AgenticVAU: Multi-Agent Explore-Verify Reasoning for Video Anomaly Understanding

Ao Li; Huining Li; Jian Zhang; Lanju Kong; Ning Liu; Shuai Feng; Xingdong Sheng; Yuntao Du; Yuxiang Duan

REVIEW 3 major objections 5 minor 46 references

AgenticVAU: Multi-Agent Explore-Verify Reasoning for Video Anomaly Understanding

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A training-free multi-agent explore–verify loop outperforms zero-shot and reinforcement-tuned video anomaly understanding baselines on VAU-Bench, with the largest gains in anomaly classification.

desk verdict A training-free multi-agent explore–verify system for video anomaly understanding reports large, consistent gains, but the headline attribution to multi-agent collaboration is confounded by a stronger reasoning model and extra observation budget. read the letter →

arxiv 2608.03779 v1 pith:ATAI77MV submitted 2026-08-04 cs.CV

Yuxiang Duan , Huining Li , Ao Li , Shuai Feng , Lanju Kong , Ning Liu , Jian Zhang , Xingdong Sheng

show 1 more author

Yuntao Du

This is my paper

classification cs.CV

keywords videoanomalyunderstandingmulti-agentreasoningexplore-verifyanchorregistrytraining-freeinferencevideo-languagemodelsVAU-Benchclassification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that video anomaly understanding—identifying what happened, where it happened, and why it is abnormal—is better cast as an explore-then-verify process than as a single pass or a single agent. It proposes a training-free system of four specialized agents: one builds scene-aware visual rules, one plans what to look at next, one inspects video frames, and one makes the final decision. The key is the anchor registry, a shared evidence memory that tags each observation as supporting, opposing, or unclear for a candidate anomaly. On VAU-Bench, the system beats zero-shot inference and a reinforcement-fine-tuned baseline, especially in binary and multi-class anomaly classification.

What carries the argument

The anchor registry is the load-bearing mechanism: a shared evidence memory that binds each observation to a temporal scope, a visual rule, and an explicit verification state (verified, rejected, or unclear). It is written by a memory update after each observation and read by the reasoning agent for planning and by the summary agent for final aggregation. Around it, four agents specialize: RuleAgent constructs supporting and opposing visual criteria, ReasonAgent plans each observation tool (Global Scan, Segment Focus, Temporal Stitch), ObserveAgent executes frame-level inspection, and SummaryAgent converts the evidence state into the task output.

What would settle it

Re-run AgenticVAU with the same four agents but replace the anchor registry with a flat, unstructured transcript of observations while keeping all other components identical; if binary accuracy on UCF-Crime remains near 94.42 percent, then the structured evidence states are not the driver of the reported gains and the central claim is falsified.

Watch

Extended reading notes

Core claim

AgenticVAU formalizes video anomaly understanding as an iterative explore–verify loop. The system first performs a prior scan and builds scene-aware visual rules, then cycles through search planning, targeted video observation at multiple temporal granularities, and structured memory updates until enough evidence is collected. The anchor registry is the central organizing device: each candidate anomaly is an anchor, and every observation is bound to a temporal interval, a visual rule, a verification state (verified, rejected, or unclear), and an evidence summary. This lets the system distinguish 'evidence against' from 'not yet determined,' and it lets the final summary agent aggregate only

Load-bearing premise

The framework's gains rest on the observation model correctly seeing the supporting and opposing visual details in coarsely sampled frames, and on the reasoning model correctly labeling evidence as verified, rejected, or unclear; if either judgment is unreliable, the loop's advantage collapses.

Editorial extensions

If this is right

If correct, training-free agentic reasoning can substitute for task-specific fine-tuning in video anomaly understanding, lowering adaptation cost to new scenes and anomaly types.
The large classification gains suggest that structured evidence collection addresses a weakness that simply scaling or tuning the underlying video-language model does not fix.
The modular design implies that improving the observation model—higher frame rates, better temporal sampling, or stronger vision encoders—should compound with the agent loop rather than require re-training the orchestration.
The verified/rejected/unclear states provide a natural confidence signal that could support abstention, explanation, or human review in deployed surveillance settings.
The explore–verify structure may transfer to other temporally dispersed video understanding tasks, such as action anticipation or causal reasoning, where evidence must be gathered from separated moments.

