Pith. sign in

REVIEW 3 major objections 5 minor 47 references

A training-free multi-agent pipeline that first judges expression difficulty and dominant modality, then reasons and reflects, lifts Ref-AVS to 69.2% J&F.

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 ↗

T0 review · grok-4.5

2026-07-15 11:45 UTC pith:FFNP6XBY

load-bearing objection Training-free multi-agent Ref-AVS that posts a real +3.4 J&F SOTA; the adaptive routing rests on an unvalidated hand taxonomy, but the ablations and reflection loop still hold up as engineering. the 3 major comments →

arxiv 2603.27706 v2 pith:FFNP6XBY submitted 2026-03-29 cs.MM

MAR3: Multi-Agent Recognition, Reasoning, and Reflection for Reference Audio-Visual Segmentation

Yuan Zhao , Zhenqi Jia , Yongqiang Zhang This is my paper
classification cs.MM
keywords Reference Audio-Visual SegmentationMultimodal ReasoningMulti-Agent SystemsDelphi consensusReflective segmentationTraining-free
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.

Reference Audio-Visual Segmentation asks a system to cut out the object that a spoken or written phrase points to inside a video that also has sound. Prior systems either fuse modalities without saying which cue should lead, or they fine-tune large models on hand-built instruction data and then segment once without checking. This paper claims that those gaps cause wrong masks. It offers MAR3, a training-free three-stage multi-agent workflow: agents first reach consensus on how hard the expression is and which modality dominates, then reason about the object with difficulty-aware collaboration, then a check agent inspects the draft mask and rewrites the text prompt for the segmenter until the mask matches the expression. On the public Ref-AVSBench test set the full pipeline reaches 69.2% J&F, 3.4 points above the previous best, and ablations show each stage contributes measurable gains.

Core claim

The paper establishes that a training-free multi-agent Recognition–Reasoning–Reflection pipeline—explicit consensus on expression difficulty and dominant modality via a Delphi-style process, adaptive collaboration between dominant- and auxiliary-modality agents under a modality-dominant difficulty rule, and iterative prompt correction by a check agent—produces higher-quality referred-object masks on audible video than prior fusion or instruction-tuned single-stage methods, reaching 69.2% J&F on Ref-AVSBench.

What carries the argument

MAR3’s three mechanisms: Consensus Multimodal Recognition (Delphi-style independent thinking, peer interaction, final decision among multimodal-cue agents), Collaborative Object Reasoning (routing that uses the authors’ modality-dominant difficulty rule so low-difficulty cases stay single-agent and moderate/high cases fuse auxiliary candidate lists), and Reflective Learning Segmentation (check agent rewrites the object text prompt of a segment agent for up to two iterations).

Load-bearing premise

The hand-defined rule that maps “low / moderate / high” difficulty exactly onto whether one modality dominates, one is auxiliary, or both dominate is assumed to be a faithful proxy for real expression hardness and the right trigger for multi-agent collaboration.

What would settle it

Replace the authors’ modality-dominant difficulty taxonomy with an independent human difficulty rating or a different automatic classifier, re-run Collaborative Object Reasoning with that routing, and check whether the 1.3-point J and F gains attributed to collaboration and the overall 69.2% J&F still appear on the same Ref-AVSBench Mix (S+U) split.

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

If this is right

  • If the three-stage pipeline is correct, Ref-AVS systems no longer need large instruction-tuning datasets or joint end-to-end training to reach state-of-the-art mask quality.
  • Explicit recognition of dominant modality becomes a necessary first step rather than an implicit byproduct of multimodal fusion.
  • A small number of reflective prompt corrections (here N=2) can close the gap left by single-stage segmentation without requiring re-localization.
  • The same Recognition–Reasoning–Reflection skeleton can be reused for other multimodal referring tasks that mix language, vision, and audio.

Where Pith is reading between the lines

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

  • The large share of high-difficulty / audio+visual-dominant expressions the authors measure on the test set suggests that many real-world referring phrases will keep needing dual-modality collaboration rather than collapsing to a single cue.
  • Because the method is training-free and agent-based, swapping newer open multimodal models into the same roles should raise the ceiling without re-collecting data.
  • The authors flag the unsolved “null” (no-object) case; the check-agent stage is a natural place to add an explicit “none” decision, which would extend the pipeline to open-world referring.
  • The performance edge is larger on Seen than Unseen categories, hinting that the consensus and reflection stages help most when the object vocabulary is familiar but the expression is hard.

