Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Keeping a scene memory across episodes lifts object-finding success

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A hierarchical room-and-object memory that persists across independent ObjectNav episodes yields small success-rate gains, but most of the gain comes from within-episode memory rather than the cross-episode component.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The protocol is new and the work is honest, but the paper's own ablation says cross-episode persistence moves SR by 0.2 points — the headline attribution doesn't hold. the 3 major comments →

arxiv 2607.14514 v2 pith:D6NKBGI4 submitted 2026-07-16 cs.CV cs.AI

VTM-Nav: Harnessing Cross-Episode Experience for Object-Goal Navigation with Hierarchical Visual-Topological Memory

Xiaoran Xu , Yupeng Wu , Tianyu Xue , Yifan Xu , Xuanran Dong , Xiaoshan Yang , Changsheng Xu This is my paper
classification cs.CV cs.AI
keywords Object-Goal NavigationCross-episode experienceVisual-topological memoryTraining-free navigationVision-language modelsPersistent memoryRe-localizationRoom topology
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper asks whether a navigation agent with fixed parameters can get better at finding objects in a scene simply by remembering what it saw on earlier visits. It proposes VTM-Nav, which stores a coarse room topology that indexes visual memories owned by each room, along with cues about where objects were seen and which approaches worked. On three ObjectNav benchmarks under a matched 40-step protocol, the agent outperforms a memory-reset version of the same pipeline and a version that keeps only free-form textual summaries. The message is that structured, scene-scoped memory can substitute for retraining or oracle knowledge in repeated deployments.

Core claim

The paper claims that a training-free object-goal navigation agent can reuse scene experience across independently initialized episodes by maintaining a two-level memory: a coarse room topology indexing fine-grained, room-owned visual records. Under matched 40-step comparisons with identical navigation components, the proposed VTM-Nav achieves success-rate gains of 4.6, 2.0, and 0.8 percentage points over the memory-reset control on HM3D v0.1, HM3D v0.2, and MP3D, and 3.1 and 5.5 points over a free-form textual-memory baseline on the two HM3D splits, with comparable or better SPL. The memory update, re-localization, retrieval, and candidate re-ranking all use fixed model parameters and no or

What carries the argument

The central mechanism is the Hierarchical Visual-Topological Memory (VTM): a room graph whose nodes own local object memories. Each object record distinguishes in-room observations from remote-visible cues seen through openings, and successful episodes add confidence and approach-direction information. At inference, the agent re-localizes in the accumulated graph via a fixed combination of semantic, topological, and spatial consistency scores, retrieves candidate target rooms with a coarse-to-fine search, and lets retrieved evidence re-rank only navigation candidates derived from the current observation. A conservative execution guard handles local failures. This design allows experience to

Load-bearing premise

The claimed gains over the memory-reset baseline rest on the assumption that the paper's reproduction of that baseline behaves like the published system; if the reproduction is weaker, the 4.6/2.0/0.8 point advantages overstate VTM-Nav's contribution.