Reading between the lines

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

The paper does not ablate the anchor registry itself while keeping the multi-agent loop intact; a comparison against a flat, unstructured memory would isolate whether structured evidence states are the driver of the gains, not just the agent decomposition.
With 48 percent of sampled errors attributed to the ObserveAgent, the fastest path to further improvement is likely better visual evidence extraction—for example adaptive frame selection or higher sampling density—rather than more sophisticated planning or memory.
The explore–verify loop is essentially a hypothesis-testing procedure; the same formulation could be applied to anomaly understanding in audio, sensor streams, or medical video where supporting and opposing evidence is sparse and temporally dispersed.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things you should know. First, this is a real attempt to restructure video anomaly understanding as an explore–verify process with a shared anchor registry that tracks supporting and opposing evidence per candidate event. That formulation is genuinely new in the agentic-VAU literature, and the case study shows the mechanism doing concrete work: an early wrong prior gets corrected through a global scan, a focused re-observation, and a verified anchor. Second, the reported numbers are too clean to take at face value, and the stress-test note is right: nothing in the paper isolates the multi-agent loop from the reasoning model.

What the paper does well. The architecture is coherent. The ablation of observation tools is sensible and shows Segment Focus carries the most weight. Table 5 at least demonstrates robustness to model choices. The manual error analysis honestly attributes 48% of failures to the observation model, which is the kind of self-criticism you want to see. And the system leads every reported metric on every dataset; the classification margins are large enough that this deserves referee time rather than a desk reject.

Where it is soft. The central attribution problem is real. Baselines use Qwen2.5-VL-3B alone, while AgenticVAU gives all text-level decisions to DeepSeek-V4-Pro. Table 5 itself shows that swapping Flash for Pro moves accuracy by about 0.8 points, so the reasoning model is not inert. Without a single-agent baseline that runs DeepSeek-V4-Pro with a fixed observation budget, you cannot say the gains come from role separation or the anchor registry. That is a load-bearing gap for the main claim, though not for the empirical fact that the pipeline scores well. Second, Table 3 compares temporal grounding only against Qwen zero-shot; VAU-R1 is missing, so the 'outperforms baselines' claim is unsupported for that task. Third, there are no seeds or error bars, and the 94.42/94.42 overlap in Table 1 on UCF-Crime looks like a copy-paste or rounding artifact worth checking. Fourth, for a training-free system, prompts and inference configuration are the method, but neither is released.

Verdict. Send it out. The mechanism is plausible and the gaps are answerable with moderate additional work. As is, I would not cite the headline claim as evidence for multi-agent collaboration, but I would cite the system as a representative agentic approach with a caveat. If I were the editor, I would ask reviewers to demand the missing control and the missing grounding baseline.

Referee Report

3 major / 5 minor

Summary. The paper introduces AgenticVAU, a training-free multi-agent framework for video anomaly understanding (VAU). The method separates the task into four specialized agents — RuleAgent, ReasonAgent, ObserveAgent, and SummaryAgent — coordinated by an anchor registry that records evidence with temporal support, visual-rule association, and a candidate/verified/rejected/unclear verification state. The system first constructs scene-aware visual rules and a query-conditioned prior, then runs an explore–verify loop that alternates global scan, segment focus, and temporal-stitch observations until enough evidence is collected, and finally aggregates the anchor registry into a task-specific output. Experiments on ECVA, UCF-Crime, and MSAD subsets of VAU-Bench report consistent improvements over Qwen2.5-VL-3B zero-shot inference and VAU-R1 reinforcement-fine-tuned baseline on multiple-choice QA and anomaly classification, and over the zero-shot baseline on temporal grounding, with the largest gains in normal/abnormal and multi-class classification. The authors attribute these gains to multi-agent collaboration enabled by role separation and structured evidence memory.

