acceptodds
Withdrawn from ICLR 2027

NEXUS: Neuro-Symbolic Event Reasoning for Training-Free Video Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Training-free video anomaly detection is attractive for real-world surveillance, where collecting representative training data is often constrained by privacy, data-governance requirements, and the inherent rarity of abnormal events. However, existing training-free approaches remain largely driven by visual affinity, either through direct vision-language similarity or by prompting multi-modal models to infer anomaly decisions from appearance-level evidence. We argue that this view is incomplete: real-world complex anomalies are not only isolated appearances, but also temporally evolving events whose interpretation depends on how semantic states emerge, persist, change, and resolve over time. To this end, we introduce Neuro-Symbolic Event eXploration with Unified Spatio-temporal VLLM Verification (NEXUS), a novel training-free framework for video anomaly detection. At its core, NEXUS introduces Neuro-symbolic Temporal Event Reasoning (N-TER), which transforms frozen vision-language evidence into five continuous event states and composes them through interpretable temporal chains capturing event onset, reversal, persistence, consequence, resolution, and prior history. Complementing this temporal reasoning, Uncertainty-guided Proposal Scoring (UPS) explicitly enforces an anomaly hypothesis to outperform both uncertain and normal alternatives, preventing visual similarity alone from becoming anomaly evidence. Then, Reasoning-Enabled Score Refinement (RESR) couples UPS proposals with N-TER consistency and contradiction, allowing temporal reasoning to refine visual hypotheses and also minimizing anomaly hallucination. A frozen VLLM Judge then performs the final spatio-temporal verification with the symbolic evidence. We evaluate NEXUS across five real-world challenging video anomaly detection benchmarks and demonstrate strong performance gain against state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_NCAdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NEXUS, a training-free framework for video anomaly detection that aims to address the limitations of appearance-based similarity and implicit temporal reasoning in existing approaches. NEXUS combines frozen OpenCLIP features with three components: an uncertainty-guided proposal scoring module (UPS), a neuro-symbolic temporal event reasoning module (N-TER), and a reasoning-enabled score refinement module (RESR), followed by a frozen Qwen3-VL-based multimodal judge for final verification. In particular, N-TER introduces a manually designed event ontology and maps visual-text similarities into five semantic states—anomaly, uncertainty, onset, continuation, and resolution—which are connected through handcrafted fuzzy temporal chains to explicitly model the evolution of anomalous events. Experiments on five video anomaly detection benchmarks demonstrate consistent improvements over the OpenCLIP baseline and strong performance against existing training-free methods, with ablations showing the contributions of uncertainty modeling and temporal event descriptions. Overall, the paper presents a well-engineered and interpretable training-free VAD pipeline, although its methodological novelty is somewhat unclear, as the framework largely combines frozen vision-language models, a handcrafted semantic ontology, deterministic temporal rules, and a large VLM-based verifier.

Critical strengths
  1. Well-motivated problem formulation. The paper identifies an important limitation of existing training-free video anomaly detection methods: anomaly detection based primarily on visual similarity may confuse anomalous events with visually similar but benign activities. The explicit modeling of uncertainty as competing semantic interpretations is a meaningful perspective and provides a reasonable motivation for the proposed framework.

  2. Explicit and interpretable temporal reasoning. NEXUS introduces an explicit event-state representation consisting of anomaly, uncertainty, onset, continuation, and resolution, and organizes these states through temporal reasoning chains. Compared with approaches that rely on a VLM to implicitly infer temporal relations, this design makes the temporal reasoning process more structured and interpretable, while allowing the contribution of different event states to be analyzed through ablations.

  3. Strong empirical performance and comprehensive component analysis. The proposed framework achieves competitive performance across multiple video anomaly detection benchmarks and consistently improves over the OpenCLIP-based baseline. The ablation studies further investigate the contributions of uncertainty modeling, temporal event states, and the proposed temporal reasoning chains, providing useful evidence that the individual components contribute to the final performance.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework appears to largely combine existing components, including frozen OpenCLIP similarity, manually designed semantic descriptions, deterministic temporal rules, and a frozen Qwen3-VL judge. While the overall pipeline is carefully engineered, the paper does not sufficiently establish that NEXUS introduces a fundamentally new neuro-symbolic reasoning mechanism beyond this combination of existing techniques.

