Pith. sign in

REVIEW 1 major objections 60 references

Four specialized agents collaborate through message passing to keep vision-language models aligned on out-of-distribution concepts.

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 Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-05-16 15:20 UTC

load-bearing objection MACL adds a four-agent message-passing setup for OOD vision-language robustness but the 1-5% gains lack the controls needed to separate signal from coordination noise. the 1 major comments →

arxiv 2601.09746 v1 submitted 2026-01-11 cs.MA

Multi-Agent Cooperative Learning for Robust Vision-Language Alignment under OOD Concepts

Philip Xu This is my paper
classification cs.MA
keywords multi-agent cooperative learningvision-language alignmentout-of-distribution conceptsmodality imbalancemessage passingfew-shot learningcross-modal alignment
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 paper introduces a multi-agent framework to stop vision-language models from losing cross-modal alignment when they meet unfamiliar concepts. Four agents specialized in images, text, naming, and coordination exchange messages and use an adaptive balancing mechanism to correct modality imbalances. The system supports feature-space name learning and a context-exchange method for few-shot scenarios. Tests on the VISTA-Beyond dataset show 1-5 percent precision gains in both few-shot and zero-shot regimes across varied visual domains.

Core claim

The central claim is that structuring vision-language alignment as cooperative learning among four agents—image, text, name, and coordination—mitigates modality imbalance for out-of-distribution concepts via structured message passing, multi-agent feature space name learning, context-exchange few-shot learning, and adaptive dynamic balancing.

What carries the argument

The four-agent message-passing architecture with adaptive dynamic balancing that regulates inter-agent contributions to mitigate modality imbalance.

Load-bearing premise

The four-agent message-passing architecture with adaptive dynamic balancing reliably mitigates modality imbalance for out-of-distribution concepts without introducing coordination overhead or instability.

What would settle it

Running the same vision-language model on the VISTA-Beyond dataset with and without the four-agent message-passing system and checking whether the reported 1-5 percent precision gains on out-of-distribution concepts disappear.

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

If this is right

  • Enables robust cross-modal alignment in few-shot and zero-shot settings for out-of-distribution concepts.
  • Delivers 1-5 percent precision improvements across diverse visual domains on the VISTA-Beyond dataset.
  • Supports multi-agent feature space name learning through coordinated agent interactions.
  • Regulates modality contributions dynamically without full model retraining.

Where Pith is reading between the lines

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

  • The coordination agent could be adapted to other multi-modal tasks where one modality tends to dominate.
  • Small gains on a single dataset suggest testing the framework on larger-scale vision-language benchmarks to confirm scalability.
  • If message passing proves stable, the approach might lower the data volume needed for aligning new concepts.

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

1 major / 0 minor

Summary. The paper introduces a Multi-Agent Cooperative Learning (MACL) framework with four agents (image, text, name, coordination) that use structured message passing, multi-agent feature space name learning, context exchange enhanced few-shot learning, and an adaptive dynamic balancing mechanism to mitigate modality imbalance and cross-modal alignment collapse in vision-language models for out-of-distribution concepts. Experiments on the VISTA-Beyond dataset report 1-5% precision gains in both few-shot and zero-shot settings across diverse visual domains.

Significance. If the reported gains prove reliable under proper statistical controls, the multi-agent architecture could provide a useful approach for handling OOD concepts in VLMs by addressing modality imbalance through coordination, extending existing alignment techniques in a structured way.

major comments (1)
  1. [Experiments] The central empirical claim rests on 1-5% precision gains on VISTA-Beyond, yet no error bars, standard deviations, number of random seeds, run counts, or statistical significance tests are provided. In a four-agent message-passing system with adaptive dynamic balancing, inter-agent weighting variations can introduce substantial run-to-run variance that may exceed the reported delta, preventing distinction between signal and noise.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment on experimental validation below and will incorporate the suggested improvements in the revised version.

read point-by-point responses

  1. Referee: [Experiments] The central empirical claim rests on 1-5% precision gains on VISTA-Beyond, yet no error bars, standard deviations, number of random seeds, run counts, or statistical significance tests are provided. In a four-agent message-passing system with adaptive dynamic balancing, inter-agent weighting variations can introduce substantial run-to-run variance that may exceed the reported delta, preventing distinction between signal and noise.