Significance. If the claims hold, AgenticVAU provides a strong demonstration that structured, role-separated evidence collection can substantially improve anomaly classification and reasoning without task-specific training. The paper has several clear strengths: the method is described in detail, the anchor registry is a sensible mechanism for organizing supporting versus opposing evidence, the case study (Figure 3) illustrates a concrete mechanism by which the explore–verify loop corrects a wrong prior, and the observation-tool ablation (Table 4) and model-variant ablation (Table 5) show sensitivity to design choices. However, the central attribution to 'multi-agent collaboration' is not yet established, because the framework replaces the base VLM's inference with a pipeline in which all text-level reasoning is performed by DeepSeek-V4-Pro, a frontier LLM not available to either baseline. The empirical margins are promising, but the current experiments do not isolate the effect of role separation, the anchor registry, or the explore–verify loop from the effect of the stronger reasoning model and the increased observation/compute budget. These are correctness-of-interpretation risks rather than

major comments (3)

[Experiment — Compared Methods; Table 5] The paper's headline claim that multi-agent collaboration is the source of improvement is underdetermined. AgenticVAU replaces Qwen2.5-VL-3B monolithic inference with a pipeline in which rule generation, planning, memory update, verification-state assignment, and final prediction are all produced by DeepSeek-V4-Pro. Table 5 shows the reasoning model is not inert: swapping DeepSeek-V4-Flash for DeepSeek-V4-Pro changes UCF-Crime binary accuracy from 93.63% to 94.42% even with the same video model. No experiment measures what DeepSeek-V4-Pro alone achieves on the same questions and the same or comparable frame budget. I request controls that isolate the multi-agent scaffold: (i) DeepSeek-V4-Pro zero-shot with a fixed sampled-frame set; (ii) a single-agent ReAct-style loop using the same reasoning model, the same three observation tools, and the same maximum number of steps but without role
[Temporal Anomaly Grounding; Table 3] Table 3 compares AgenticVAU only against Qwen2.5-VL-3B for temporal grounding; VAU-R1 is absent from this task. Since the abstract and conclusion claim that AgenticVAU outperforms both zero-shot inference and reinforcement-learning-based baselines across multiple VAU tasks, the grounding comparison is incomplete for the cross-task claim. Moreover, no variance information is reported anywhere in the paper: all metrics appear to be from single runs without seeds or error bars. While the classification gains are large enough to be credible, the MCQ improvements (e.g., 91.10 vs. 89.53 on ECVA) are small enough that single-run reporting makes them impossible to assess. Please add VAU-R1 grounding numbers and report means over multiple seeds with error bars, or at least state explicitly that only one run was performed and indicate which conclusions should be treated as preliminary.
[Insightful Analysis — Error Analysis; Figure 4] The core of the method is the anchor registry, yet the verification states (candidate/verified/rejected/unclear) are assigned by the same reasoning model that also performs planning and summarization. There is no external evaluation of whether these self-assigned states agree with the video content. The error analysis in Figure 4 attributes 48% of failures to ObserveAgent and 18% to MemoryUpdate, but this is a manual categorization of 50 sampled errors and does not measure the reliability of the registry states themselves. If perception or self-adjudication is unreliable, the entire explore–verify mechanism can collapse regardless of agent architecture. I request a state-accuracy measurement: sample a set of anchors, have human annotators judge whether the assigned verification state and evidence summary are consistent with the ground-truth interval and anomaly label, and report precisio

minor comments (5)