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 / 5 minor

Summary. The paper proposes MAR3, a training-free multi-agent framework for Reference Audio-Visual Segmentation (Ref-AVS) that decomposes the task into Consensus Multimodal Recognition (CMR), Collaborative Object Reasoning (COR), and Reflective Learning Segmentation (RLS). CMR uses heterogeneous LLM agents and a Delphi-style independent-thinking / peer-interaction / final-decision process to label each reference expression with a difficulty level and dominant/auxiliary modality under an author-defined modality-dominant difficulty rule (Low = single dominant modality; Moderate = one dominant + one auxiliary; High = both modalities dominant). COR then routes reasoning accordingly (independent dominant-modality agent for Low; auxiliary candidate lists fed to the dominant agent for Moderate/High). RLS has a check agent inspect intermediate SAM3 masks and iteratively revise the object text prompt (max N=2). On Ref-AVSBench, MAR3 reports 69.2% J&F on Mix (S+U), +3.4 absolute over TGS-Agent, with ablations (Tables 2–3) and qualitative examples supporting each module.

Significance. If the empirical gains hold under fairer controls, the work is a useful contribution to multimodal reasoning and multi-agent systems for Ref-AVS: it is training-free, decomposes a tightly coupled audio-visual-language-pixel task into specialized agents, and shows that reflective prompt correction can recover from wrong initial object names. Strengths include clear module ablations (Table 2), a controlled study of reflection steps N (Table 3, best at N=2), and qualitative traces that make the pipeline inspectable. The Delphi-inspired consensus step and the check–segment loop are concrete, reusable design patterns. The result is of interest to the audio-visual segmentation and multi-agent communities even if some design choices remain under-validated.

major comments (3)
  1. [§2.2–2.3, Table 2, Figures 4–5] §2.2–2.3 and Table 2: The adaptive COR path is gated by the hand-defined modality-dominant difficulty rule (Low / Moderate / High). CMR only classifies expressions into this taxonomy; it does not validate that the labels match human difficulty/modality judgments or that the prescribed collaboration is the right intervention. The “w/o COR” ablation collapses collaboration into pure dominant-modality reasoning (+1.3 J on Mix) but does not test always-collaborate, never-collaborate, random, or inverted routing. Figures 4–5 report only the distribution of CMR’s own labels (e.g., 66.8% High / Audio+Visual on Mix), which cannot establish that the taxonomy is faithful. Without such controls, the claim that modality-dominant adaptive collaboration is responsible for part of the SOTA margin is under-supported.
  2. [Table 1, Abstract] Table 1: On Unseen, MAR3 is slightly below TGS-Agent (76.2* vs 76.7 J&F; asterisked in the paper). The abstract and main claim emphasize “superior performance … outperforming SOTA by 3.4%” on Mix. The manuscript should either qualify the SOTA claim by subset, report variance / multiple seeds / statistical tests, or analyze why the training-free pipeline does not transfer as strongly to Unseen categories. As written, the headline number overstates robustness across the official splits.
  3. [§3.1 Dataset, §4 Conclusion] §3.1 and Conclusion: Null expressions (referents absent from the video) are excluded from evaluation, and the conclusion notes the method was not designed for them. Because Ref-AVSBench includes a Null set and real deployments must handle empty referents, the paper should either evaluate a simple “no object” decision path (e.g., check-agent rejection) or clearly bound the claim to non-null expressions rather than leaving it only as future work.
minor comments (5)
  1. [Table 1] Table 1: Crab [6] reports incomplete metrics (only Mix J&F-style numbers partially filled). Either complete the row from the original paper or mark missing entries consistently so the comparison table is readable.
  2. [§3.1] §3.1 Implementation Details: Agent–model assignments (Qwen3-30B-A3B, GLM4-32B, Deepseek-R1-32B, Qwen3-Omni, SAM3) and FPS/audio rate are free parameters. A short sensitivity note (or appendix) on swapping one MCA model would strengthen reproducibility claims for a multi-agent system.
  3. [Figure 3] Figure 3 and qualitative text: Intermediate COR outputs sometimes still name the wrong object before RLS corrects it; making the “Match: False → Revised” step more visually prominent (e.g., side-by-side initial vs revised masks) would help readers see the reflection gain without reading long agent transcripts.
  4. [Throughout / §2.2] Typos / polish: “arXiv Preprint, March, 2026” headers; occasional spacing issues (“Ref-A VS”, “J &F ”); “modality-dominant difficulty rule” is introduced before being fully formalized—add a short numbered definition box in §2.2 for clarity.
  5. [§1.2] Related Work §1.2: Multi-agent citations are broad; a tighter comparison to other training-free agent pipelines for grounding/segmentation (beyond TGS-Agent) would better position the novelty of Delphi-style CMR + reflective segmentation.