What would settle it

Re-run the matched 40-step comparisons against the actual published baseline implementation (or verify that the paper's reproduction matches its reported numbers on the same episodes); if the advantage over memory-reset falls within a bootstrap confidence interval that includes zero, the cross-episode memory claim fails. Also check whether the persistence ablation—removing only the across-episode carry-over—reproduces a meaningful success-rate change.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training-free ObjectNav agents can improve with experience in a scene without any parameter updates or oracle maps.
  • Structured memory (topology plus room-owned visuals) beats retaining free-form textual summaries.
  • The gains appear early in a scene's episode sequence, with late-half success improving more for VTM-Nav than for baselines, suggesting a cumulative benefit.
  • The memory organization transfers across dataset scales (HM3D and MP3D) with fixed thresholds.
  • The approach opens an alternative axis for ObjectNav research: instead of better models or training, reuse of cross-episode experience.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The persistence-specific ablation is small (59.6 vs 59.4 SR), so part of the benefit may come from the within-episode memory structure rather than cross-episode persistence; the causal attribution to persistence is weaker than the headline gains suggest.
  • The setting could be extended to lifelong deployment where episodes are not pre-shuffled; one might test whether gains grow with the number of visits and whether memory ever saturates or needs forgetting.
  • The room-ownership principle might apply to other embodied tasks, such as object search with mobile manipulators or vision-language navigation, where a coarse spatial index over episodic memories could be reused.
  • A testable prediction: if memory is corrupted by mislocalized rooms, late-episode performance should degrade more than early-episode performance, which could be probed by injecting noisy room labels.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes VTM-Nav, a training-free object-goal navigation system that maintains a persistent hierarchical visual-topological memory across independently initialized episodes within the same scene. A coarse room graph indexes room-owned visual records; at each step the agent re-localizes in this graph, updates memory, retrieves target-relevant room/object evidence, and grounds VLM candidate selection in the current observation, with a conservative execution guard. Evaluations on HM3D v0.1, HM3D v0.2, and MP3D under a matched 40-step protocol report SR gains over a memory-reset WMNav* control (4.6, 2.0, and 0.8 points) and over persistent textual memory (3.1 and 5.5 points on the HM3D benchmarks), plus component ablations and a cross-episode progression analysis. The paper claims these results demonstrate effective reuse of cross-episode scene experience.

Significance. If the central claim were fully supported, the paper would make a useful contribution: showing that a training-free agent can reuse scene-scoped experience across independent episodes without retraining or oracle information is a timely and practical problem. The controlled matched-component protocol, fixed thresholds across datasets, honest reporting of confidence intervals, and explicit labeling of the progression analysis as descriptive are methodological strengths. However, the attribution of the headline gains to cross-episode persistence is the paper's load-bearing claim, and, as detailed below, the paper's own ablation and statistics do not currently support that attribution. The contribution is still potentially valuable as a within-episode hierarchical visual-topological memory system, but the claims and framing need substantial revision or additional evidence.

major comments (3)
  1. [§4.5, Table 2; Abstract/§4.3] Table 2 shows that removing cross-episode persistence (clearing VTM at episode boundaries while retaining all other components) changes SR from 59.6 to 59.4, a 0.2-point difference. In contrast, removing topology, the guard, or visual evidence changes SR by 3.1, 1.6, and 1.9 points respectively. The full-system gain over the memory-reset control is 59.6 vs 55.0, i.e., 4.6 points, of which at most 0.2 is attributable to persistence. The Abstract's statement that the results 'demonstrate effective reuse of cross-episode scene experience' is therefore not supported by the paper's own controlled comparison. The authors need either a persistence-specific experiment with adequate statistical power (e.g., multiple scene-order/target schedules, or a paired confidence interval for the persistence ablation), or a reframing of the contribution as within-episode hierarchical visual-topological memor
  2. [§4.4 and Table 1] The only confidence interval reported for the SR difference against the memory-reset control is the HM3D v0.2 paired scene-cluster bootstrap: [−0.60, 3.87] percentage points, which includes zero. The headline SR gains on HM3D v0.1 and MP3D are reported without any uncertainty quantification. Thus the statistical support for 'VTM-Nav exceeds the memory-reset WMNav control' is currently weak, especially given the 0.2-point persistence contribution in the ablation. The paper should report paired scene-cluster bootstrap intervals (or equivalent) for all matched comparisons, including the persistence-only contrast, and should reconcile the Abstract's definitive tone with §4.4's statement that the remaining differences are 'unresolved point-estimate trends.'
  3. [§4.2, Table 1] The control WMNav* is described as 'our WMNav reproduction or extension.' No evidence is provided that this reproduction faithfully matches the published WMNav system on the same episodes. Differences in the VLM prompt, candidate pipeline, step-budget handling, or other implementation details could make the control artificially weak and inflate VTM-Nav's measured advantage. Since the headline comparisons are all against this in-house control, the authors should calibrate WMNav* against the original WMNav (e.g., by reproducing published numbers on the same evaluation episodes or using the released code), or report the original WMNav results under the same protocol. Without this, the controlled-comparison claim is not fully interpretable.
minor comments (6)
  1. [§4.3 heading] Typo: 'Comparision' should be 'Comparison.'
  2. [§2.1 heading] Typo: 'T raining' should be 'Training.'
  3. [Eqs. (13), (15)] The superscript on the room set is missing: should be V^room_s, not V rooms_s. Also check the corresponding notation in Eq. (15).
  4. [Table 2] The row labels 'Guard' and 'Visual' are not fully defined in the table caption. Clarify exactly which components are included/excluded in each ablation, especially whether 'Visual' includes both in-room and remote-visible cues and whether 'Guard' includes the lexical re-ranking threshold τ_bias.
  5. [§4.6, Table 4] The progression analysis is appropriately labeled descriptive, but the early/late split is confounded by episode order, target composition, and difficulty. The paper should state more explicitly that this analysis does not provide causal evidence for memory accumulation, and should avoid using it as supporting evidence for the main claim.
  6. [General] No code or data release is mentioned. For reproducibility of the controlled comparisons, please provide the implementation, prompts, lexical dictionaries, and exact threshold values used in the Appendix or as supplementary material.