[Explore–Verify Loop — Loop Control] The termination criterion 'several consecutive observations produce no useful registry update' is vague. Define 'several' and 'useful' operationally, or replace with a concrete threshold such as the number of unchanged anchors.
[Implementation Details; Figure 2] Several sampling hyperparameters appear in the case study but are not formalized in the experimental setup, including Global Scan slice count and fps, Segment Focus fps, and the registry confidence thresholds (0.55, 0.90 in Figure 3). Please provide a table of these presets and the registry confidence threshold used during updates.
[Implementation Details] The maximum number of Explore–Verify steps is fixed at 15, but no analysis is given of how performance varies with T, nor is the average number of steps used per dataset/task reported. A short sensitivity analysis or average-step table would help readers understand the inference cost.
[Figure 4 caption and text] The error analysis is based on 50 manually categorized errors. Please specify how the 50 cases were sampled, the criteria used for assigning the primary failure source, and whether any inter-annotator agreement was measured.
[Table 2] There are formatting errors in Table 2: values such as '8.715.98' and '6.846.03' appear to be missing separators between FLU and INF columns. Please correct these.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark evaluation is externally grounded and the claimed gains are empirical, not derived from fitted inputs or self-citations.

full rationale

AgenticVAU's central claim is an empirical performance comparison on VAU-Bench (ECVA, UCF-Crime, MSAD) against zero-shot Qwen2.5-VL-3B and reinforcement-fine-tuned VAU-R1. The reported accuracies, mIoU, and VAU-Eval scores come from external benchmark labels and human-defined metric rubrics, not from the method's own equations or fitted parameters. No parameter is estimated on the test set and then reported as a prediction; Tables 4 and 5 are ablations of tools and model variants, not fitted inputs. The framework uses DeepSeek-V4-Pro for multiple reasoning roles, which creates an attribution confound: no single-agent control with the same reasoning model and frame budget isolates the explore-verify loop, so the claim that multi-agent collaboration specifically drives the gains is underdetermined. That is a correctness-of-interpretation risk, not circularity: the paper never defines its final answer as the same quantity as its rules, anchors, or verification states, and the final prediction is constrained by external video frames and ground-truth labels during evaluation. Citations to prior benchmarks and baselines (e.g., Zhu et al. 2025; Du et al. 2024a) are not load-bearing self-citations from this paper's author set, and no uniqueness theorem or ansatz is smuggled in via citation. The paper's own error analysis (Figure 4) and one-tool-at-a-time ablation treat the architecture components as empirically separable rather than asserting their contribution by construction. Therefore no circular step meeting the quoted-reduction standard is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

No fitted constants or closed-form derivation appear in this paper; the central claim is an empirical performance claim validated on external benchmark labels. The ledger instead records the operational parameters set by hand and the trust assumptions placed in the two underlying LLMs and in the benchmark metrics. The absence of an ablation isolating the anchor registry and the absence of single-agent agentic baselines mean two central mechanism claims rest on unmeasured assumptions.

free parameters (3)

Maximum Explore-Verify steps T = 15
Hand-set loop bound (Implementation Details); no sensitivity analysis on T is reported, yet T interacts with early-stop criteria and with how much evidence the SummaryAgent receives.
Observation sampling presets (Global Scan slice count and fps, Segment Focus fps) = e.g., slice_num=6, fps=1.0 for Global Scan; fps=2.0 for Segment Focus in the Figure 3 case
Sampling density is chosen by hand and stated to be identical across datasets for fairness, but no sweep shows robustness to density; coarser sampling would weaken evidence capture, denser sampling would raise cost.
Registry confidence thresholds shown in the case study = conf 0.55 (candidate), 0.90 (verified) in Figure 3
The status is unclear: these values may be illustrative annotations rather than enforced thresholds; the paper does not specify how confidence is computed or used in the real system.