Circularity Check

0 steps flagged

No circular derivation: SOTA J&F is an external held-out mask metric, not a quantity forced by the authors' difficulty taxonomy or self-citation.

full rationale

MAR3 is a training-free multi-agent engineering paper. Its central claim (69.2% J&F on Ref-AVSBench Mix, +3.4 over TGS-Agent) is produced by running the full pipeline and scoring predicted masks against independent pixel annotations; nothing in that number is algebraically or statistically forced by the modality-dominant difficulty rule of §2.2. That rule is an explicit design taxonomy (Low = single dominant modality; Moderate = dominant+auxiliary; High = both dominant) used to route COR, not a fitted parameter re-labeled as a prediction, nor a uniqueness theorem imported from the same authors. CMR classifies expressions into the taxonomy via Delphi-style multi-agent consensus citing external sociological sources (Linstone/Turoff et al.), then COR and RLS act; ablations (Table 2–3) remove modules and re-measure the same external J/F, rather than re-deriving a fitted quantity. Figures 4–5 only report the distribution of CMR’s own labels for analysis; they are not offered as a first-principles derivation of the SOTA margin. No self-definitional loop, no fit-then-predict, no load-bearing self-citation uniqueness claim, and no renaming of a known closed-form result. Any concern that the hand-crafted taxonomy is unvalidated is a correctness/specification risk, not circularity of the reported derivation chain.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The central performance claim rests on (1) an author-defined difficulty/modality taxonomy, (2) the assumption that heterogeneous off-the-shelf MLLMs plus Delphi interaction produce reliable labels, (3) fixed hyper-parameters (reflection steps, model choices), and (4) the invented pipeline modules themselves. No physical constants; free parameters are engineering knobs; invented entities are the three mechanisms and the difficulty rule.

free parameters (3)
  • max_reflective_learning_steps_N = 2
    Chosen by ablation (Table 3); N=2 is best on the test set and is used for the headline number.
  • MCA_agent_model_assignment = Qwen3-30B-A3B, GLM4-32B, Deepseek-R1-32B, Qwen3-Omni-30B, SAM3
    Qwen3-30B / GLM4-32B / Deepseek-R1-32B for recognition agents and Qwen3-Omni / SAM3 for reasoning and segmentation are hand-selected; results depend on these choices.
  • video_FPS_and_audio_rate = FPS=1, 22050 Hz
    FPS=1 and 22050 Hz are fixed preprocessing choices that affect what agents see/hear.
axioms (4)
  • ad hoc to paper The modality-dominant difficulty rule (Low / Moderate / High as defined in §2.2) correctly stratifies Ref-AVS expressions and should control collaboration.
    Author-defined taxonomy used to route COR; not derived from external theory beyond informal human analogy.
  • domain assumption Sociological Delphi consensus among heterogeneous LLM agents yields more robust difficulty and modality labels than a single agent.
    Imported from Delphi literature (§2.2) and supported only by the paper’s own ablation (w/o Delphi).
  • domain assumption Off-the-shelf MLLMs can name referred objects and judge mask–expression consistency without Ref-AVS fine-tuning.
    Core training-free premise; performance inherits whatever biases those models have.
  • domain assumption SAM3 produces usable masks from free-form object text prompts on Ref-AVS frames.
    Segment agent is treated as a black-box oracle.
invented entities (3)
  • Consensus Multimodal Recognition (CMR) with Delphi peer interaction no independent evidence
    purpose: Explicitly label expression difficulty and dominant/auxiliary modality before reasoning.
    New module; independent evidence limited to internal ablation.
  • Collaborative Object Reasoning (COR) under modality-dominant difficulty rule no independent evidence
    purpose: Adaptively fuse dominant and auxiliary modality agents for object naming.
    New routing strategy; evidence is ablation vs single dominant agent.
  • Reflective Learning Segmentation (RLS) check–segment loop no independent evidence
    purpose: Iteratively rewrite object text prompt from intermediate masks.
    New reflection loop for Ref-AVS; N chosen on test ablations.

