arxiv: 2605.05640 · v1 · submitted 2026-05-07 · 💻 cs.CV

Recognition: unknown

AffectSeek: Agentic Affective Understanding in Long Videos under Vague User Queries

Zhen Zhang , Yuhang Yang , Yunxiang Jiang , Yuhuan Lu , Haifeng Lu , Zheng Lian , Runhao Zeng , Xiping Hu

Authors on Pith no claims yet

Pith reviewed 2026-05-08 14:54 UTC · model grok-4.3

classification 💻 cs.CV

keywords affective understandinglong video analysisvague queriesagentic frameworkemotion localizationvideo benchmarkmulti-step reasoningrationale generation

0 comments

The pith

AffectSeek uses five staged steps to locate and explain emotions in long videos from vague user queries.

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

The paper shifts affective understanding away from pre-clipped short videos or images toward real-world long videos where users pose vague natural-language questions. It defines the VQAU task, which requires localizing affective moments, classifying their emotions, and generating evidence-based rationales. To solve the multi-step challenges, the authors build VQAU-Bench for unified testing and introduce AffectSeek, an agentic system that decomposes the work into intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation. Specialized roles and cross-stage checks let the system progressively match vague intent to temporal evidence. A sympathetic reader would care because everyday video interaction involves long, unsegmented content and imprecise queries that current passive models cannot handle.

Core claim

The paper claims that vague-query-driven video affective understanding remains unsolved by existing affective recognition models and single-step vision-language models, and that AffectSeek solves it by decomposing the task into intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation, then progressively aligning vague user intent with long-video evidence through role-specialized reasoning and cross-stage verification.

What carries the argument

The five-stage decomposition of intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation, together with role-specialized reasoning and cross-stage verification.

If this is right

VQAU-Bench enables systematic measurement of semantic-temporal-affective alignment, moment localization, emotion classification, and rationale generation.
Existing affective recognition models and single-step vision-language models perform poorly on the VQAU task.
AffectSeek supplies a simple, effective framework that improves agentic handling of long-video affective queries.

Where Pith is reading between the lines

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

The staged verification approach could apply to other video tasks that start from vague instructions, such as action or event localization.
Breaking queries into verifiable stages may lower error accumulation in any multi-step video reasoning pipeline.
Interactive video systems could adopt similar decomposition to support real-time user clarifications during affective searches.

Load-bearing premise

The five-stage process with role-specialized reasoning and cross-stage verification will reliably align vague user intent to long-video affective evidence without introducing compounding errors.

What would settle it

If AffectSeek applied to VQAU-Bench produces no measurable gains in affective moment localization accuracy, emotion classification precision, or rationale quality over single-step vision-language models.

Figures

Figures reproduced from arXiv: 2605.05640 by Haifeng Lu, Runhao Zeng, Xiping Hu, Yuhang Yang, Yuhuan Lu, Yunxiang Jiang, Zheng Lian, Zhen Zhang.

**Figure 1.** Figure 1: VQAU task-solving pipeline. Given a query and a full video as inputs, view at source ↗

**Figure 2.** Figure 2: Illustration of partial dataset scenarios. Specifically, (a)–(f) show the view at source ↗

**Figure 4.** Figure 4: Overview of the proposed annotation pipeline. The framework identifies affective event clips with temporal boundaries and emotion labels, generates view at source ↗

**Figure 5.** Figure 5: Distribution of samples across emotion categories in the dataset. The view at source ↗

**Figure 6.** Figure 6: Dataset-level statistics of the constructed benchmark. The four donut charts report the distributions of query length, reason explanation length, affective view at source ↗

**Figure 7.** Figure 7: Overall workflow of AffectSeek framework. Given a vague query and a full video, the CoreAgent dynamically invokes the LocalizeAgent for temporal view at source ↗

**Figure 8.** Figure 8: Visualization of representative prediction results. The comparison shows that AffectSeek effectively localizes affective event clips, identifies the view at source ↗