Circularity Check

0 steps flagged

No circularity: VTM-Nav is an empirical systems paper with hand-set thresholds and matched controls; the main internal-validity concerns are not derivation-circularity.

full rationale

VTM-Nav does not derive its results from fitted parameters or from the quantities it claims to predict. The memory update and retrieval formulas (Eqs. 8–24) use fixed, hand-set thresholds and an unweighted sum of three consistency scores, shared across datasets; there is no training loop, no parameter fit, and no optimization objective whose optimum is the reported SR/SPL. The controlled comparisons share backbone, candidate pipeline, action space, controller, and 40-step budget with the WMNav* control, and the control is explicitly a reproduction/extension (Table 1 note). If that reproduction were unfaithful, the gains would be overestimated, but that is a baseline-fidelity/correctness concern, not circularity: no equation in the paper reduces the claimed improvement to the method's own inputs. The component ablation (Table 2) is the correct control for isolating cross-episode persistence; the small persistence contribution (59.6 vs 59.4 on HM3D v0.1) weakens the causal attribution in the abstract, but the paper itself states in §4.6 that the early/late progression analysis is 'descriptive rather than causal.' There are no load-bearing self-citations, no imported uniqueness theorems, no ansatz smuggled in via citation, and no renaming of a known result as a derivation. The central empirical claim is independently testable against external implementations, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

No new physical entities. The method depends on a number of hand-set thresholds (τobj, τbias, Δe, Δ0, Δ+, remote cap, η weights, equal localization weights) whose values are not stated in the main text and for which no sensitivity analysis is reported. The central empirical claim also rests on the unverified fidelity of the in-house WMNav* control.

free parameters (8)
  • tau_obj
    Object-association threshold for writing observations to memory nodes; hand-set, value deferred to appendix.
  • tau_bias
    Threshold for memory-guidance to override the VLM's candidate selection; hand-set.
  • delta_e
    Confidence increment for room-transition edges; hand-set.
  • delta_0
    Confidence increment for direct in-room object observations; hand-set.
  • delta_plus
    Additional confidence increment for success-backed object records; hand-set, required to exceed delta_0.
  • remote_visible_cap
    Lower confidence cap applied to remote-visible evidence; hand-set.
  • retrieval_weights_eta1_eta2
    Weights on observation count and success count in the retrieval ranking equation (21); hand-set.
  • localization_term_weights = 1,1,1
    Equal weighting of semantic, topological, and spatial consistency in Sloc (equation 8); a fixed modeling choice.
axioms (4)
  • domain assumption WMNav* faithfully represents the published WMNav system
    All controlled gains are computed against WMNav*, described as 'our WMNav reproduction or extension' (Table 1); no evidence is provided that it matches official WMNav performance.
  • domain assumption VLM-generated room labels and object sets are consistent across episodes
    Re-localization and retrieval in §3.3 rely on Qwen3-VL-Plus room/object perception being stable enough across independently initialized episodes.
  • domain assumption Pose estimates are accurate enough for spatial consistency and memory indexing
    Ssp in equation (12) and the room-memory association in §3.3 treat estimated positions Pt as reliable for comparing stored room positions.
  • domain assumption Hand-set thresholds generalize across datasets without per-dataset tuning
    §4.2 states 'fixed thresholds across datasets,' but no sensitivity analysis is provided to show the thresholds are not overfit to HM3D v0.1.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VTM-Nav: Harnessing Cross-Episode Experience for Object-Goal Navigation with Hierarchical Visual-Topological Memory." pith.science (2026). https://pith.science/paper/D6NKBGI4

