use cosine similarity with a SentenceTransformer embedding, and §5.1 evaluates QA with the same Sentence-BERT all-MiniLM-L6-v2 cosine similarity. Since Pano-R1 is RL-trained to maximize this composite reward (Eq. 9), the reported QA improvements are partly a measurement of how well the model optimizes the evaluation function, not an independent estimate of open-ended VQA quality. This is a fitted-input-called-prediction overlap. Separately, the benchmark's ground-truth answers are Kimi-generated with only 'random sampling checks,' which is a label-validity threat but not a circularity in the derivation. No load-bearing self-citation or imported uniqueness theorem was found; the duplicated 360-R1 self-reference [33][34] is not used to justify any central step.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the reliability of AI-generated labels and on embedding similarity as a correctness measure, neither of which is independently validated in the paper. The paper introduces no new particle, force, dimension, or mathematical object; CFpano and Pano-R1 are artifacts, so the invented-entities ledger is empty.

free parameters (3)
  • GRPO sampling group size N and KL/clip hyperparameters (beta, epsilon) = Not reported
    These control the RL training in §4.1; without them the training is not reproducible and the reported results cannot be audited.
  • Frame-pair sampling window = Delta ID <= 20
    Defines correlated frames in §3.3; no ablation shows how the cutoff changes dataset difficulty or results.
  • QA scoring embedding model = all-MiniLM-L6-v2 (Sentence-BERT)
    The open-ended QA metric and the QA reward use this same embedding (§5.1, Eq. 6); QA scores are defined relative to this choice, and different embeddings would shift numbers.
assumptions (5)
  • domain assumption ReplicaPano's 3D annotations are complete and accurate enough to support generated VQA ground truth
    CFpano labels are derived from ReplicaPano's depth, layout, object boxes, and meshes (§3). If these annotations are noisy, the benchmark inherits the noise.
  • domain assumption Kimi vision model (moonshot-v1-8k-vision-preview) generates correct captions and answers
    All fine-grained captions and QA answers are produced by this closed model (§3.3); correctness is only spot-checked by human annotators, with no reported error rate.
  • domain assumption Random sampling checks by human annotators guarantee dataset quality
    The paper states that human annotators perform random sampling checks but does not report how many items were checked, inter-annotator agreement, or the correction rate (§3.3).
  • domain assumption Sentence-BERT cosine similarity measures answer correctness for open-ended VQA
    QA evaluation and the QA reward both assume embedding cosine similarity tracks semantic correctness; no human correlation study is provided (§5.1, Eq. 6).
  • standard math GRPO objective (Eq. 2) is the standard estimator and is correctly applied to the MLLM
    The training relies on the DeepSeekMath GRPO formulation (§4.1, ref [24]); this is a standard result, but the specific implementation is not verifiable without code.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Omnidirectional Spatial Modeling from Correlated Panoramas." pith.science (2026). https://pith.science/paper/DBWFVLBR

@misc{pith2026250902164,
  author       = {Pith},
  title        = {Pith review of: Omnidirectional Spatial Modeling from Correlated Panoramas},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBWFVLBR}},
  note         = {Machine review of arXiv:2509.02164}
}
read the original abstract

Omnidirectional scene understanding is vital for various downstream applications, such as embodied AI, autonomous driving, and immersive environments, yet remains challenging due to geometric distortion and complex spatial relations in 360{\deg} imagery. Existing omnidirectional methods achieve scene understanding within a single frame while neglecting cross-frame correlated panoramas. To bridge this gap, we introduce \textbf{CFpano}, the \textbf{first} benchmark dataset dedicated to cross-frame correlated panoramas visual question answering in the holistic 360{\deg} scenes. CFpano consists of over 2700 images together with over 8000 question-answer pairs, and the question types include both multiple choice and open-ended VQA. Building upon our CFpano, we further present \methodname, a multi-modal large language model (MLLM) fine-tuned with Group Relative Policy Optimization (GRPO) and a set of tailored reward functions for robust and consistent reasoning with cross-frame correlated panoramas. Benchmark experiments with existing MLLMs are conducted with our CFpano. The experimental results demonstrate that \methodname achieves state-of-the-art performance across both multiple-choice and open-ended VQA tasks, outperforming strong baselines on all major reasoning categories (\textbf{+5.37\%} in overall performance). Our analyses validate the effectiveness of GRPO and establish a new benchmark for panoramic scene understanding.

Figures

Figures reproduced from arXiv: 2509.02164 by the authors.

Figure 1
Figure 1. The radar chart comparing MCQ Accuracy (%) and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Hierarchical structure and question distribution of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall framework for dataset construction and Reinforcement Learning training of the Pano-R1 model. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The first case of model inference between our model (Pano-R1) and other leading MLLMs. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The second case of model inference between our model (Pano-R1) and other leading MLLMs. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The third case of model inference between our model (Pano-R1) and other leading MLLMs [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. PanoWorld: Towards Spatial Supersensing in 360$^\circ$ Panorama World

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    PanoWorld adds spherical geometry to MLLMs via cross-attention and pano-specific instruction data, yielding better performance on panoramic spatial reasoning benchmarks than standard perspective-based pipelines.