**Figure 9.** Figure 9: Ablation Study of Individual Tools in AffectSeek. The comparison view at source ↗

**Figure 10.** Figure 10: Ablation study on Each Agent in AffectSeek. The experimental view at source ↗

read the original abstract

Existing affective understanding studies have mainly focused on recognizing emotions from images, audio signals, or pre-cliped video clips, where the affective evidence is already given. This passive and clip-centered setting does not fully reflect real-world scenarios, in which users often interact with long videos and express their needs through natural-language queries. In this paper, we study \textbf{Vague-Query-driven video Affective Understanding (VQAU)}, a new task that requires models to localize affective moments in long videos, predict their emotion categories, and generate evidence-grounded rationales under vague user queries. To support this task, we construct \textbf{VQAU-Bench}, a benchmark that integrates long videos, vague affective queries, temporal clip annotations, emotion labels, and rationale explanations into a unified evaluation framework. VQAU-Bench enables systematic assessment of semantic-temporal-affective alignment, affective moment localization, emotion classification, and rationale generation. To address the multi-step reasoning challenges of VQAU, we further propose \textbf{AffectSeek}, an agentic framework that actively seeks, verifies, and explains affective moments in long videos. AffectSeek decomposes VQAU into intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation, and progressively aligns vague user intent with long-video evidence through role-specialized reasoning and cross-stage verification. Experiments show that VQAU remains challenging for existing affective recognition models and single-step vision-language models, while AffectSeek provides a simple yet effective framework for agentic long-video affective understanding.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper defines VQAU as a new task for localizing affective moments in long videos from vague queries, releases VQAU-Bench, and presents AffectSeek as a five-stage agentic pipeline that claims better results than standard models.

read the letter

The main point is that this work moves affective video analysis from short pre-clipped segments to an active, query-driven search over longer footage. They formalize VQAU, build a benchmark that bundles long videos, vague natural-language queries, temporal annotations, emotion labels, and rationales, and then give AffectSeek as a decomposition into intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation, with role-specific agents and cross-stage checks. That setup directly targets the gap between passive recognition and real user interactions, and the benchmark looks like a practical addition for testing semantic-temporal-affective alignment in one place. The abstract reports that existing affective models and single-step vision-language models struggle while AffectSeek improves performance, which aligns with the motivation if the numbers are backed by proper controls and ablations in the full text. The staged approach is a straightforward way to manage multi-step reasoning without inventing new architectures from scratch. The soft spot is the risk of error accumulation across stages, especially on vague queries where early localization mistakes could cascade; the cross-verification is meant to catch this, but without detailed failure-case breakdowns or stage-wise metrics it is hard to judge how much it actually helps. The evaluation of rationale quality also seems subjective by nature and could benefit from clearer inter-annotator agreement or human preference data. This is aimed at researchers in affective computing and long-video understanding who want benchmarks or agentic baselines for query-driven tasks. A reader working on video search tools or multi-agent systems would find the task definition and data useful even if they adapt the method. It deserves peer review because the new task and unified benchmark are concrete contributions that the community can build on, regardless of how the specific pipeline performs in revisions.

Referee Report

0 major / 3 minor

Summary. The paper introduces Vague-Query-driven video Affective Understanding (VQAU) as a new task requiring localization of affective moments in long videos, emotion category prediction, and evidence-grounded rationale generation from vague natural-language queries. It constructs VQAU-Bench, a unified benchmark combining long videos, vague affective queries, temporal clip annotations, emotion labels, and rationale explanations to evaluate semantic-temporal-affective alignment. The proposed AffectSeek framework decomposes the task into five stages—intent interpretation, candidate localization, clip verification, emotion reasoning, and rationale generation—employing role-specialized agents and cross-stage verification to progressively align vague user intent with long-video evidence. Experiments indicate that existing affective recognition models and single-step vision-language models struggle on VQAU, while AffectSeek offers an effective agentic solution.

Significance. If the experimental results and ablations hold, this work meaningfully extends affective understanding beyond passive, pre-clipped settings to realistic interactive long-video scenarios with vague queries. The VQAU-Bench benchmark enables systematic, multi-dimensional evaluation that was previously unavailable, and the five-stage agentic decomposition with verification provides a concrete, reproducible template for handling multi-step reasoning in video-language tasks. These contributions could influence agentic frameworks in computer vision and multimodal AI.

minor comments (3)

[Abstract] Abstract: the statement that 'AffectSeek provides a simple yet effective framework' would benefit from a brief quantitative highlight (e.g., improvement margins on key metrics) to substantiate the effectiveness claim without requiring the reader to reach the experiments section.
The operational definition of 'vague' queries and the criteria used to generate or select them in VQAU-Bench should be stated more explicitly (e.g., ambiguity level, lexical features, or annotation protocol) to support reproducibility and future extensions.
Ensure all stage-specific prompts or role instructions for the specialized agents are either included in the main text or clearly referenced to an appendix so that the cross-stage verification mechanism can be exactly replicated.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and constructive review. We are pleased that the referee recognizes the novelty of the VQAU task, the utility of VQAU-Bench, and the practical value of the five-stage agentic decomposition in AffectSeek for handling vague queries in long videos.