    Authors: We agree that the current presentation lacks sufficient statistical rigor to substantiate the reported gains, particularly given the potential for variance introduced by the adaptive balancing and message-passing mechanisms. In the revised manuscript, we will rerun all experiments using at least five independent random seeds, report mean performance with standard deviations, include error bars in figures and tables, and conduct paired statistical significance tests (e.g., t-tests) against baselines to confirm that the 1-5% improvements are reliable and not attributable to noise. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical result on external dataset with no self-referential derivations

full rationale

The paper introduces a multi-agent framework (MACL) with four agents and reports precision gains on the VISTA-Beyond dataset in few-shot and zero-shot settings. No equations, fitted parameters, or derivation steps appear in the provided text that reduce by construction to self-defined inputs, self-citations, or renamed known results. The central claim is a direct experimental outcome on an independent external benchmark, satisfying the criteria for a self-contained non-circular finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; all such elements remain unknown.

pith-pipeline@v0.9.0 · 5393 in / 1021 out tokens · 42213 ms · 2026-05-16T15:20:09.214866+00:00 · methodology

0 comments
read the original abstract

This paper introduces a novel Multi-Agent Cooperative Learning (MACL) framework to address cross-modal alignment collapse in vision-language models when handling out-of-distribution (OOD) concepts. Four core agents, including image, text, name, and coordination agents, collaboratively mitigate modality imbalance through structured message passing. The proposed framework enables multi-agent feature space name learning, incorporates a context exchange enhanced few-shot learning algorithm, and adopts an adaptive dynamic balancing mechanism to regulate inter-agent contributions. Experiments on the VISTA-Beyond dataset demonstrate that MACL significantly improves performance in both few-shot and zero-shot settings, achieving 1-5% precision gains across diverse visual domains.

Figures

Figures reproduced from arXiv: 2601.09746 by Philip Xu.

Figure 1
Figure 1. Figure 1: Cross-Dataset Generalization Results addressed in this work. 2.2. Few-shot and Zero-shot Learning Methods In order to improve the adaptability of visual language mod￾els in data-scarce scenarios, various few-shot and zero-shot learning methods have been proposed, such as FLAN [37] et al.[24, 39] have been proposed. Approaches such as CoOp[60] and CoCoOp[59] achieve domain-specific adap￾tation by learning t… view at source ↗
Figure 2
Figure 2. Figure 2: Cross-Modal Alignment Visualization main adaptation, concept generalization, and representa￾tion calibration. Methods[36, 58] such as DAPT [5] and AdaVLT [25, 47, 49] mitigate domain shift through domain￾specific fine-tuning strategies, while WiSE-FT [38, 48, 50] proposes a weight-space ensemble approach to balance pre￾trained knowledge with domain-specific information. In the realm of concept generalizati… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of models’ accuracy on various OOD [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of models’ accuracy on various SC [PITH_FULL_IMAGE:figures/full_fig_p007_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.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages

  1. [1]

    LLM2Vec: Large language models are secretly powerful text encoders

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. LLM2Vec: Large language models are secretly powerful text encoders. InFirst Conference on Language Modeling, 2024. 1

    work page 2024

  2. [2]

    Internvl: Scal- ing up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scal- ing up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 24185–24198, 2024. 1

    work page 2024

  3. [3]

    Reproducible scaling laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), page 2818–2829. IEEE, 2023. 2, 6

    work page 2023

  4. [4]

    The case for human–AI interaction as system 0 thinking.Nature Human Behaviour, 8(10):1829–1830, 2024

    Massimo Chiriatti, Marianna Ganapini, Enrico Panai, Mario Ubiali, and Giuseppe Riva. The case for human–AI interaction as system 0 thinking.Nature Human Behaviour, 8(10):1829–1830, 2024. 1

    work page 2024

  5. [5]

    Distribution-aware prompt tuning for vision-language models

    Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 22004–22013, 2023. 3

    work page 2023

  6. [6]

    Exploring large language models for multi-modal out-of-distribution detection, 2023

    Yi Dai, Hao Lang, Kaisheng Zeng, Fei Huang, and Yongbin Li. Exploring large language models for multi-modal out-of-distribution detection, 2023. 1

    work page 2023

  7. [7]

    Qlora: Efficient finetuning of quan- tized llms, 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quan- tized llms, 2023. 1

    work page 2023

  8. [8]

    Unveiling encoder- free vision-language models, 2024

    Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder- free vision-language models, 2024. 1

    work page 2024

  9. [9]