  2. Heavy reliance on a handcrafted ontology. NEXUS relies on a manually constructed ontology containing anomaly, uncertainty, onset, continuation, and resolution descriptions. This ontology encodes substantial prior knowledge about anomaly mechanisms and their benign alternatives. The paper should provide more evidence regarding how sensitive performance is to the design of these descriptions and whether the same approach can generalize to anomaly types or domains that are not anticipated during ontology construction.

  3. The contribution of temporal reasoning is not sufficiently isolated. The description-bank ablation shows that progressively adding onset, continuation, and resolution descriptions already produces substantial performance gains. This makes it unclear how much of the improvement comes from the proposed N-TER temporal reasoning itself versus simply providing additional temporally related semantic prompts. A stronger baseline using the same description bank but without the temporal graph and fuzzy reasoning would be necessary to isolate the contribution of N-TER.

  4. Insufficiently controlled comparison with direct VLM reasoning. N-TER exploits a broader temporal neighborhood of up to five intervals, whereas the Qwen3-VL judge normally receives only the Past/Target/Future visual groups. The paper therefore does not fully rule out the possibility that the gains arise from providing additional temporal information through the symbolic pathway rather than from neuro-symbolic reasoning itself. A direct Qwen3-VL baseline with the same temporal context would make the comparison substantially more convincing.

  5. Limited validation of the handcrafted temporal rules. The temporal chains in N-TER are manually specified, and the reported chain ablation does not sufficiently demonstrate that these rules are robust across different datasets and anomaly categories. It remains unclear whether the selected temporal relations reflect general properties of anomalous events or are particularly effective for the benchmark distributions used in the experiments. Per-dataset and per-anomaly-category analyses would strengthen the claim of general temporal reasoning.

  6. High inference cost despite being training-free. Although NEXUS requires no trainable parameters, its inference cost is dominated by the Qwen3-VL judge, which accounts for the vast majority of the reported runtime. This raises concerns about the scalability of the method to long-duration or real-time surveillance scenarios. A more detailed analysis of the accuracy–latency trade-off and the performance of lighter verification models would help clarify the practical value of the proposed training-free design.

Questions for authors
  1. Terminology: The paper uses “VLLM” to refer to the vision-language model. Could the authors clarify and consistently use “VLM” or “LVLM” instead, to avoid confusion with VLLM, the commonly used inference engine for large language models?

  2. Citation format: The manuscript currently uses \cite for citations. Could the authors consistently use \citep where parenthetical citations are intended, to ensure consistency with the paper's citation style?

Reviewer_vnsAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a training-free video anomaly detection framework named NEXUS. The authors point out that existing training-free methods largely rely on appearance-level vision-language similarity matching or implicit temporal reasoning by vision-language models, making it difficult to distinguish visually similar ambiguous normal behaviors (such as running for exercise vs. escaping), and lacking explicit modeling of the dynamic evolutionary process of anomalous events.

Critical strengths

NEXUS proposes a training-free video anomaly detection framework that suppresses false positives from visually similar ambiguous normal behaviors via Uncertainty-guided Proposal Scoring (UPS) and transforms CLIP features into interpretable five-state fuzzy temporal logic evidence using Neuro-symbolic Temporal Event Reasoning (N-TER). Building upon this, it performs multi-scale spatio-temporal verification by combining Reasoning-Enabled Score Refinement (RESR) with a frozen large vision-language model (Qwen3-VL-8B). This method demonstrates robust generalization performance across five public datasets.