Circularity Check

0 steps flagged

No significant circularity; empirical agentic pipeline with no derivations

full rationale

The paper defines a new task (VQAU) and proposes AffectSeek as a five-stage agentic framework (intent interpretation, candidate localization, clip verification, emotion reasoning, rationale generation) with role-specialized agents and cross-stage verification. No equations, fitted parameters, predictions, or first-principles derivations are present that could reduce to their own inputs by construction. The benchmark VQAU-Bench is constructed to support evaluation, and claims rest on experimental results rather than self-referential definitions or self-citation chains. This is a standard empirical proposal of a pipeline, self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only view reveals no explicit free parameters, mathematical axioms, or newly postulated entities; the work relies on standard assumptions in computer vision, NLP, and agentic systems.

pith-pipeline@v0.9.0 · 5596 in / 1182 out tokens · 38244 ms · 2026-05-08T14:54:41.977677+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 6 canonical work pages · 2 internal anchors

[1]

Multimodal emotion recognition in conversations: A survey of methods, trends, challenges and prospects,

C. Wu, Y . Cai, Y . Liu, P. Zhu, Y . Xue, Z. Gong, J. Hirschberg, and B. Ma, “Multimodal emotion recognition in conversations: A survey of methods, trends, challenges and prospects,”arXiv preprint arXiv:2505.20511, 2025

work page arXiv 2025
[2]

Facial expression analysis and its potentials in iot systems: A contemporary survey,

Z. Shangguan, Y . Dong, S. Guo, V . C. Leung, M. J. Deen, and X. Hu, “Facial expression analysis and its potentials in iot systems: A contemporary survey,”ACM Computing Surveys, vol. 58, no. 2, pp. 1–39, 2025

2025
[3]

Emotion recognition from skeleton data: A comprehensive survey,

H. Lu, J. Chen, Z. Zhang, R. Liu, R. Zeng, and X. Hu, “Emotion recognition from skeleton data: A comprehensive survey,”arXiv preprint arXiv:2507.18026, 2025

work page arXiv 2025
[4]

Explain- able ai for audio and visual affective computing: A scoping review,

D. S. Johnson, O. Hakobyan, J. Paletschek, and H. Drimalla, “Explain- able ai for audio and visual affective computing: A scoping review,” IEEE Transactions on Affective Computing, vol. 16, no. 2, pp. 518–536, 2025

2025
[5]

Poster++: A simpler and stronger facial expression recognition net- work,

J. Mao, R. Xu, X. Yin, Y . Chang, B. Nie, A. Huang, and Y . Wang, “Poster++: A simpler and stronger facial expression recognition net- work,”Pattern Recognition, vol. 157, p. 110951, 2025

2025
[6]

Mser: Multimodal speech emotion recognition using cross-attention with deep fusion,

M. Khan, W. Gueaieb, A. El Saddik, and S. Kwon, “Mser: Multimodal speech emotion recognition using cross-attention with deep fusion,” Expert Systems with Applications, vol. 245, p. 122946, 2024

2024
[7]

Towards stable cross-domain depression recognition under missing modalities,

J. Chen, M. Tan, H. Lu, Q. Xu, Z. Wang, R. Zeng, and X. Hu, “Towards stable cross-domain depression recognition under missing modalities,”Pattern Recognition, p. 113367, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0031320326003328 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

2026
[8]

Rmer-dt: robust multimodal emotion recognition in conversational contexts based on diffusion and transformers,

X. Zhu, Y . Wang, E. Cambria, I. Rida, J. S. L ´opez, L. Cui, and R. Wang, “Rmer-dt: robust multimodal emotion recognition in conversational contexts based on diffusion and transformers,”Information Fusion, vol. 123, p. 103268, 2025

2025
[9]

Understanding emotional body expressions via large language models,

H. Lu, J. Chen, F. Liang, M. Tan, R. Zeng, and X. Hu, “Understanding emotional body expressions via large language models,” vol. 39, no. 2, Apr. 2025, pp. 1447–1455. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/32135

2025
[10]

Explainable affective body expression recognition with multi-scale spatiotemporal encoding and llm-based reasoning,

T. Wang, H. Lu, J. Duan, T. Meng, R. Mao, S. Liu, and D. Ming, “Explainable affective body expression recognition with multi-scale spatiotemporal encoding and llm-based reasoning,”IEEE Transactions on Affective Computing, pp. 1–13, 2026

2026
[11]

Skeleton-based pre-training with discrete labels for emotion recognition in iot environments,

Z. Zhang, F. Liang, W. Wang, R. Zeng, V . C. Leung, and X. Hu, “Skeleton-based pre-training with discrete labels for emotion recognition in iot environments,”IEEE Internet of Things Journal, 2025