    A survey on in-context learning,

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiy- ong Wu, Tianyu Liu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. A survey on in-context learning,

    work page

  10. [10]

    CCG: Rare-label prediction via neu- ral SEM–driven causal game

    Yijia Fan, Jusheng Zhang, Kaitong Cai, Jing Yang, and Keze Wang. CCG: Rare-label prediction via neu- ral SEM–driven causal game. InFindings of the Asso- ciation for Computational Linguistics: EMNLP 2025, pages 6243–6256, Suzhou, China, 2025. Association for Computational Linguistics. 1

    work page 2025

  11. [11]

    Clip-adapter: Better vision-language models with feature adapters,

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters.arXiv preprint arXiv:2110.04544, 2021. 2, 6

    work page arXiv 2021

  12. [12]

    Context encoders as a simple but powerful extension of word2vec, 2017

    Franziska Horn. Context encoders as a simple but powerful extension of word2vec, 2017. 1

    work page 2017

  13. [13]

    Parameter-efficient transfer learning for nlp, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzeb- ski, Bruna Morrone, Quentin de Laroussilhe, An- drea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp, 2019. 1

    work page 2019

  14. [14]

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

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large lan- guage models, 2021. 1

    work page 2021

  15. [15]

    Le, Yunhsuan Sung, Zhen Li, and Tom Duerig

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V . Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. 2

    work page 2021

  16. [16]

    Visual prompt tuning, 2022

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser- Nam Lim. Visual prompt tuning, 2022. 1

    work page 2022

  17. [17]

    The power of scale for parameter-efficient prompt tuning,

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning,

    work page

  18. [18]

    Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi

    Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation, 2021. 2

    work page 2021

  19. [19]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and gener- ation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and gener- ation. InICML, 2022. 1

    work page 2022

  20. [20]

    Im- proving the accuracy of diabetes diagnosis applica- tions through a hybrid feature selection algorithm

    Xiaohua Li, Jusheng Zhang, and Fatemeh Safara. Im- proving the accuracy of diabetes diagnosis applica- tions through a hybrid feature selection algorithm. Neural Process. Lett., 55(1):153–169, 2021. 1

    work page 2021

  21. [21]

    Distilling large vision- language model with out-of-distribution generalizabil- ity, 2023

    Xuanlin Li, Yunhao Fang, Minghua Liu, Zhan Ling, Zhuowen Tu, and Hao Su. Distilling large vision- language model with out-of-distribution generalizabil- ity, 2023. 1

    work page 2023

  22. [22]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1

    work page 2023

  23. [23]

    Towards out-of- distribution generalization: A survey, 2023

    Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of- distribution generalization: A survey, 2023. 2

    work page 2023

  24. [24]

    Mta: Multimodal task alignment for bev perception and captioning, 2024

    Yunsheng Ma, Burhaneddin Yaman, Xin Ye, Feng Tao, Abhirup Mallik, Ziran Wang, and Liu Ren. Mta: Multimodal task alignment for bev perception and captioning, 2024. 2

    work page 2024

  25. [25]

    Adavit: Adaptive vision transformers for efficient im- age recognition, 2021

    Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient im- age recognition, 2021. 3

    work page 2021

  26. [26]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024. 1

    work page 2024

  27. [27]

    Learning to name classes for vision and language models, 2023

    Sarah Parisot, Yongxin Yang, and Steven McDon- agh. Learning to name classes for vision and language models, 2023. 6

    work page 2023

  28. [28]

    Qwen2.5 technical report, 2025

    Qwen. Qwen2.5 technical report, 2025. 1

    work page 2025

  29. [29]

    Learning trans- ferable visual models from natural language supervi- sion, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning trans- ferable visual models from natural language supervi- sion, 2021. 1

    work page 2021

  30. [30]

    Denseclip: Language-guided dense predic- tion with context-aware prompting

    Yongming Rao, Wenliang Zhao, Guangyi Chen, Yan- song Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense predic- tion with context-aware prompting. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1

    work page 2022

  31. [31]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021

    Christoph Schuhmann, Richard Vencu, Romain Beau- mont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komat- suzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. 3

    work page 2021

  32. [32]

    How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021

    Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip ben- efit vision-and-language tasks?arXiv preprint arXiv:2107.06383, 2021. 3

    work page arXiv 2021

  33. [33]

    Human cognition involves the dynamic integra- tion of neural activity and neuromodulatory systems