assumptions (5)

domain assumption The video observation model (Qwen2.5-VL-3B) can reliably perceive supporting/opposing visual criteria from coarsely sampled frames (roughly 1-2 fps)
Section 3.3: ObserveAgent judgments on sampled frames determine every verification state; Figure 4 concedes 48% of failures come from ObserveAgent, so this perception reliability is load-bearing.
domain assumption The reasoning model (DeepSeek-V4-Pro) correctly assigns verification states (verified/rejected/unclear), merges anchors, and never corrupts evidence in the registry
Section 3.3 (Anchor Registry Update) and Eq. 3: MemoryUpdate is an LLM text operation of the same model family that decides the final answer, so the evidence record is self-adjudicated. Figure 4 attributes 18% of failures to MemoryUpdate.
domain assumption VAU-Eval (CLS, KM, FLU, INF, FAC) scores are valid proxies for anomaly reasoning quality, including LLM-judged factual consistency (FAC)
Table 2 reasoning improvements are measured solely by this automated rubric inherited from Zhu et al. 2025; no human agreement or variance is reported.
domain assumption Ground-truth labels and temporal intervals in the ECVA, UCF-Crime, and MSAD subsets of VAU-Bench are correct
All accuracy and mIoU numbers in Tables 1-3 assume benchmark labels are ground truth.
domain assumption The initial prior P is fallible and used only to initialize search; the loop can and should override it
Section 3.1 (Prior Scan): 'P is used only to initialize exploration. It is not treated as verified evidence.' The case study (Figure 3) depends on the system recovering from a wrong prior.

invented entities (2)

Anchor registry with anchors in states candidate/verified/rejected/unclear
purpose: Shared evidence memory binding each observation to a visual rule, temporal interval, and verification state, coordinating the four agents
The registry's unique contribution (verification-state bookkeeping) is never isolated: Table 4 ablates observation tools, not the registry semantics, and the error analysis assumes the registry is reliable. No falsifiable handle on the registry's contribution outside the paper is given.
Four specialized agent roles (RuleAgent, ReasonAgent, ObserveAgent, SummaryAgent)
purpose: Role specialization over a shared evidence state to replace the unified single-agent loop
The agents are prompts over the same two LLMs rather than independent models, and no single-agent agentic baseline (PANDA, QVAD, Anom-pi) is benchmarked, so the marginal value of the role split itself is unmeasured; the 'value of multi-agent collaboration' claim is asserted from the full-system comparison only.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AgenticVAU: Multi-Agent Explore-Verify Reasoning for Video Anomaly Understanding." pith.science (2026). https://pith.science/paper/ATAI77MV

@misc{pith2026260803779,
  author       = {Pith},
  title        = {Pith review of: AgenticVAU: Multi-Agent Explore-Verify Reasoning for Video Anomaly Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ATAI77MV}},
  note         = {Machine review of arXiv:2608.03779}
}

read the original abstract

Video anomaly understanding (VAU) focuses on comprehensively interpreting abnormal events in videos, requiring models to identify anomalous occurrences, discover their supporting evidence, and explain the underlying causes beyond simple anomaly detection. Existing VAU methods often rely on specialized training or limited observations, restricting generalization or evidence coverage. Although single-agent alternatives support adaptive video observation, they still integrate exploration, observation, and decision-making within a unified reasoning process, offering limited role specialization and structured evidence coordination. To address these limitations, we present AgenticVAU, a training-free multi-agent framework that casts VAU as an explore--verify process, where the system first discovers potential anomalies and then verifies them through targeted observations. To achieve this, four specialized agents are introduced to handle visual-rule construction, search planning, video observation, and final decision, respectively. These agents communicate through an anchor registry, a shared evidence memory that binds each observation. Guided by this agent framework, AgenticVAU interleaves broad temporal exploration, dense local verification, and cross-interval comparison until sufficient evidence is collected. We conduct extensive experiments on the ECVA, UCF-Crime, and MSAD subsets of VAU-Bench, the results show that AgenticVAU outperforms zero-shot inference and reinforcement learning-based baselines, demonstrating the value of multi-agent collaboration for video anomaly understanding.

Figures

Figures reproduced from arXiv: 2608.03779 by the authors.