2025
[12]

From long videos to engaging clips: A human-inspired video editing framework with multimodal narrative understanding,

X. Wang, X. Li, Y . Wei, Y . Song, F. Zeng, Z. Chen, G. Xu, T. Xuet al., “From long videos to engaging clips: A human-inspired video editing framework with multimodal narrative understanding,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2025, pp. 2764–2781

2025
[13]

Joint moment retrieval and highlight detection via natural language queries,

R. Luo, A. Peng, H. Yap, and K. Beard, “Joint moment retrieval and highlight detection via natural language queries,”arXiv preprint arXiv:2305.04961, 2023

work page arXiv 2023
[14]

Reliable crowdsourcing and deep locality- preserving learning for expression recognition in the wild,

S. Li, W. Deng, and J. Du, “Reliable crowdsourcing and deep locality- preserving learning for expression recognition in the wild,” inProceed- ings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2852–2861

2017
[15]

Affectnet: A database for facial expression, valence, and arousal computing in the wild,

A. Mollahosseini, B. Hasani, and M. H. Mahoor, “Affectnet: A database for facial expression, valence, and arousal computing in the wild,”IEEE transactions on affective computing, vol. 10, no. 1, pp. 18–31, 2017

2017
[16]

A database of german emotional speech,

F. Burkhardt, “A database of german emotional speech,” 2000

2000
[17]

Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings,

R. Lotfian and C. Busso, “Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings,”IEEE Transactions on Affective Computing, vol. 10, no. 4, pp. 471–483, 2017

2017
[18]

Affectgpt: A new dataset, model, and benchmark for emotion understanding with multimodal large language models,

Z. Lian, H. Chen, L. Chen, H. Sun, L. Sun, Y . Ren, Z. Cheng, B. Liu, R. Liu, X. Penget al., “Affectgpt: A new dataset, model, and benchmark for emotion understanding with multimodal large language models,” in Proceedings of the 42nd International Conference on Machine Learning, 2025

2025
[19]

Temporal sentiment localization: Listen and look in untrimmed videos,

Z. Zhang and J. Yang, “Temporal sentiment localization: Listen and look in untrimmed videos,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 199–208

2022
[20]

Moviegraphs: Towards understanding human-centric situations from videos,

P. Vicol, M. Tapaswi, L. Castrejon, and S. Fidler, “Moviegraphs: Towards understanding human-centric situations from videos,” inPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018

2018
[21]

Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,

Z. Cheng, Z.-Q. Cheng, J.-Y . He, J. Sun, K. Wang, Y . Lin, Z. Lian, X. Peng, and A. G. Hauptmann, “Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning,”Advances in Neural Information Processing Systems, vol. 37, pp. 110 805–110 853, 2024

2024
[22]

Emotion-llamav2 and mmeverse: A new framework and benchmark for multimodal emotion understanding,

X. Peng, J. Chen, Z. Cheng, B. Peng, F. Wu, Y . Dong, S. Tu, Q. Hu, H. Huang, Y . Linet al., “Emotion-llamav2 and mmeverse: A new framework and benchmark for multimodal emotion understanding,” arXiv preprint arXiv:2601.16449, 2026

work page arXiv 2026
[23]

Agent-mer: A cognitive agent with hierarchical deliberation for open-vocabulary multimodal emotion recognition,

Z. Lai, Z. Zhu, X. Hong, and Y . Wang, “Agent-mer: A cognitive agent with hierarchical deliberation for open-vocabulary multimodal emotion recognition,” inProceedings of the 33rd ACM International Conference on Multimedia, MM 2025, Dublin, Ireland, October 27-31, 2025, C. Gurrin, K. Schoeffmann, M. Zhang, L. Rossetto, S. Rudinac, D. Dang-Nguyen, W. Cheng,...

2025
[24]

Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene- enhanced video large language model,

J. Zhao, J. Wang, Y . Jin, J. Luo, and G. Zhou, “Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene- enhanced video large language model,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 592–601

2024
[25]

Omni-sila: Towards omni-scene driven visual sentiment identifying, locating and attributing in videos,

J. Luo, J. Wang, J. Ma, Y . Jin, S. Li, and G. Zhou, “Omni-sila: Towards omni-scene driven visual sentiment identifying, locating and attributing in videos,” inProceedings of the ACM on Web Conference 2025, 2025, pp. 188–197

2025
[26]

Emodetective: Detecting, exploring, and thinking emotional cause in videos,

X. Huang, Y . Zhou, J. Li, S. Lu, and S. Wang, “Emodetective: Detecting, exploring, and thinking emotional cause in videos,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 5735– 5744