  2. PanoWorld: Towards Spatial Supersensing in 360$^\circ$ Panorama World

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    PanoWorld adds spherical spatial cross-attention and pano-native training data to MLLMs for improved spatial reasoning on ERP panoramas, outperforming baselines on new and existing benchmarks.

Reference graph

Works this paper leans on

119 extracted references · 56 canonical work pages · cited by 1 Pith paper

  1. [1]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-VL Technical Rep...

    arXiv 2025

  2. [2]

    Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. 2024. The Revolution of Multimodal Large Language Models: A Survey. In Findings of the Association for Computational Linguistics: ACL 2024 . Association for Computational Linguistics, Bangkok, Thailand, 13590–13618. doi...

    doi:10.18653/v1/2024.findings-acl.807 2024

  3. [3]

    Zeyu Cai, Zhelong Huang, Xu Zheng, Yexin Liu, Chao Liu, Zeyu Wang, and Lin Wang. 2024. Interact360: Interactive identity-driven text to 360 panorama generation. In 2024 IEEE Conference on Artificial Intelligence (CAI). IEEE, 728–736

    2024

  4. [4]

    Yihong Cao, Jiaming Zhang, Xu Zheng, Hao Shi, Kunyu Peng, Hang Liu, Kailun Yang, and Hui Zhang. 2025. Unlocking Constraints: Source-Free Occlusion-Aware Seamless Segmentation. arXiv preprint arXiv:2506.21198 (2025)

    work page Pith review arXiv 2025

  5. [5]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al . 2024. Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling. arXiv preprint arXiv:2412.05271 (2024)

    arXiv 2024

  6. [6]

    Shih-Han Chou, Wei-Lun Chao, Wei-Sheng Lai, Min Sun, and Ming-Hsuan Yang. 2020. Visual Question Answering on 360 ° Images. In 2020 IEEE Win- ter Conference on Applications of Computer Vision (W ACV) . IEEE, 1596–1605. doi:10.1109/WACV45572.2020.9093452

    arXiv 2020

  7. [7]

    Fabien Danieau, Thomas Lopez, Nicolas Mollet, Bertrand Leroy, Olivier Dumas, and Jean-François Vial. 2017. Enabling embodiment and interaction in omnidi- rectional videos. In 2017 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 697–702. doi:10.1109/ICME.2017.8019388

    work page arXiv 2017

  8. [8]

    Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. 2017. Visual Dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

Show all 119 references
  1. [11]

    Zihao Dongfang, Xu Zheng, Ziqiao Weng, Yuanhuiyi Lyu, Danda Pani Paudel, Luc Van Gool, Kailun Yang, and Xuming Hu. 2025. Are Multimodal Large Lan- guage Models Ready for Omnidirectional Spatial Reasoning? arXiv preprint arXiv:2505.11907 (2025)

    2025 arXiv

  2. [12]

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. 2025. Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models. arXiv preprint arXiv:2503.06749 (2025). https://arxiv.org/abs/2503.06749

    2025 arXiv

  3. [13]

    Varun Ravi Kumar, Senthil Yogamani, Hazem Rashed, Ganesh Sitsu, Christian Witt, Isabelle Leang, Stefan Milz, and Patrick Mäder. 2021. OmniDet: Surround View Cameras Based Multi-Task Visual Perception Network for Autonomous Driving. IEEE Robotics and Automation Letters 6, 2 (20...

    2021

  4. [14]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: bootstrap- ping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning (ICML’23), Vol. 202. JMLR.org, 19730–19742

    2023

  5. [15]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In Proceedings of the 39th International Conference on Machine Learn- ing (Proceedings of Machine Learning Resea...

    2022

  6. [16]

    Zhigang Li, Guanglei Qi, Wenkai Hu, Xiangyu Ma, and Qinsheng Guo. 2021. Application of Augmented Reality in Campus Navigation. In2021 6th International Conference on Intelligent Computing and Signal Processing (ICSP) . IEEE, 889–893. doi:10.1109/ICSP51882.2021.9408770

    2021

  7. [17]

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Visual-RFT: Visual Reinforcement Fine-Tuning. arXiv preprint arXiv:2503.01785 (2025)

    2025 arXiv

  8. [18]

    Ziyu Liu, Yuhang Zang, Yushan Zou, Zijian Liang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Visual Agentic Reinforcement Fine-Tuning. arXiv preprint arXiv:2505.14246 (2025). https://arxiv.org/abs/2505. 14246

    2025 arXiv

  9. [19]

    Yuanhuiyi Lyu, Xu Zheng, Dahun Kim, and Lin Wang. 2024. OmniBind: Teach to Build Unequal-Scale Modality Interaction for Omni-Bind of All. arXiv preprint arXiv:2405.16108 (2024). https://arxiv.org/abs/2405.16108

    2024 arXiv

  10. [21]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. In Ad- vances in Neural Information Processing Syste...