Critical weaknesses

1.The framework achieves strong detection performance, but the gains are not clearly attributable to explicit event reasoning rather than richer descriptions, broader context, or multi-scale processing. Existing ablations do not fully isolate these factors; for example, UPS and the VLLM Judge are introduced together. Comparisons with simple temporal aggregation, structured prompting, and temporally shuffled controls under matched inputs and inference settings would better establish the distinctive contribution. 2.The method relies on eight predefined anomaly mechanisms and handcrafted event-phase descriptions, constraining its semantic coverage. Sharing the ontology across datasets is valuable, but the reported results do not establish recognition of uncovered anomalies or adaptation to context-dependent normality. Evaluations involving unseen mechanisms and behaviors whose abnormality changes across settings would help distinguish cross-domain transfer from generalization to unknown anomalies. 3.Several temporal rules share or subsume the same conditions; for example, Chain-I adds only an onset condition to Chain-II, yet probabilistic OR counts both toward higher consistency. Six valid rule scores of 0.5 already produce approximately 0.984, suggesting that high consistency may partly reflect rule quantity rather than stronger event evidence. Redundant-rule removal, maximum aggregation, and sensitivity to rule count would help validate the score’s interpretation. 4.Equation (1) defines RESR additively, whereas Appendix C.2 uses a multiplicative gate (minimum 0.5). This discrepancy affects small-margin behavior and reproducibility. The authors must clarify the exact formulation and reconcile the main text, appendix, and ablations. 5.The paper lacks evaluation of operational false-alarm rates, scene transfer, and real-time latency (due to future-interval dependence). To substantiate practical reliability, the authors should include fixed-threshold false alarm metrics, cross-scene threshold transfer tests, and accuracy-latency evaluations using strictly past and current observations.

Questions for authors

1.Explain the data contradiction between Table 2(a) and Table 2(c): Why does using only the A+N description banks (without U, O, C, R) in Table 2(c) achieve 82.06 AUC, whereas Table 2(a) yields only 81.33 AUC despite using + UPS + VLLM-Judge (which includes A, N, U)? 2.Clarify the true mathematical formula of RESR: Please provide the exact mathematical expression of RESR in the actual code, and explain why Equation (1) in the main text is an additive formula, whereas a multiplicative gate g_{min} = 0.5 appears in Appendix C.2 and Table 6? 3.Verify the fairness of the test set split and baseline comparison: Please clarify whether the test set splits for UBnormal (211 videos) and MSAD (360 videos) in Table 3 are completely identical (with the exact same frame counts and video lists) to all comparative methods cited in Table 1 (such as VadCLIP, GS-MoE, VADTree, etc.)? 4.The issues mentioned in the weaknesses need to be addressed.

Reviewer_a7YTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NEXUS, a video anomaly detection framework that requires no training on the target dataset. It first uses frozen OpenCLIP and fixed description banks to compute support for anomalous, competing, and normal interpretations, and forms anomaly proposals through UPS. The semantic evidence is then converted into five continuous event states. Six types of fuzzy temporal chains are evaluated over a fixed temporal neighborhood, and RESR refines the proposal scores. A frozen Qwen3-VL-8B performs final verification using video frames and structured evidence, followed by multi-scale prediction fusion.

The paper evaluates NEXUS on five benchmarks: UCF-Crime, XD-Violence, MSAD, SVAD, and UBnormal. It reports 88.32 AUC on UCF-Crime and 76.44 AP on XD-Violence. The experiments include direct frozen-VLLM inference, existing training-free methods, component ablations, description-bank ablations, and multi-scale analysis. The results support the effectiveness of the combined pipeline, but important questions remain about the definition and independent contribution of its core reasoning mechanism.

Critical strengths
  1. The evaluation covers diverse benchmarks and obtains competitive results with a fixed configuration.

    The paper evaluates five heterogeneous benchmarks using the same description banks, rules, and coefficients. In Table 1, NEXUS achieves 88.32 AUC on UCF-Crime and 76.44 AP on XD-Violence. The authors also explicitly acknowledge that its UBnormal AUC and overall MSAD AUC are below the strongest training-free results listed in the table. These results support the applicability of a fixed semantic interface across benchmarks.

  2. The existing ablations provide substantive support for the main design choices.

    In Table 2(a), adding N-TER to UPS and the Judge improves UCF-Crime AUC from 81.33 to 87.54, and adding RESR further raises it to 88.32. In Table 2(c), introducing descriptions of competing interpretations improves AUC from 82.06 to 84.57. The appendix also distinguishes single-scale from multi-scale results. These experiments suggest that the combined design merits further investigation, although the sources of the individual gains are not yet fully established.

Critical weaknesses
  1. The definition of the core algorithm is inconsistent with its implementation description.