    James M Shine, Michael Breakspear, Peter T Bell, Kaylena A Ehgoetz Martens, Richard Shine, Oluwasanmi Koyejo, Olaf Sporns, and Russell A Pol- drack. Human cognition involves the dynamic integra- tion of neural activity and neuromodulatory systems. Nature Neuroscience, 22(2):289–296, 2019. 1

    work page 2019

  34. [34]

    Fastcontext: Han- dling out-of-vocabulary words using the word struc- ture and context

    Renato M Silva, Johannes V Lochter, Tiago A Almeida, and Akebo Yamakami. Fastcontext: Han- dling out-of-vocabulary words using the word struc- ture and context. InBrazilian Conference on Intelli- gent Systems, pages 539–557. Springer, 2022. 1

    work page 2022

  35. [35]

    Out- of-distribution generalization via composition: a lens through induction heads in transformers, 2024

    Jiajun Song, Zhuoyan Xu, and Yiqiao Zhong. Out- of-distribution generalization via composition: a lens through induction heads in transformers, 2024. 2

    work page 2024

  36. [36]

    Robust fine-tuning of vision-language models for domain generalization,

    Kevin V ogt-Lowell, Noah Lee, Theodoros Tsiligkaridis, and Marc Vaillant. Robust fine-tuning of vision-language models for domain generalization,

    work page

  37. [37]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, An- drew M

    Jason Wei, Maarten Bosma, Vincent Y . Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, An- drew M. Dai, and Quoc V . Le. Finetuned language models are zero-shot learners, 2022. 2

    work page 2022

  38. [38]

    Robust fine-tuning of zero-shot models, 2022

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models, 2022. 3

    work page 2022

  39. [39]

    Cascade prompt learning for vision-language model adaptation, 2024

    Ge Wu, Xin Zhang, Zheng Li, Zhaowei Chen, Jia- jun Liang, Jian Yang, and Xiang Li. Cascade prompt learning for vision-language model adaptation, 2024. 2

    work page 2024

  40. [40]

    Depthssc: Monocular 3d seman- tic scene completion via depth-spatial alignment and voxel adaptation, 2024

    Jiawei Yao, Jusheng Zhang, Xiaochao Pan, Tong Wu, and Canran Xiao. Depthssc: Monocular 3d seman- tic scene completion via depth-spatial alignment and voxel adaptation, 2024. 1

    work page 2024

  41. [41]

    Task residual for tuning vision-language models

    Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xin- chao Wang. Task residual for tuning vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899–10909, 2023. 6

    work page 2023

  42. [42]

    Boosting vision-language models with transduction,

    Maxime Zanella, Beno ˆıt G´erin, and Ismail Ben Ayed. Boosting vision-language models with transduction,

    work page

  43. [43]

    MAT-agent: Adaptive multi- agent training optimization

    Jusheng Zhang, Kaitong Cai, Yijia Fan, Ningyuan Liu, and Keze Wang. MAT-agent: Adaptive multi- agent training optimization. InThe Thirty-ninth An- nual Conference on Neural Information Processing Systems, 2025. 1

    work page 2025

  44. [44]

    CF-VLM: Counterfactual vision- language fine-tuning

    Jusheng Zhang, Kaitong Cai, Yijia Fan, Jian Wang, and Keze Wang. CF-VLM: Counterfactual vision- language fine-tuning. InThe Thirty-ninth Annual Con- ference on Neural Information Processing Systems,

    work page

  45. [45]

    Mm-cot:a benchmark for probing visual chain- of-thought reasoning in multimodal models, 2025

    Jusheng Zhang, Kaitong Cai, Xiaoyang Guo, Sidi Liu, Qinhan Lv, Ruiqi Chen, Jing Yang, Yijia Fan, Xiaofei Sun, Jian Wang, Ziliang Chen, Liang Lin, and Keze Wang. Mm-cot:a benchmark for probing visual chain- of-thought reasoning in multimodal models, 2025. 1

    work page 2025

  46. [46]

    Top-down semantic refinement for image captioning, 2025

    Jusheng Zhang, Kaitong Cai, Jing Yang, Jian Wang, Chengpei Tang, and Keze Wang. Top-down semantic refinement for image captioning, 2025. 1

    work page 2025

  47. [47]