    2022

  11. [22]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, Vol. 36. 53728–53741

    2023

  12. [23]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  13. [24]

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

    2024 arXiv

  14. [25]

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615 (2025)

    2025 arXiv

  15. [26]

    DeepSeek-AIv Team. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv preprint arXiv:2501.12948 (2025). https://arxiv.org/abs/2501.12948

    2025 arXiv

  16. [27]

    Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. 2024. LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. arXiv preprint arXiv:2407.15754 (2024). https://arxiv.org/abs/2407.15754

    2024 arXiv

  17. [28]

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Jun- yang Lin. 2025. Qwen2.5-Omni Technical Report. arXiv preprint arXiv:2503.20215 (2025). https://arxiv.org/abs/2503.20215

    2025 arXiv

  18. [29]

    Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. 2015. Show, Attend and Tell: Neural Image Caption Generation with Visual Attention. In Proceedings of the 32nd International Conference on Machine Learning (Pro...

    2015

  19. [30]

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. 2025. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization. arXiv preprint arXiv:2503.10615 (2025)

    2025 arXiv

  20. [31]

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. 2024. MMM...

    2024

  21. [32]

    Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. 2021. Pano-AVQA: Grounded Audio-Visual Question Answering on 360° Videos. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) . IEEE, 2011–2021. doi:10.1109/ICCV48922.2021.00204

    2021

  22. [34]

    Xinshen Zhang, Zhen Ye, and Xu Zheng. 2025. Towards Omnidirectional Rea- soning with 360-R1: A Dataset, Benchmark, and GRPO-based Method. arXiv Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Zhang et al. preprint arXiv:2505.14197 (2025). https://arxiv.org/abs/2505.14197

    2025 arXiv

  23. [35]

    Xu Zheng, Chenfei Liao, Yuqian Fu, Kaiyu Lei, Yuanhuiyi Lyu, Lutao Jiang, Bin Ren, Jialei Chen, Jiawen Wang, Chengxin Li, et al . 2025. MLLMs are Deeply Affected by Modality Bias. arXiv preprint arXiv:2505.18657 (2025)

    2025 arXiv

  24. [36]

    Xu Zheng, Tianbo Pan, Yunhao Luo, and Lin Wang. 2023. Look at the Neighbor: Distortion-aware Unsupervised Domain Adaptation for Panoramic Semantic Segmentation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 18641–18652. doi:10.1109/ICCV51070.2023.01713

    2023

  25. [37]

    Vasilakos, and Lin Wang

    Xu Zheng, Pengyuan Zhou, Athanasios V. Vasilakos, and Lin Wang. 2024. Se- mantics, Distortion, and Style Matter: Towards Source-Free UDA for Panoramic Segmentation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). IEEE, 27885–27895. doi:10.1109/...

    2024

  26. [38]

    Vasilakos, and Lin Wang

    Xu Zheng, Peng Yuan Zhou, Athanasios V. Vasilakos, and Lin Wang. 2025. 360SFUDA++: Towards Source-Free UDA for Panoramic Segmentation by Learn- ing Reliable Category Prototypes. IEEE Transactions on Pattern Analysis and Machine Intelligence 47, 2 (2025), 1190–1204. doi:10.1109...

    2025

  27. [39]

    Xu Zheng, Jinjing Zhu, Yexin Liu, Zidong Cao, Chong Fu, and Lin Wang. 2023. Both Style and Distortion Matter: Dual-Path Unsupervised Domain Adaptation for Panoramic Semantic Segmentation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 128...

    2023

  28. [40]

    Ding Zhong, Xu Zheng, Chenfei Liao, Yuanhuiyi Lyu, Jialei Chen, Shengyang Wu, Linfeng Zhang, and Xuming Hu. 2025. Omnisam: Omnidirectional segment anything model for uda in panoramic semantic segmentation. arXiv preprint arXiv:2503.07098 (2025)

    2025

  29. [41]

    Frame A" and

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Nianchen Deng, Songze Li,...

    2025 arXiv

  30. [43]

    question

    **Format**: This is a Multiple-Choice Question (MCQ). 4. **JSON Output**: Your final output must be a single, valid JSON object with the following keys: - "question": The generated question string. - "options": A JSON object containing four plausible options, with keys "A", "B...

  31. [44]

    question

    **Format**: This is an open-ended Question-Answer (QA). 4. **JSON Output**: Your final output must be a single, valid JSON object with the following keys: - "question": The generated question string. - "answer": A concise, factual, open-ended answer string. The JSON object for...