pith-pipeline@v1.1.0-grok45 · 19668 in / 3457 out tokens · 31945 ms · 2026-07-15T11:45:27.858255+00:00 · methodology

0 comments
read the original abstract

Reference Audio-Visual Segmentation (Ref-AVS) aims to segment objects in audible videos based on multimodal cues in reference expressions. Previous methods overlook the explicit recognition of expression difficulty and dominant modality in multimodal cues, over-rely on the quality of the instruction-tuning dataset for object reasoning, and lack reflective validation of segmentation results, leading to erroneous mask predictions. To address these issues, in this paper, we propose a novel training-free Multi-Agent Recognition, Reasoning, and Reflection framework to achieve high-quality Reference Audio-Visual Segmentation, termed MAR3. Incorporating the sociological Delphi theory to achieve robust analysis, a Consensus Multimodal Recognition mechanism is proposed that enables LLM agents to explicitly recognize the difficulty of reference expressions and the dominant modality of multimodal cues. Based on our modality-dominant difficulty rule, we propose an adaptive Collaborative Object Reasoning strategy to reliably reason about the referred object. To further ensure precise mask prediction, we develop a Reflective Learning Segmentation mechanism, in which a check agent examines intermediate segmentation results and iteratively corrects the object text prompt of the segment agent. Experiments demonstrate that MAR3 achieves superior performance (69.2% in J&F) on the Ref-AVSBench dataset, outperforming SOTA by 3.4% absolutely.

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

47 extracted references · 1 canonical work pages

  1. [1]

    Maite Barrios, Georgina Guilera, Laura Nuño, and Juana Gómez-Benito. 2021. Consensus in the Delphi method: what makes a decision change?Technological Forecasting and Social Change163 (2021), 120484

    2021

  2. [2]

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, et al. 2025. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719(2025)

    Pith/arXiv arXiv 2025

  3. [3]

    Boyu Chen, Zhengrong Yue, Siran Chen, Zikang Wang, Yang Liu, Peng Li, and Yali Wang. 2025. Lvagent: Long video understanding by multi-round dynami- cal collaboration of mllm agents. InProceedings of the IEEE/CVF International Conference on Computer Vision. 20237–20246

    2025

  4. [4]

    Chieh-Yun Chen, Min Shi, Gong Zhang, and Humphrey Shi. 2025. T2i-copilot: A training-free multi-agent text-to-image system for enhanced prompt interpre- tation and interactive generation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 19396–19405

    2025

  5. [5]

    Yanbo Ding, Shaobin Zhuang, Kunchang Li, Zhengrong Yue, Yu Qiao, and Yali Wang. 2025. Muses: 3d-controllable image generation via multi-modal agent collaboration. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 2753–2761

    2025

  6. [6]

    Henghui Du, Guangyao Li, Chang Zhou, Chunjie Zhang, Alan Zhao, and Di Hu

  7. [7]

    InProceedings of the Computer Vision and Pattern Recognition Conference

    Crab: A unified audio-visual scene understanding model with explicit coop- eration. InProceedings of the Computer Vision and Pattern Recognition Conference. 18804–18814

  8. [8]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A fam- ily of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793(2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  10. [10]

    Shanshan Han, Qifan Zhang, Yuhang Yao, Weizhao Jin, and Zhaozhuo Xu

  11. [11]

    LLM multi-agent systems: Challenges and open problems.arXiv preprint arXiv:2402.03578(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Erik Hohmann, Mark P Cote, and Jefferson C Brand. 2018. Research pearls: expert consensus based evidence using the Delphi method.Arthroscopy: the journal of arthroscopic & related surgery34, 12 (2018), 3278–3282

    2018

  13. [13]