@misc{pith2026260714514,
  author       = {Pith},
  title        = {Pith review of: VTM-Nav: Harnessing Cross-Episode Experience for Object-Goal Navigation with Hierarchical Visual-Topological Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6NKBGI4}},
  note         = {Machine review of arXiv:2607.14514}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Training-free ObjectNav agents increasingly use vision-language models (VLMs), yet typically discard acquired scene knowledge after each request. We study cross-episode ObjectNav, where each request is an independently initialized, single-goal episode and only self-acquired, scene-scoped memory persists across episodes. We ask whether an agent with fixed model parameters and navigation components can reuse such experience without retraining or oracle information. We introduce \method, a training-free framework with a persistent Hierarchical Visual-Topological Memory (VTM). VTM uses a coarse room topology to index room-owned visual memories, distinguishes in-room from remote-visible evidence, and retains successful approach cues. For each request, VTM-Nav re-localizes the agent in accumulated scene structure, retrieves target-relevant records from plausible rooms, and grounds memory guidance in candidates derived from the current observation. A conservative execution guard further handles local failures. Under matched 40-step comparisons, VTM-Nav exceeds the memory-reset WMNav control by 4.6, 2.0, and 0.8 SR points on HM3D v0.1, HM3D v0.2, and MP3D, respectively, with comparable or higher SPL. On HM3D, it also exceeds WMNav harnessed by textual memory by 3.1 and 5.5 SR points. These results demonstrate effective reuse of cross-episode scene experience through hierarchical visual-topological memory.

Figures

Figures reproduced from arXiv: 2607.14514 by Changsheng Xu, Tianyu Xue, Xiaoran Xu, Xiaoshan Yang, Xuanran Dong, Yifan Xu, Yupeng Wu.

Figure 1
Figure 1. Figure 1: Harnessing cross-episode experience for Object [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VTM-Nav. The upper branch shows re-localization, Hierarchical Visual-Topological Memory [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Failure-to-success comparison on HM3D v0.2. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Anderson, Peter and Chang, Angel X. and Chaplot, Devendra Singh and Dosovitskiy, Alexey and Gupta, Saurabh and Koltun, Vladlen and Kosecka, Jana and Malik, Jitendra and Mottaghi, Roozbeh and Savva, Manolis and Zamir, Amir R. , title =. arXiv preprint arXiv:1807.06757 , year =

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2006.13171 , year =

    Batra, Dhruv and Gokaslan, Aaron and Kembhavi, Aniruddha and Maksymets, Oleksandr and Mottaghi, Roozbeh and Savva, Manolis and Toshev, Alexander and Wijmans, Erik , title =. arXiv preprint arXiv:2006.13171 , year =

    Pith/arXiv arXiv 2006

  4. [4]

    International Conference on 3D Vision , year =

    Chang, Angel and Dai, Angela and Funkhouser, Thomas and Halber, Maciej and Niessner, Matthias and Savva, Manolis and Song, Shuran and Zeng, Andy and Zhang, Yinda , title =. International Conference on 3D Vision , year =

  5. [5]

    and Undersander, Eric and Galuba, Wojciech and Westbury, Andrew and Chang, Angel X

    Ramakrishnan, Santhosh Kumar and Gokaslan, Aaron and Wijmans, Erik and Maksymets, Oleksandr and Clegg, Alexander and Turner, John M. and Undersander, Eric and Galuba, Wojciech and Westbury, Andrew and Chang, Angel X. and Savva, Manolis and Zhao, Yili and Batra, Dhruv , title =. Advances in Neural Information Processing Systems Datasets and Benchmarks Trac...