    2018

  32. [45]

    Basic Understanding 37.50% 0.7383 1.1 Perspective Definition & Identification 29.63% 0.9089 1.2 Effect of Perspective on Object Shape 56.82% 0.8198 1.3 Perspective & Occlusion 37.50% 0.5382

  33. [46]

    Image Characteristics 44.32% 0.7254 2.1 Field of View & Information 48.96% 0.6403 2.2 Distortion & Perspective Effects 38.75% 0.8336 2.3 Multi-view Fusion – 0.7727

  34. [47]

    Perspective Question Design 59.90% 0.7470 3.1 Position Inference 60.53% 0.7063 3.2 Spatial Relationships 64.44% 0.7356 3.3 Dynamic Perspective Change 63.49% 0.7868 3.4 Size Judgment 72.73% – 3.5 Perspective Transformation 21.74% 0.7334

  35. [48]

    Advanced Reasoning 72.82% 0.7184 4.1 Multi-level Spatial Understanding 56.82% 0.6910 4.2 Inferring Implied Relationships 85.14% 0.9076 4.3 Inferring Scene Narrative 70.13% 0.7648

  36. [49]

    Quantitative Reasoning 44.49% 0.8940 5.1 Occlusion-Based Counting 52.50% – 5.2 Disambiguation & Total Count 38.10% 0.8779 5.3 Conditional & Comparative Counting 41.94% 0.9041 B.1 InternVL3-9B B.2 InternVL3-8B Table 5: Performance Evaluation of InternVL3-8B Category/Sub-categor...

  37. [50]

    Basic Understanding 55.56% 0.7750 1.1 Perspective Definition & Identification 53.70% 0.4999 1.2 Effect of Perspective on Object Shape 50.00% 0.8197 1.3 Perspective & Occlusion 62.50% 0.6839

  38. [51]

    Image Characteristics 48.86% 0.6858 2.1 Field of View & Information 61.46% 0.5985 2.2 Distortion & Perspective Effects 33.75% 0.6816 2.3 Multi-view Fusion – 0.7420

  39. [52]

    Perspective Question Design 49.50% 0.6846 3.1 Position Inference 63.16% 0.5401 3.2 Spatial Relationships 68.89% 0.8109 3.3 Dynamic Perspective Change 17.46% 0.7051 3.4 Size Judgment 69.70% – 3.5 Perspective Transformation 47.83% 0.6807

  40. [53]

    Advanced Reasoning 65.64% 0.7574 4.1 Multi-level Spatial Understanding 77.27% 0.7402 4.2 Inferring Implied Relationships 55.41% 0.8915 4.3 Inferring Scene Narrative 68.83% 0.7857

  41. [54]

    Quantitative Reasoning 22.03% 0.9173 5.1 Occlusion-Based Counting 8.75% – 5.2 Disambiguation & Total Count 6.35% 0.8962 5.3 Conditional & Comparative Counting 44.09% 0.9306 B.3 InternVL3-2B Table 6: Performance Evaluation of InternVL3-2B Category/Sub-category MCQ Accuracy QA S...

  42. [55]

    Basic Understanding 52.31% 0.6978 1.1 Perspective Definition & Identification 56.48% 0.3506 1.2 Effect of Perspective on Object Shape 56.82% 0.7828 1.3 Perspective & Occlusion 42.19% 0.5154

  43. [56]

    Image Characteristics 42.05% 0.6793 2.1 Field of View & Information 53.12% 0.6198 2.2 Distortion & Perspective Effects 28.75% 0.5474 2.3 Multi-view Fusion – 0.7262

  44. [57]

    Perspective Question Design 38.12% 0.6303 3.1 Position Inference 47.37% 0.5772 3.2 Spatial Relationships 57.78% 0.7974 3.3 Dynamic Perspective Change 20.63% 0.5080 3.4 Size Judgment 48.48% – 3.5 Perspective Transformation 17.39% 0.7153

  45. [58]

    Advanced Reasoning 56.41% 0.7416 4.1 Multi-level Spatial Understanding 59.09% 0.7187 4.2 Inferring Implied Relationships 56.76% 0.8524 4.3 Inferring Scene Narrative 54.55% 0.7827

  46. [59]

    Quantitative Reasoning 23.73% 0.7070 5.1 Occlusion-Based Counting 22.50% – 5.2 Disambiguation & Total Count 3.17% 0.3876 5.3 Conditional & Comparative Counting 38.71% 0.9084 B.4 InternVL2.5-8B Table 7: Performance Evaluation of InternVL2.5-8B Category/Sub-category MCQ Accuracy...