2025
[27]

Vad: A video affective dataset with danmu,

S. Wang, X. Li, F. Zheng, J. Pan, X. Li, Y . Chang, Q. Li, J. Wang, Y . Xiao et al., “Vad: A video affective dataset with danmu,”IEEE Transactions on Affective Computing, 2024

2024
[28]

Dense- captioning events in videos,

R. Krishna, K. Hata, F. Ren, L. Fei-Fei, and J. Carlos Niebles, “Dense- captioning events in videos,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 706–715

2017
[29]

Tall: Temporal activity local- ization via language query,

J. Gao, C. Sun, Z. Yang, and R. Nevatia, “Tall: Temporal activity local- ization via language query,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 5267–5275

2017
[30]

Dfew: A large-scale database for recognizing dynamic facial expressions in the wild,

X. Jiang, Y . Zong, W. Zheng, C. Tang, W. Xia, C. Lu, and J. Liu, “Dfew: A large-scale database for recognizing dynamic facial expressions in the wild,” inProceedings of the 28th ACM international conference on multimedia, 2020, pp. 2881–2889

2020
[31]

Ferv39k: A large-scale multi-scene dataset for facial expres- sion recognition in videos,

Y . Wang, Y . Sun, Y . Huang, Z. Liu, S. Gao, W. Zhang, W. Ge, and W. Zhang, “Ferv39k: A large-scale multi-scene dataset for facial expres- sion recognition in videos,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 20 922–20 931

2022
[32]

Mer 2023: Multi-label learning, modality robustness, and semi-supervised learning,

Z. Lian, H. Sun, L. Sun, K. Chen, M. Xu, K. Wang, K. Xu, Y . He, Y . Li, J. Zhaoet al., “Mer 2023: Multi-label learning, modality robustness, and semi-supervised learning,” inProceedings of the 31st ACM international conference on multimedia, 2023, pp. 9610–9614

2023
[33]

Meld: A multimodal multi-party dataset for emotion recognition in conversations,

S. Poria, D. Hazarika, N. Majumder, G. Naik, E. Cambria, and R. Mihal- cea, “Meld: A multimodal multi-party dataset for emotion recognition in conversations,” inProceedings of the 57th annual meeting of the association for computational linguistics, 2019, pp. 527–536

2019
[34]

Emovit: Revolutionizing emotion insights with visual instruction tuning,

H. Xie, C.-J. Peng, Y .-W. Tseng, H.-J. Chen, C.-F. Hsu, H.-H. Shuai, and W.-H. Cheng, “Emovit: Revolutionizing emotion insights with visual instruction tuning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 596–26 605

2024
[35]

Mafw: A large-scale, multi-modal, compound affective database for dynamic facial expression recognition in the wild,

Y . Liu, W. Dai, C. Feng, W. Wang, G. Yin, J. Zeng, and S. Shan, “Mafw: A large-scale, multi-modal, compound affective database for dynamic facial expression recognition in the wild,” inProceedings of the 30th ACM international conference on multimedia, 2022, pp. 24–32

2022
[36]

Open-vocabulary multimodal emotion recog- nition: Dataset, metric, and benchmark,

Z. Lian, H. Sun, L. Sun, L. Chen, H. Chen, H. Gu, Z. Wen, S. Chen, Z. Siyuan, H. Yaoet al., “Open-vocabulary multimodal emotion recog- nition: Dataset, metric, and benchmark,” 2024

2024
[37]

Building machines that learn and think like people,

B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman, “Building machines that learn and think like people,”Behavioral and brain sciences, vol. 40, p. e253, 2017

2017
[38]

Qwen3.5: Accelerating productivity with native multimodal agents,

Q. Team, “Qwen3.5: Accelerating productivity with native multimodal agents,” February 2026. [Online]. Available: https://qwen.ai/blog?id= qwen3.5

2026
[39]

Gemini: A Family of Highly Capable Multimodal Models

G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millicanet al., “Gemini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

work page internal anchor Pith review arXiv 2023
[40]

Introducing gpt-5.4,

OpenAI, “Introducing gpt-5.4,” https://openai.com/index/ introducing-gpt-5-4/, 2026, accessed: 2026-04-29

2026
[41]

Kimi K2.5: Visual Agentic Intelligence

K. Team, T. Bai, Y . Bai, Y . Bao, S. Cai, Y . Cao, Y . Charles, H. Che, C. Chen, G. Chenet al., “Kimi k2. 5: Visual agentic intelligence,”arXiv preprint arXiv:2602.02276, 2026

work page internal anchor Pith review arXiv 2026