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Savva, Manolis and Kadian, Abhishek and Maksymets, Oleksandr and Zhao, Yili and Wijmans, Erik and Jain, Bhavana and Straub, Julian and Liu, Jia and Koltun, Vladlen and Malik, Jitendra and Parikh, Devi and Batra, Dhruv , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2019 , doi =

    2019

  7. [7]

    2025 , eprint =

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and others , title =. 2025 , eprint =. doi:10.48550/arXiv.2511.21631 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631 2025

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Ramrakhya, Ram and Undersander, Eric and Batra, Dhruv and Das, Abhishek , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  9. [9]

    Workshop on Reincarnating Reinforcement Learning at ICLR , year =

    Yadav, Karmesh and Ramrakhya, Ram and Majumdar, Arjun and Berges, Vincent-Pierre and Kuhar, Sachit and Batra, Dhruv and Baevski, Alexei and Maksymets, Oleksandr , title =. Workshop on Reincarnating Reinforcement Learning at ICLR , year =

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Gadre, Samir Yitzhak and Wortsman, Mitchell and Ilharco, Gabriel and Schmidt, Ludwig and Song, Shuran , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  11. [11]

    IEEE International Conference on Robotics and Automation , pages =

    Yokoyama, Naoki and Ha, Sehoon and Batra, Dhruv and Wang, Jiuguang and Bucher, Bernadette , title =. IEEE International Conference on Robotics and Automation , pages =. 2024 , eprint =. doi:10.48550/arXiv.2312.03275 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.03275 2024

  12. [12]

    Proceedings of the International Conference on Neuro-symbolic Systems , series =

    Goetting, Dylan and Singh, Himanshu Gaurav and Loquercio, Antonio , title =. Proceedings of the International Conference on Neuro-symbolic Systems , series =. 2025 , eprint =. doi:10.48550/arXiv.2411.05755 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.05755 2025

  13. [13]

    arXiv preprint arXiv:2411.16425 , year =

    Zhong, Linqing and Gao, Chen and Ding, Zihan and Liao, Yue and Ma, Huimin and Zhang, Shifeng and Zhou, Xu and Liu, Si , title =. arXiv preprint arXiv:2411.16425 , year =. doi:10.48550/arXiv.2411.16425 , url =. 2411.16425 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.16425

  14. [14]

    Open Scene Graphs for Open-World Object-Goal Navigation

    Loo, Joel and Wu, Zhanxin and Hsu, David , title =. 2025 , eprint =. doi:10.48550/arXiv.2508.04678 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.04678 2025

  15. [15]

    arXiv preprint arXiv:2503.02247 , year =

    Nie, Dujun and Guo, Xianda and Duan, Yiqun and Zhang, Ruijun and Chen, Long , title =. arXiv preprint arXiv:2503.02247 , year =. doi:10.48550/arXiv.2503.02247 , url =. 2503.02247 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.02247

  16. [16]

    MerNav: A Highly Generalizable Memory-Execute-Review Framework for Zero-Shot Object Goal Navigation

    Qi, Dekang and Zeng, Shuang and Chang, Xinyuan and Xiong, Feng and Xie, Shichao and Wu, Xiaolong and Xu, Mu , title =. 2026 , eprint =. doi:10.48550/arXiv.2602.05467 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.05467 2026

  17. [17]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    An, Dong and Wang, Hanqing and Wang, Wenguan and Wang, Zun and Huang, Yan and He, Keji and Wang, Liang , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2025 , doi =

    2025

  18. [18]