  47. [60]

    Basic Understanding 44.91% 0.8230 1.1 Perspective Definition & Identification 38.89% 0.5179 1.2 Effect of Perspective on Object Shape 59.09% 0.8528 1.3 Perspective & Occlusion 45.31% 0.7682

  48. [61]

    Image Characteristics 44.32% 0.7558 2.1 Field of View & Information 38.54% 0.5961 2.2 Distortion & Perspective Effects 51.25% 0.8493 2.3 Multi-view Fusion – 0.8518

  49. [62]

    Perspective Question Design 53.96% 0.7055 3.1 Position Inference 63.16% 0.6672 3.2 Spatial Relationships 57.78% 0.6103 3.3 Dynamic Perspective Change 52.38% 0.7277 3.4 Size Judgment 51.52% – 3.5 Perspective Transformation 39.13% 0.7695

  50. [63]

    Advanced Reasoning 72.82% 0.7103 4.1 Multi-level Spatial Understanding 50.00% 0.7001 4.2 Inferring Implied Relationships 87.84% 0.6823 4.3 Inferring Scene Narrative 71.43% 0.7326

  51. [64]

    B.5 InternVL2.5-4B Table 8: Performance Evaluation of InternVL2.5-4B Category/Sub-category MCQ Accuracy QA Similarity Overall Performance 46.93% 0.7748

    Quantitative Reasoning 41.53% 0.8918 5.1 Occlusion-Based Counting 43.75% – 5.2 Disambiguation & Total Count 28.57% 0.8495 5.3 Conditional & Comparative Counting 48.39% 0.9185 Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Zhang et al. B.5 InternVL2.5-4B Table 8: Perfo...

    2018

  52. [65]

    Basic Understanding 49.54% 0.7942 1.1 Perspective Definition & Identification 55.56% 0.6769 1.2 Effect of Perspective on Object Shape 54.55% 0.8121 1.3 Perspective & Occlusion 35.94% 0.7579

  53. [66]

    Image Characteristics 44.89% 0.7650 2.1 Field of View & Information 45.83% 0.6450 2.2 Distortion & Perspective Effects 43.75% 0.7042 2.3 Multi-view Fusion – 0.8459

  54. [67]

    Perspective Question Design 47.03% 0.7411 3.1 Position Inference 50.00% 0.7143 3.2 Spatial Relationships 55.56% 0.7867 3.3 Dynamic Perspective Change 41.27% 0.7116 3.4 Size Judgment 39.39% – 3.5 Perspective Transformation 52.17% 0.7687

  55. [68]

    Advanced Reasoning 70.26% 0.7799 4.1 Multi-level Spatial Understanding 56.82% 0.7700 4.2 Inferring Implied Relationships 71.62% 0.8058 4.3 Inferring Scene Narrative 76.62% 0.7988

  56. [69]

    Quantitative Reasoning 26.69% 0.8882 5.1 Occlusion-Based Counting 21.25% – 5.2 Disambiguation & Total Count 25.40% 0.8485 5.3 Conditional & Comparative Counting 32.26% 0.9132 B.6 InternVL2.5-2B Table 9: Performance Evaluation of InternVL2.5-2B Category/Sub-category MCQ Accurac...

  57. [70]

    Basic Understanding 44.44% 0.7536 1.1 Perspective Definition & Identification 50.93% 0.8985 1.2 Effect of Perspective on Object Shape 45.45% 0.7982 1.3 Perspective & Occlusion 32.81% 0.6417

  58. [71]

    Image Characteristics 42.61% 0.6519 2.1 Field of View & Information 60.42% 0.5565 2.2 Distortion & Perspective Effects 21.25% 0.5092 2.3 Multi-view Fusion – 0.7225

  59. [72]

    Perspective Question Design 40.10% 0.5919 3.1 Position Inference 28.95% 0.6784 3.2 Spatial Relationships 68.89% 0.7400 3.3 Dynamic Perspective Change 31.75% 0.4405 3.4 Size Judgment 45.45% – 3.5 Perspective Transformation 17.39% 0.6232

  60. [73]

    Advanced Reasoning 64.10% 0.7002 4.1 Multi-level Spatial Understanding 61.36% 0.7019 4.2 Inferring Implied Relationships 54.05% 0.7924 4.3 Inferring Scene Narrative 75.32% 0.6920

  61. [74]

    Quantitative Reasoning 22.88% 0.7249 5.1 Occlusion-Based Counting 10.00% – 5.2 Disambiguation & Total Count 11.11% 0.5116 5.3 Conditional & Comparative Counting 41.94% 0.8594 B.7 Qwen2.5-Omni-7B Table 10: Performance Evaluation of Qwen2.5-Omni-7B Category/Sub-category MCQ Accu...