**Figure 1.** Comparison of VAU paradigms. (a) Single-model methods make one-shot predictions from fixed observations. (b) Single-agent methods rely on one agent to perform exploration, observation, evidence organization, and decisionmaking. (c) AgenticVAU distributes these responsibilities among specialized agents within an explore–verify process. its temporal search, revisit selected intervals with denser observations, or co… view at source ↗

**Figure 2.** Overview of AgenticVAU. Given a video V and a query Q, the framework first obtains a video prior and constructs scene-aware visual rules, where E + j,k and E − j,k denote the supporting and opposing visual criteria of rule rj . During the Explore– Verify Loop, ReasonAgent plans each observation, ObserveAgent inspects the requested video content, and the resulting evidence is organized in the anchor registry. Summary… view at source ↗

**Figure 3.** Case study of AgenticVAU on a shoplifting video. The base VLM mistakes early normal behavior for the target event, whereas AgenticVAU uses global exploration and local verification to revise the initial prior. The resulting verified anchor supports the correct answer, temporal grounding, anomaly class, and evidence-based explanation. Reasoning Model Video Model Acc. (%) DeepSeek-V4-Flash Qwen2.5-VL-3B 93.63 Qwen2.5-… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Distribution of the primary failure sources among [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 20 canonical work pages

[1]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Learning temporal regularity in video sequences , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

work page
[2]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Real-world anomaly detection in surveillance videos , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=
[3]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Future frame prediction for anomaly detection--a new baseline , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=
[4]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Weakly-supervised video anomaly detection with robust temporal feature magnitude learning , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=
[5]

MGFN: magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection , author=. Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence , pages=

work page
[6]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Dual memory units with uncertainty regulation for weakly supervised video anomaly detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[7]

IEEE Transactions on Image Processing , volume=

Learning prompt-enhanced context features for weakly-supervised video anomaly detection , author=. IEEE Transactions on Image Processing , volume=. 2024 , publisher=

work page 2024
[8]

Proceedings of the AAAI conference on artificial intelligence , volume=

Vadclip: Adapting vision-language models for weakly supervised video anomaly detection , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

work page

Show all 46 references

[9]

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

Open-vocabulary video anomaly detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[10]

arXiv preprint arXiv:2503.13160 , year=

Language-guided open-world video anomaly detection under weak supervision , author=. arXiv preprint arXiv:2503.13160 , year=
[11]

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

Harnessing large language models for training-free video anomaly detection , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[12]

European Conference on Computer Vision , pages=

Follow the rules: reasoning for video anomaly detection with large language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

2024
[13]

Proceedings of the computer vision and pattern recognition conference , pages=

Holmes-vau: Towards long-term video anomaly understanding at any granularity , author=. Proceedings of the computer vision and pattern recognition conference , pages=
[14]

arXiv preprint arXiv:2505.23504 , year=

Vau-r1: Advancing video anomaly understanding via reinforcement fine-tuning , author=. arXiv preprint arXiv:2505.23504 , year=

arXiv
[15]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

FineVAU: A Novel Human-Aligned Benchmark for Fine-Grained Video Anomaly Understanding , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[16]

Advances in Neural Information Processing Systems , volume=

Panda: Towards generalist video anomaly detection via agentic ai engineer , author=. Advances in Neural Information Processing Systems , volume=
[17]

arXiv preprint arXiv:2604.03040 , year=

Qvad: A question-centric agentic framework for efficient and training-free video anomaly detection , author=. arXiv preprint arXiv:2604.03040 , year=

arXiv
[18]

European Conference on Computer Vision , pages=

Videoagent: Long-form video understanding with large language model as agent , author=. European Conference on Computer Vision , pages=. 2024 , organization=

2024
[19]

European Conference on Computer Vision , pages=

Videoagent: A memory-augmented multimodal agent for video understanding , author=. European Conference on Computer Vision , pages=. 2024 , organization=

2024
[20]

arXiv preprint arXiv:2504.04471 , year=

Videoagent2: Enhancing the llm-based agent system for long-form video understanding by uncertainty-aware cot , author=. arXiv preprint arXiv:2504.04471 , year=

arXiv
[21]

Advances in Neural Information Processing Systems , volume=

Deep video discovery: Agentic search with tool use for long-form video understanding , author=. Advances in Neural Information Processing Systems , volume=
[22]