    Advances in Neural Information Processing Systems , volume =

    Chaplot, Devendra Singh and Gandhi, Dhiraj and Gupta, Saurabh and Gupta, Abhinav and Salakhutdinov, Ruslan , title =. Advances in Neural Information Processing Systems , volume =

  19. [19]

    International Conference on Learning Representations , year =

    Savinov, Nikolay and Dosovitskiy, Alexey and Koltun, Vladlen , title =. International Conference on Learning Representations , year =

  20. [20]

    International Journal of Computer Vision , year =

    Zheng, Qi and Liu, Daqing and Wang, Chaoyue and Zhang, Jing and Wang, Dadong and Tao, Dacheng , title =. International Journal of Computer Vision , year =. doi:10.1007/s11263-024-02159-8 , eprint =

    work page doi:10.1007/s11263-024-02159-8

  21. [21]

    Advances in Neural Information Processing Systems , volume =

    Yin, Hang and Xu, Xiuwei and Wu, Zhenyu and Zhou, Jie and Lu, Jiwen , title =. Advances in Neural Information Processing Systems , volume =. 2024 , eprint =. doi:10.48550/arXiv.2410.08189 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.08189 2024

  22. [22]

    and Kemker, Ronald and Part, Jose L

    Parisi, German I. and Kemker, Ronald and Part, Jose L. and Kanan, Christopher and Wermter, Stefan , title =. Neural Networks , volume =. 2019 , doi =

    2019

  23. [23]

    Information Fusion , volume =

    Lesort, Timothee and Lomonaco, Vincenzo and Stoian, Andrei and Maltoni, Davide and Filliat, David and Diaz-Rodriguez, Natalia , title =. Information Fusion , volume =. 2020 , doi =

    2020

  24. [24]

    Proceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation , pages =

    Yamauchi, Brian , title =. Proceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation , pages =. 1997 , doi =

    1997

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Li, Hongxin and Wang, Zeyu and Yang, Xu and Yang, Yuran and Mei, Shuqi and Zhang, Zhaoxiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  26. [26]

    2025 , eprint =

    Zhang, Lingfeng and Liu, Yuecheng and Zhang, Zhanguang and Aghaei, Matin and Hu, Yaochen and Gu, Hongjian and Alomrani, Mohammad Ali and Bravo, David Gamaliel Arcos and Karimi, Raika and Hamidizadeh, Atia and Xu, Haoping and Huang, Guowei and Zhang, Zhanpeng and Cao, Tongtong and Qiu, Weichao and Quan, Xingyue and Hao, Jianye and Zhuang, Yuzheng and Zhang...

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.14254 2025

  27. [27]

    arXiv preprint arXiv:2506.15096 , year =

    Ji, Zihe and Lin, Huangxuan and Gao, Yue , title =. arXiv preprint arXiv:2506.15096 , year =. doi:10.48550/arXiv.2506.15096 , url =. 2506.15096 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.15096

  28. [28]

    Advances in Neural Information Processing Systems , year =

    Yu, Ming-Ming and Zhu, Fei and Liu, Wenzhuo and Yang, Yirong and Wang, Qunbo and Wu, Wenjun and Liu, Jing , title =. Advances in Neural Information Processing Systems , year =. doi:10.48550/arXiv.2510.20685 , url =. 2510.20685 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.20685

  29. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Khanna, Mukul and Ramrakhya, Ram and Chhablani, Gunjan and Yenamandra, Sriram and Gervet, Theophile and Chang, Matthew and Kira, Zsolt and Chaplot, Devendra Singh and Batra, Dhruv and Mottaghi, Roozbeh , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  30. [30]

    2026 , eprint =

    Niu, Haochen and Zhang, Lantao and Ji, Xingwu and Ying, Rendong and Liu, Peilin and Wen, Fei , title =. 2026 , eprint =. doi:10.48550/arXiv.2603.01813 , url =

    work page doi:10.48550/arxiv.2603.01813 2026

  31. [31]