  62. [75]

    Basic Understanding 43.52% 0.6329 1.1 Perspective Definition & Identification 37.96% 0.4448 1.2 Effect of Perspective on Object Shape 54.55% 0.7747 1.3 Perspective & Occlusion 45.31% 0.3091

  63. [76]

    Image Characteristics 35.23% 0.6503 2.1 Field of View & Information 39.58% 0.6211 2.2 Distortion & Perspective Effects 30.00% 0.7135 2.3 Multi-view Fusion – 0.6648

  64. [77]

    Perspective Question Design 50.99% 0.5316 3.1 Position Inference 44.74% 0.6156 3.2 Spatial Relationships 68.89% 0.5515 3.3 Dynamic Perspective Change 39.68% 0.4477 3.4 Size Judgment 54.55% – 3.5 Perspective Transformation 52.17% 0.5644

  65. [78]

    Advanced Reasoning 67.69% 0.6936 4.1 Multi-level Spatial Understanding 63.64% 0.7152 4.2 Inferring Implied Relationships 77.03% 0.6633 4.3 Inferring Scene Narrative 61.04% 0.6510

  66. [79]

    Quantitative Reasoning 47.03% 0.7238 5.1 Occlusion-Based Counting 56.25% – 5.2 Disambiguation & Total Count 34.92% 0.5455 5.3 Conditional & Comparative Counting 47.31% 0.8363 B.8 Qwen2.5-Omni-3B Table 11: Performance Evaluation of Qwen2.5-Omni-3B Category/Sub-category MCQ Accu...

  67. [80]

    Basic Understanding 38.89% 0.6296 1.1 Perspective Definition & Identification 41.67% 0.4448 1.2 Effect of Perspective on Object Shape 46.59% 0.7782 1.3 Perspective & Occlusion 29.69% 0.2898

  68. [81]

    Image Characteristics 30.11% 0.6063 2.1 Field of View & Information 32.29% 0.6257 2.2 Distortion & Perspective Effects 27.50% 0.5556 2.3 Multi-view Fusion – 0.5973

  69. [82]

    Perspective Question Design 43.56% 0.5537 3.1 Position Inference 44.74% 0.6198 3.2 Spatial Relationships 61.36% 0.6202 3.3 Dynamic Perspective Change 33.33% 0.4427 3.4 Size Judgment 43.43% – 3.5 Perspective Transformation 34.78% 0.6031

  70. [83]

    Advanced Reasoning 72.31% 0.6627 4.1 Multi-level Spatial Understanding 54.55% 0.6894 4.2 Inferring Implied Relationships 77.03% 0.6581 4.3 Inferring Scene Narrative 76.62% 0.6081

  71. [84]

    Quantitative Reasoning 35.17% 0.7950 5.1 Occlusion-Based Counting 37.50% – 5.2 Disambiguation & Total Count 33.33% 0.7538 5.3 Conditional & Comparative Counting 35.48% 0.8210 Omnidirectional Spatial Modeling from Correlated Panoramas Conference acronym ’XX, June 03–05, 2018, W...

    2018

  72. [85]

    Basic Understanding 39.35% 0.7838 1.1 Perspective Definition & Identification 29.63% 0.7109 1.2 Effect of Perspective on Object Shape 52.27% 0.7932 1.3 Perspective & Occlusion 46.88% 0.7654

  73. [86]

    Image Characteristics 41.48% 0.7923 2.1 Field of View & Information 41.67% 0.6675 2.2 Distortion & Perspective Effects 41.25% 0.8668 2.3 Multi-view Fusion – 0.8672

  74. [87]

    Perspective Question Design 62.87% 0.7743 3.1 Position Inference 68.42% 0.7487 3.2 Spatial Relationships 64.44% 0.8602 3.3 Dynamic Perspective Change 61.90% 0.7262 3.4 Size Judgment 66.67% – 3.5 Perspective Transformation 47.83% 0.7976

  75. [88]

    Advanced Reasoning 73.33% 0.7583 4.1 Multi-level Spatial Understanding 72.73% 0.7496 4.2 Inferring Implied Relationships 85.14% 0.8892 4.3 Inferring Scene Narrative 62.34% 0.7693

  76. [89]

    Quantitative Reasoning 37.71% 0.8798 5.1 Occlusion-Based Counting 38.75% – 5.2 Disambiguation & Total Count 36.51% 0.8174 5.3 Conditional & Comparative Counting 37.63% 0.9192 B.10 Qwen2.5-VL-3B Table 13: Performance Evaluation of Qwen2.5-VL-3B Category/Sub-category MCQ Accurac...