    Li Hu, Guoqiang Chen, Xiuwei Shang, Shaoyin Cheng, Benlong Wu, LiGangyang LiGangyang, Xu Zhu, Weiming Zhang, and Nenghai Yu. 2025. CompileAgent: Automated real-world repo-level compilation with tool-integrated LLM-based agent system. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2078–2091

    2025

  14. [14]

    Panwen Hu, Jin Jiang, Jianqi Chen, Mingfei Han, Shengcai Liao, Xiaojun Chang, and Xiaodan Liang. 2024. Storyagent: Customized storytelling video generation via multi-agent collaboration.arXiv preprint arXiv:2411.04925(2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Kaiyi Huang, Yukun Huang, Xuefei Ning, Zinan Lin, Yu Wang, and Xihui Liu. 2026. Genmac: compositional text-to-video generation with multi-agent collaboration. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 5049–5057

    2026

  16. [16]

    Yubin Kim, Chanwoo Park, Hyewon Jeong, Yik S Chan, Xuhai Xu, Daniel Mc- Duff, Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, and Hae W Park

  17. [17]

    Advances in Neural Information Processing Systems37 (2024), 79410–79452

    Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems37 (2024), 79410–79452

    2024

  18. [18]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  19. [19]

    InProceedings of the IEEE/CVF international conference on computer vision

    Segment anything. InProceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  20. [20]

    Mingcheng Li, Xiaolu Hou, Ziyang Liu, Dingkang Yang, Ziyun Qian, Jiawei Chen, Jinjie Wei, Yue Jiang, Qingyao Xu, and Lihua Zhang. 2025. MCCD: Multi-Agent Collaboration-based Compositional Diffusion for Complex Text-to-Image Gener- ation. InProceedings of the Computer Vision and Pattern Recognition Conference. 13263–13272

    2025

  21. [21]

    Xiang Li, Duyi Pan, Hongru Xiao, Jiale Han, Jing Tang, Jiabao Ma, Wei Wang, and Bo Cheng. 2025. Dialogueagents: A hybrid agent-based speech synthesis framework for multi-party dialogue. In2025 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

    2025

  22. [22]

    1975.The delphi method

    Harold A Linstone, Murray Turoff, et al . 1975.The delphi method. Vol. 1975. Addison-Wesley Reading, MA

    1975

  23. [23]

    Pei Liu, Xin Liu, Ruoyu Yao, Junming Liu, Siyuan Meng, Ding Wang, and Jun Ma. 2025. Hm-rag: Hierarchical multi-agent multimodal retrieval augmented generation. InProceedings of the 33rd ACM international conference on multimedia. 2781–2790

    2025

  24. [24]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. 2024. Grounding dino: Marry- ing dino with grounded pre-training for open-set object detection. InEuropean conference on computer vision. Springer, 38–55

    2024

  25. [25]

    Geliang Ouyang, Jingyao Chen, Zhihe Nie, Yi Gui, Yao Wan, Hongyu Zhang, and Dongping Chen. 2025. nvagent: Automated data visualization from natural language via collaborative agent workflow. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 19534–19567

    2025

  26. [26]

    Qi Peng, Jialin Cui, Jiayuan Xie, Yi Cai, and Qing Li. 2025. Tree-of-Reasoning: Towards Complex Medical Diagnosis via Multi-Agent Reasoning with Evidence Tree. InProceedings of the 33rd ACM International Conference on Multimedia. 1744–1753

    2025

  27. [27]

    Abduljalil Radman and Jorma Laaksonen. 2025. TSAM: Temporal SAM Aug- mented with Multimodal Prompts for Referring Audio-Visual Segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference. 23947– 23956

    2025

  28. [28]

    Wolf Rauch. 1979. The decision delphi.Technological forecasting and social change 15, 3 (1979), 159–169

    1979

  29. [29]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. 2024. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714(2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Yan Rong, Jinting Wang, Guangzhi Lei, Shan Yang, and Li Liu. 2025. Audiogenie: A training-free multi-agent framework for diverse multimodality-to-multiaudio generation. InProceedings of the 33rd ACM International Conference on Multimedia. 8872–8881

    2025

  31. [31]