    Proceedings of the International Conference on Machine Learning , year =

    Gao, Junyu and Yao, Xuan and Xu, Changsheng , title =. Proceedings of the International Conference on Machine Learning , year =. 2311.13209 , archivePrefix =

    Pith/arXiv arXiv

  32. [32]

    2025 , eprint =

    Ko, Heeju and Kim, Sungjune and Oh, Gyeongrok and Yoon, Jeongyoon and Lee, Honglak and Jang, Sujin and Kim, Seungryong and Kim, Sangpil , title =. 2025 , eprint =. doi:10.48550/arXiv.2506.06630 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.06630 2025

  33. [33]

    2023 , eprint =

    Yadav, Karmesh and Majumdar, Arjun and Ramrakhya, Ram and Yokoyama, Naoki and Baevski, Alexei and Kira, Zsolt and Maksymets, Oleksandr and Batra, Dhruv , title =. 2023 , eprint =. doi:10.48550/arXiv.2303.07798 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.07798 2023

  34. [34]

    Advances in Neural Information Processing Systems , year =

    Majumdar, Arjun and Aggarwal, Gunjan and Devnani, Bhavika and Hoffman, Judy and Batra, Dhruv , title =. Advances in Neural Information Processing Systems , year =. doi:10.48550/arXiv.2206.12403 , url =. 2206.12403 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.12403

  35. [35]

    European Conference on Computer Vision , year =

    Sun, Xinyu and Liu, Lizhao and Zhi, Hongyan and Qiu, Ronghe and Liang, Junwei , title =. European Conference on Computer Vision , year =. doi:10.48550/arXiv.2403.11650 , url =. 2403.11650 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.11650

  36. [36]

    2023 , eprint =

    Cai, Wenzhe and Huang, Siyuan and Cheng, Guangran and Long, Yuxing and Gao, Peng and Sun, Changyin and Dong, Hao , title =. 2023 , eprint =. doi:10.48550/arXiv.2309.10309 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.10309 2023

  37. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Zhang, Sixian and Yu, Xinyao and Song, Xinhang and Wang, Xiaohan and Jiang, Shuqiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  38. [38]

    International Conference on Machine Learning , year =

    Zhou, Kaiwen and Zheng, Kaizhi and Pryor, Connor and Shen, Yilin and Jin, Hongxia and Getoor, Lise and Wang, Xin Eric , title =. International Conference on Machine Learning , year =. doi:10.48550/arXiv.2301.13166 , url =. 2301.13166 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.13166

  39. [39]

    IEEE/RSJ International Conference on Intelligent Robots and Systems , year =

    Yu, Bangguo and Kasaei, Hamidreza and Cao, Ming , title =. IEEE/RSJ International Conference on Intelligent Robots and Systems , year =. doi:10.1109/IROS55552.2023.10342512 , eprint =

    arXiv 2023

  40. [40]

    2024 , eprint =

    Wu, Pengying and Mu, Yao and Wu, Bingxian and Hou, Yi and Ma, Ji and Zhang, Shanghang and Liu, Chang , title =. 2024 , eprint =. doi:10.48550/arXiv.2401.02695 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.02695 2024

  41. [41]

    Findings of the Association for Computational Linguistics: NAACL , year =

    Kuang, Yuxuan and Lin, Hai and Jiang, Meng , title =. Findings of the Association for Computational Linguistics: NAACL , year =. doi:10.48550/arXiv.2402.10670 , url =. 2402.10670 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.10670

  42. [42]

    2025 , eprint =

    Cai, Yuxin and He, Xiangkun and Wang, Maonan and Guo, Hongliang and Yau, Wei-Yun and Lv, Chen , title =. 2025 , eprint =. doi:10.48550/arXiv.2504.09000 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.09000 2025

  43. [43]

    2026 , eprint =

    James, MoniJesu and Habel, Amir Atef and Fedoseev, Aleksey and Tsetserokou, Dzmitry , title =. 2026 , eprint =. doi:10.48550/arXiv.2603.18210 , url =

    work page doi:10.48550/arxiv.2603.18210 2026

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.