    Learning dynamics of vlm finetuning, 2025

    Jusheng Zhang, Kaitong Cai, Jing Yang, and Keze Wang. Learning dynamics of vlm finetuning, 2025. 3

    work page 2025

  48. [48]

    Failure-driven workflow refinement, 2025

    Jusheng Zhang, Kaitong Cai, Qinglin Zeng, Ningyuan Liu, Stephen Fan, Ziliang Chen, and Keze Wang. Failure-driven workflow refinement, 2025. 3

    work page 2025

  49. [49]

    Drdiff: Dynamic routing diffusion with hierarchical attention for breaking the efficiency- quality trade-off, 2025

    Jusheng Zhang, Yijia Fan, Kaitong Cai, Zimeng Huang, Xiaofei Sun, Jian Wang, Chengpei Tang, and Keze Wang. Drdiff: Dynamic routing diffusion with hierarchical attention for breaking the efficiency- quality trade-off, 2025. 3

    work page 2025

  50. [50]

    Osc: Cognitive orchestration through dynamic knowledge alignment in multi-agent llm col- laboration, 2025

    Jusheng Zhang, Yijia Fan, Kaitong Cai, Xiaofei Sun, and Keze Wang. Osc: Cognitive orchestration through dynamic knowledge alignment in multi-agent llm col- laboration, 2025. 3

    work page 2025

  51. [51]

    Kolmogorov-arnold fourier networks, 2025

    Jusheng Zhang, Yijia Fan, Kaitong Cai, and Keze Wang. Kolmogorov-arnold fourier networks, 2025. 1

    work page 2025

  52. [52]

    GAM-agent: Game-theoretic and uncertainty- aware collaboration for complex visual reasoning

    Jusheng Zhang, Yijia Fan, Wenjun Lin, Ruiqi Chen, Haoyi Jiang, Wenhao Chai, Jian Wang, and Keze Wang. GAM-agent: Game-theoretic and uncertainty- aware collaboration for complex visual reasoning. In The Thirty-ninth Annual Conference on Neural Infor- mation Processing Systems, 2025. 1

    work page 2025

  53. [53]

    Tri-MARF: A tri-modal multi-agent re- sponsive framework for comprehensive 3d object an- notation

    Jusheng Zhang, Yijia Fan, Zimo Wen, Jian Wang, and Keze Wang. Tri-MARF: A tri-modal multi-agent re- sponsive framework for comprehensive 3d object an- notation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 1

    work page 2025

  54. [54]

    Hybridtoken-vlm: Hybrid token compression for vision-language models, 2025

    Jusheng Zhang, Xiaoyang Guo, Kaitong Cai, Qinhan Lv, Yijia Fan, Wenhao Chai, Jian Wang, and Keze Wang. Hybridtoken-vlm: Hybrid token compression for vision-language models, 2025. 1

    work page 2025

  55. [55]

    KABB: Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems

    Jusheng Zhang, Zimeng Huang, Yijia Fan, Ningyuan Liu, Mingyan Li, Zhuojie Yang, Jiawei Yao, Jian Wang, and Keze Wang. KABB: Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems. InForty-second International Conference on Machine Learning, 2025. 1

    work page 2025

  56. [56]

    Llm-cas: Dynamic neuron perturbation for real-time hallucination correction, 2025

    Jensen Zhang, Ningyuan Liu, Yijia Fan, Zihao Huang, Qinglin Zeng, Kaitong Cai, Jian Wang, and Keze Wang. Llm-cas: Dynamic neuron perturbation for real-time hallucination correction, 2025. 1

    work page 2025

  57. [57]

    Llava-read: Enhanc- ing reading ability of multimodal language models,

    Ruiyi Zhang, Yufan Zhou, Jian Chen, Jiuxiang Gu, Changyou Chen, and Tong Sun. Llava-read: Enhanc- ing reading ability of multimodal language models,

    work page

  58. [58]

    Learn- ing from models beyond fine-tuning.Nature Machine Intelligence, 7(1):6–17, 2025

    Hongling Zheng, Li Shen, Anke Tang, Yong Luo, Han Hu, Bo Du, Yonggang Wen, and Dacheng Tao. Learn- ing from models beyond fine-tuning.Nature Machine Intelligence, 7(1):6–17, 2025. 3

    work page 2025

  59. [59]

    Conditional prompt learning for vision- language models, 2022

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision- language models, 2022. 2, 6

    work page 2022

  60. [60]

    Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022. 2, 6

    work page 2022