    Haoyang Su, Renqi Chen, Shixiang Tang, Zhenfei Yin, Xinzhe Zheng, Jinzhe Li, Biqing Qi, Qi Wu, Hui Li, Wanli Ouyang, Philip Torr, Bowen Zhou, and Nanqing Dong. 2025. Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistic...

    doi:10.18653/v1/2025.acl-long.1368 2025

  32. [32]

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530(2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Qwen Team. 2025. Qwen3 Technical Report. arXiv:2505.09388 [cs.CL] https: //arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  34. [34]

    Murray Turoff. 1970. The design of a policy Delphi.Technological forecasting and social change2, 2 (1970), 149–171

    1970

  35. [35]

    Wiebe Van der Hoek and Michael Wooldridge. 2008. Multi-agent systems.Foun- dations of Artificial Intelligence3 (2008), 887–928

    2008

  36. [36]

    Haochun Wang, Sendong Zhao, Jingbo Wang, Zewen Qiang, Bing Qin, and Ting Liu. 2025. Beyond Frameworks: Unpacking Collaboration Strategies in Multi- Agent Systems. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (Ed...

    doi:10.18653/v1/ 2025

  37. [37]

    Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. 2024. Ref-avs: Refer and segment objects in audio-visual scenes. In European Conference on Computer Vision. Springer, 196–213

    2024

  38. [38]

    Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. 2025. SAM2- LOVE: Segment Anything Model 2 in Language-aided Audio-Visual Scenes. In Proceedings of the Computer Vision and Pattern Recognition Conference. 28932– 28941

    2025

  39. [39]

    Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang. 2024. Macrec: A multi-agent collaboration framework for recommendation. InProceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2760–2764

    2024

  40. [40]

    Dawei Xiang, Wenyan Xu, Kexin Chu, Tianqi Ding, Zixu Shen, Yiming Zeng, Jianchang Su, and Wei Zhang. 2025. Promptsculptor: Multi-agent based text-to- image prompt optimization. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 774–786

    2025

  41. [41]

    Fan Zhou, Bowen Yu, Jianxin Yang, Le Yu, Jingren Zhou, and Junyang Lin

    Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, Yuanjun Lv, Yongqi Wang, Dake Guo, He Wang, Linhan Ma, Pei Zhang, Xinyu Zhang, Hongkun Hao, Zishan Guo, Baosong Yang, Bin Zhang, Ziyang Ma, Xipin Wei, Shuai Bai, Keqin Chen, Xuejing Liu, Peng Wang, Mingkun Yang, Dayiheng Liu, Xingzhang Ren, Bo ...

    Pith/arXiv arXiv 2026

  42. [42]

    Zekang Yang, Wang Zeng, Sheng Jin, Chen Qian, Ping Luo, and Wentao Liu. 2025. Nader: Neural architecture design via multi-agent collaboration. InProceedings of the Computer Vision and Pattern Recognition Conference. 4452–4461

    2025

  43. [43]

    Kaining Ying, Henghui Ding, Guangquan Jie, and Yu-Gang Jiang. 2025. Towards omnimodal expressions and reasoning in referring audio-visual segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 22575– 22585

    2025

  44. [44]

    Muhammad Imran Yousuf. 2007. Using expertsopinions through Delphi technique. Practical assessment, research, and evaluation12, 1 (2007)

    2007

  45. [45]

    Yehang Zhang, Xinli Xu, Xiaojie Xu, Doudou Zhang, Li Liu, and Ying-Cong Chen. 2025. Orchestrating Audio: Multi-Agent Framework for Long-Video Audio Synthesis. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Christos Christodoulopoulos, Tanmoy Chakraborty, Car- olyn Rose, and Violet Peng (Eds.). Association for Com...

    work page doi:10.18653/v1/2025.emnlp-main.1133 2025

  46. [46]

    Jinxing Zhou, Yanghao Zhou, Mingfei Han, Tong Wang, Xiaojun Chang, Hisham Cholakkal, and Rao Muhammad Anwer. 2025. Think before you segment: An object-aware reasoning agent for referring audio-visual segmentation.arXiv preprint arXiv:2508.04418(2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Yucheng Zhou, Lingran Song, and Jianbing Shen. 2025. Mam: Modular multi- agent framework for multi-modal medical diagnosis via role-specialized collab- oration. InFindings of the Association for Computational Linguistics: ACL 2025. 25319–25333

    2025