    Equation (1) on page 7 uses additive score refinement, whereas Appendix C.2 describes a multiplicative gate with a lower bound of g=0.5g=0.5. It is therefore unclear which version was used in the experiments. Moreover, leaving non-positive proposals unchanged does not ensure that positive proposals retain their sign, nor does it constrain the final Judge's output. Please reconcile the equation, implementation, and ablation configurations, and clarify the scope of the conservativeness claim.

  2. The interpretation of the reasoning scores and the independent role of the rules are not sufficiently validated.

    The definitions on page 5 imply A+U≤1A+U\leq1, so AUAU is closer to an ambiguity penalty when competing interpretations are tied than to simultaneous strong support for two hypotheses. When all six chains have the neutral value 0.50.5, fuzzy OR can nevertheless produce κ=0.984375\kappa=0.984375. This suggests that a high score may partly reflect the accumulation of overlapping rules and cannot directly be interpreted as evidence of a coherent event. This example does not imply that a tie produces a positive proposal. The existing ablations have also not separated the effects of additional symbolic context from rule computation. I suggest controls using rule deduplication or max aggregation, together with a Judge that receives the same context without explicit rules.

  3. The comparison omits important strong baselines, limiting the assessment of competitive advantages.

    The paper cites PRISM and SphereVAD but does not compare against them in Table 1. The Qwen3-VL-Embedding-2B version of PRISM (ICML 2026) reports 88.85 UCF AUC and 77.48 XD AP, both higher than this paper's 88.32 and 76.44. Its PerceptionEncoder-L14 version achieves 80.31 XD AP. SphereVAD also reports a higher XD AP of 86.99, although it uses synthetic calibration and cross-test-video statistics. These complete configurations should be included, with differences in backbones, textual priors, and additional data resources distinguished, and the claimed performance advantages reassessed.

Questions for authors
  1. Were the results in Tables 1 and 2 obtained using the additive RESR in the main text or the multiplicative gate in the appendix? Please provide the complete definition, clarify whether a gate lower bound and clipping are used, and explain whether the sign-preservation and hallucination-suppression guarantees apply only to intermediate proposals.

  2. How do the authors interpret A+U≤1A+U\leq1, AU≤1/4AU\leq1/4, and the case where neutral states yield κ=0.984375\kappa=0.984375? Are the main gains retained after removing overlapping rules, using max aggregation, or replacing the penalty with a simple margin-based alternative?

  3. With the same five-node semantic context, visual inputs, and scale settings, how does providing raw states directly to the Judge compare with explicit rule computation? Are these conditions held fixed across the rows of Table 2 and for the direct Qwen baseline?

  4. Could the authors provide all eight mechanisms, the complete set of 74 descriptions, and the actual Judge prompt templates, together with the rationale for selecting these texts? If cross-model applicability is emphasized, could the complete pipeline be evaluated with one different Judge, rather than only comparing its direct inference results? A full grid of model combinations is unnecessary.

  5. Which UNKNOWN-handling switch accounts for the improvement from I–VI to I–VII in Table 2(b)? Does the end-to-end runtime in the appendix include the four-scale inference and fusion used for the main results? Please explicitly map the timing configuration to the main-result configuration.

  6. Why does Table 1 omit comparisons with the cited PRISM and SphereVAD? Could the authors include their complete configurations and originally reported results, explain differences in textual priors, backbones, synthetic calibration, and test-pool statistics, and provide comparisons under matched protocols where feasible?

Reviewer_W2sXOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NEXUS, a training-free video anomaly detection framework combining uncertainty-guided proposal scoring, neuro-symbolic temporal event reasoning, and VLLM-based verification. Experiments across five datasets demonstrate competitive performance, with ablations supporting the benefits of explicit temporal reasoning.

Critical strengths

Explicit temporal event reasoning offers an interpretable alternative to appearance-based anomaly detection.

Evaluations across five benchmarks and detailed ablations demonstrate consistent benefits from the proposed reasoning components.

Critical weaknesses

The framework relies on manually designed semantic descriptions and temporal rules, raising concerns about scalability and generalization to unseen anomaly types.

The multi-stage pipeline involves repeated VLLM inference, but computational overhead and practical efficiency are insufficiently evaluated.

Questions for authors

How sensitive is NEXUS to changes in the manually constructed ontology and temporal rules?

Can the authors report inference latency and computational cost compared with existing training-free methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.