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

Videoarm: Agentic reasoning over hierarchical memory for long-form video understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[23]

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

LensWalk: Agentic video understanding by planning how you see in videos , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[24]

arXiv preprint arXiv:2504.20091 , year=

Videomultiagents: A multi-agent framework for video question answering , author=. arXiv preprint arXiv:2504.20091 , year=

arXiv
[25]

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

Lvagent: Long video understanding by multi-round dynamical collaboration of mllm agents , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=
[26]

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Longvideoagent: Multi-agent reasoning with long videos , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=
[27]

arXiv preprint arXiv:2506.18071 , year=

Mupa: Towards multi-path agentic reasoning for grounded video question answering , author=. arXiv preprint arXiv:2506.18071 , year=

arXiv
[28]

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

SVAgent: Storyline-Guided Long Video Understanding via Cross-Modal Multi-Agent Collaboration , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[29]

arXiv preprint arXiv:2605.00444 , year=

Scaling Video Understanding via Compact Latent Multi-Agent Collaboration , author=. arXiv preprint arXiv:2605.00444 , year=

arXiv
[30]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=
[31]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Text prompt with normality guidance for weakly supervised video anomaly detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=
[32]

Advances in Neural Information Processing Systems , volume=

Hawk: Learning to understand open-world video anomalies , author=. Advances in Neural Information Processing Systems , volume=
[33]

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

Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[34]

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

Moviechat: From dense token to sparse memory for long video understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[35]

Advances in Neural Information Processing Systems , volume=

Longvideobench: A benchmark for long-context interleaved video-language understanding , author=. Advances in Neural Information Processing Systems , volume=
[36]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

Videotree: Adaptive tree-based video representation for llm reasoning on long videos , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
[37]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

Adaptive keyframe sampling for long video understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
[38]

arXiv preprint arXiv:2210.03629 , year=

React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

arXiv
[39]

Advances in neural information processing systems , volume=

Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=
[40]

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

Vca: Video curious agent for long video understanding , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=
[41]

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

VideoSeek: Long-Horizon Video Agent with Tool-Guided Seeking , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=
[42]

arXiv preprint arXiv:2502.13923 , year=

Qwen2.5-vl technical report , author=. arXiv preprint arXiv:2502.13923 , year=

arXiv
[43]

arXiv preprint arXiv:2606.19348 , year=

Deepseek-v4: Towards highly efficient million-token context intelligence , author=. arXiv preprint arXiv:2606.19348 , year=
[44]

arXiv preprint arXiv:2412.07183 , year=

Exploring what why and how: A multifaceted benchmark for causation understanding of video anomaly , author=. arXiv preprint arXiv:2412.07183 , year=

arXiv
[45]

Advances in Neural Information Processing Systems , volume=

Advancing video anomaly detection: A concise review and a new dataset , author=. Advances in Neural Information Processing Systems , volume=
[46]

arXiv preprint arXiv:2607.00622 , year=

Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization , author=. arXiv preprint arXiv:2607.00622 , year=

arXiv

Pith tools

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

[1] [1]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Learning temporal regularity in video sequences , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

work page

[2] [2]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Real-world anomaly detection in surveillance videos , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

[3] [3]

Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

Future frame prediction for anomaly detection--a new baseline , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

[4] [4]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Weakly-supervised video anomaly detection with robust temporal feature magnitude learning , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

[5] [5]

MGFN: magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection , author=. Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence , pages=

work page

[6] [6]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Dual memory units with uncertainty regulation for weakly supervised video anomaly detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

[7] [7]

IEEE Transactions on Image Processing , volume=

Learning prompt-enhanced context features for weakly-supervised video anomaly detection , author=. IEEE Transactions on Image Processing , volume=. 2024 , publisher=

work page 2024

[8] [8]

Proceedings of the AAAI conference on artificial intelligence , volume=

Vadclip: Adapting vision-language models for weakly supervised video anomaly detection , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

work page