  77. [90]

    Basic Understanding 40.74% 0.7975 1.1 Perspective Definition & Identification 35.19% 0.7773 1.2 Effect of Perspective on Object Shape 59.09% 0.8125 1.3 Perspective & Occlusion 37.50% 0.7632

  78. [91]

    Image Characteristics 55.68% 0.7945 2.1 Field of View & Information 47.92% 0.7210 2.2 Distortion & Perspective Effects 65.00% 0.7823 2.3 Multi-view Fusion – 0.8424

  79. [92]

    Perspective Question Design 47.52% 0.7895 3.1 Position Inference 39.47% 0.7506 3.2 Spatial Relationships 64.44% 0.8723 3.3 Dynamic Perspective Change 41.27% 0.7749 3.4 Size Judgment 57.58% – 3.5 Perspective Transformation 30.43% 0.7813

  80. [93]

    Advanced Reasoning 55.38% 0.7634 4.1 Multi-level Spatial Understanding 34.09% 0.7554 4.2 Inferring Implied Relationships 72.97% 0.8141 4.3 Inferring Scene Narrative 50.65% 0.7772

  81. [94]

    Quantitative Reasoning 36.02% 0.8512 5.1 Occlusion-Based Counting 31.25% – 5.2 Disambiguation & Total Count 30.16% 0.7245 5.3 Conditional & Comparative Counting 44.09% 0.9310 B.11 Pano-R1-7B Table 14: Performance Evaluation of Pano-R1-7B Category/Sub-category MCQ Accuracy QA S...

  82. [95]

    Basic Understanding 42.13% 0.8255 1.1 Perspective Definition & Identification 30.56% 0.9660 1.2 Effect of Perspective on Object Shape 65.91% 0.8380 1.3 Perspective & Occlusion 45.31% 0.7892

  83. [96]

    Image Characteristics 61.93% 0.8264 2.1 Field of View & Information 67.71% 0.7370 2.2 Distortion & Perspective Effects 55.00% 0.8693 2.3 Multi-view Fusion – 0.8807

  84. [97]

    Perspective Question Design 65.35% 0.8304 3.1 Position Inference 78.95% 0.8003 3.2 Spatial Relationships 68.89% 0.8843 3.3 Dynamic Perspective Change 66.67% 0.8355 3.4 Size Judgment 63.64% – 3.5 Perspective Transformation 34.78% 0.8099

  85. [98]

    Advanced Reasoning 76.41% 0.7545 4.1 Multi-level Spatial Understanding 68.18% 0.7400 4.2 Inferring Implied Relationships 90.54% 0.8730 4.3 Inferring Scene Narrative 67.53% 0.7780

  86. [99]

    Quantitative Reasoning 42.80% 0.9110 5.1 Occlusion-Based Counting 45.00% – 5.2 Disambiguation & Total Count 41.27% 0.8807 5.3 Conditional & Comparative Counting 41.94% 0.9301 B.12 Pano-R1-3B Table 15: Performance Evaluation of Pano-R1-3B Category/Sub-category MCQ Accuracy QA S...

  87. [100]

    Basic Understanding 44.91% 0.7893 1.1 Perspective Definition & Identification 42.59% 0.9159 1.2 Effect of Perspective on Object Shape 65.91% 0.7964 1.3 Perspective & Occlusion 34.38% 0.7680

  88. [101]

    Image Characteristics 55.11% 0.7979 2.1 Field of View & Information 55.21% 0.7682 2.2 Distortion & Perspective Effects 55.00% 0.8481 2.3 Multi-view Fusion – 0.8130

  89. [102]

    Perspective Question Design 48.51% 0.8054 3.1 Position Inference 50.00% 0.7645 3.2 Spatial Relationships 66.67% 0.8463 3.3 Dynamic Perspective Change 39.68% 0.8320 3.4 Size Judgment 60.61% – 3.5 Perspective Transformation 17.39% 0.7703

  90. [103]

    Advanced Reasoning 64.10% 0.7531 4.1 Multi-level Spatial Understanding 50.00% 0.7493 4.2 Inferring Implied Relationships 83.78% 0.7605 4.3 Inferring Scene Narrative 53.25% 0.7600

  91. [104]

    Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Zhang et al

    Quantitative Reasoning 38.98% 0.8639 5.1 Occlusion-Based Counting 42.50% – 5.2 Disambiguation & Total Count 28.57% 0.8256 5.3 Conditional & Comparative Counting 43.01% 0.8881 C Dataset Information This appendix provides more detailed information about the datasets we introduce...

    2018

  92. [105]

    Basic Understanding 1,534 18.95% 1.1 Perspective Definition & Identification 569 7.03% 1.2 Effect of Perspective on Object Shape 500 6.18% 1.3 Perspective & Occlusion 465 5.74%

  93. [106]

    Image Characteristics 1,466 18.11% 2.1 Field of View & Information 670 8.28% 2.2 Distortion & Perspective Effects 465 5.74% 2.3 Multi-view Fusion 331 4.09%

  94. [107]

    Perspective Question Design 2,331 28.80% 3.1 Position Inference 450 5.56% 3.2 Spatial Relationships 452 5.58% 3.3 Dynamic Perspective Change 723 8.93% 3.4 Size Judgment 231 2.85% 3.5 Perspective Transformation 475 5.87%

  95. [108]

    Advanced Reasoning 1,223 15.11% 4.1 Multi-level Spatial Understanding 450 5.56% 4.2 Inferring Implied Relationships 322 3.98% 4.3 Inferring Scene Narrative 451 5.57%

  96. [109]

    Quantitative Reasoning 1,540 19.03% 5.1 Occlusion-Based Counting 384 4.74% 5.2 Disambiguation & Total Count 481 5.94% 5.3 Conditional & Comparative Counting 675 8.34% Table 18: Training Set Category Distribution Category/Sub-category Count Percentage Total Training Questions 6...

  97. [110]

    Basic Understanding 1,250 19.31% 1.1 Perspective Definition & Identification 460 7.10% 1.2 Effect of Perspective on Object Shape 409 6.32% 1.3 Perspective & Occlusion 381 5.88%

  98. [111]

    Image Characteristics 1,162 17.95% 2.1 Field of View & Information 526 8.12% 2.2 Distortion & Perspective Effects 380 5.87% 2.3 Multi-view Fusion 256 3.95%

  99. [112]

    Perspective Question Design 1,865 28.80% 3.1 Position Inference 359 5.54% 3.2 Spatial Relationships 359 5.54% 3.3 Dynamic Perspective Change 568 8.77% 3.4 Size Judgment 198 3.06% 3.5 Perspective Transformation 381 5.88%

  100. [113]

    Advanced Reasoning 969 14.97% 4.1 Multi-level Spatial Understanding 367 5.67% 4.2 Inferring Implied Relationships 247 3.81% 4.3 Inferring Scene Narrative 355 5.48%

  101. [114]

    Flex means both question type are applied in this question tempate text

    Quantitative Reasoning 1,229 18.98% 5.1 Occlusion-Based Counting 304 4.69% 5.2 Disambiguation & Total Count 389 6.01% 5.3 Conditional & Comparative Counting 536 8.28% D Question Templates This appendix includes a large table of question templates. Flex means both question type...

  102. [115]

    Basic Understanding 284 17.54% 1.1 Perspective Definition & Identification 109 6.73% 1.2 Effect of Perspective on Object Shape 91 5.62% 1.3 Perspective & Occlusion 84 5.19%

  103. [116]

    Image Characteristics 304 18.78% 2.1 Field of View & Information 144 8.89% 2.2 Distortion & Perspective Effects 85 5.25% 2.3 Multi-view Fusion 75 4.63%

  104. [117]

    Perspective Question Design 466 28.78% 3.1 Position Inference 91 5.62% 3.2 Spatial Relationships 93 5.74% 3.3 Dynamic Perspective Change 155 9.57% 3.4 Size Judgment 33 2.04% 3.5 Perspective Transformation 94 5.81%

  105. [118]

    Advanced Reasoning 254 15.69% 4.1 Multi-level Spatial Understanding 83 5.13% 4.2 Inferring Implied Relationships 75 4.63% 4.3 Inferring Scene Narrative 96 5.93%

  106. [119]

    Each case compares the per- formance of Qwen2.5VL-7B-Instruct, InternVL3-9B, and Pano-R1 on panoramic images, highlighting differences in reasoning and output accuracy

    Quantitative Reasoning 311 19.21% 5.1 Occlusion-Based Counting 80 4.94% 5.2 Disambiguation & Total Count 92 5.68% 5.3 Conditional & Comparative Counting 139 8.59% E Model Inference Cases This appendix presents three model inference cases, each in a large figure spanning two co...

    2007

  107. [120]

    Basic Understanding 1.1 Perspective Definition & Identification T1.1.1 MCQ Comparing {frame_X} and {frame_Y}, which one was taken from a higher viewpoint? T1.1.2 MCQ Which frame, {frame_X} or {frame_Y}, provides a more direct top-down view of {object_A}? 1.2 Effect of Perspect...

  108. [121]

    Image Characteristics 2.1 Field of View & Information T2.1.1 MCQ Comparing the two images, which one ( {frame_X} or {frame_Y}) captures a wider field of view of the scene? T2.1.2 MCQ Considering an object located at the edge of the view in {frame_X}, does this same object appe...

  109. [122]

    Describe the camera’s direction of movement from {frame_X} to {frame_Y}

    Perspec- tive Question Design 3.1 Position Inference T3.1.1 QA {frame_X} shows the front of {object_A}, while {frame_Y} shows its side. Describe the camera’s direction of movement from {frame_X} to {frame_Y}. T3.1.2 MCQ Using both views, determine if {object_A} is freestanding...

  110. [123]

    Describe how the apparent shape of the room changes between the perspective of Frame A and Frame B

    Quanti- tative Reasoning 5.1 Occlusion-Based Counting T5.1.1 MCQ In {frame_X}, some {object_type_plural} are hidden. Using {frame_Y} to see the full scene, what is the true total number of {object_type_plural}? T5.1.2 MCQ {frame_Y} provides a clearer view of an area that is pa...

    2018

  111. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.