SRVAU-R1: Enhancing Video Anomaly Understanding via Reflection-Aware Learning

TL;DR

The paper tackles the gap in video anomaly understanding (VAU) where multi-modal LLM-based systems rely on surface-level descriptions. It introduces SRVAU-R1, a reflection-aware framework that embeds explicit self-reflection into reasoning via a reflection-oriented CoT data construction and a two-stage training pipeline (reflection-enhanced supervised fine-tuning and reflection-aware reinforcement fine-tuning). A first-of-its-kind reflection CoT dataset structures samples into initial reasoning, self-reflection, and revised reasoning, enabling robust self-diagnosis and correction. The RL stage uses GRPO with a composite reward that includes task accuracy, reflection quality, and a temporal IoU component to align reasoning with temporal anomaly spans. Empirical results across VAU benchmarks show SRVAU-R1 achieves significant gains in both temporal localization and reasoning quality, demonstrating the value of explicit reflection for reliable, interpretable VAU in complex video scenarios.

Abstract

Multi-modal large language models (MLLMs) have demonstrated significant progress in reasoning capabilities and shown promising effectiveness in video anomaly understanding (VAU) tasks. However, existing MLLM-based approaches remain largely focused on surface-level descriptions of anomalies, lacking deep reasoning over abnormal behaviors like explicit self-reflection and self-correction. To address that, we propose Self-Reflection-Enhanced Reasoning for Video Anomaly Understanding (SRVAU-R1), a reflection-aware learning framework that incorporates reflection in MLLM reasoning. Specifically, SRVAU-R1 introduces the first reflection-oriented Chain-of-Thought dataset tailored for VAU, providing structured supervision with initial reasoning, self-reflection, and revised reasoning. Based on that, it includes a novel reflection-aware learning paradigm with supervised fine-tuning and reinforcement fine-tuning to enhance multi-modal reasoning for VAU. Extensive experiments on multiple video anomaly benchmarks demonstrate that SRVAU-R1 consistently outperforms existing methods, achieving significant improvements in both temporal anomaly localization accuracy and reasoning quality.

Figures (4)

• Figure 1: Existing VAU methods such as VAU-R1 provide only surface-level analysis and lack reflection on the generated results. In this paper, we propose a new method named SRVAU-R1 to include explicit self-reflection in VAU to correct or refine initial reasoning.
• Figure 2: Overview of the SRVAU-R1 framework. First, SRVAU-R1 includes the (a) reflection-aware data construction pipeline, where initial chain-of-thought candidates and explicit self-reflection annotations are generated under the guidance of an advanced model. SRVAU-R1 can be flexibly applied in VAU including multi-choice question answering and temporal anomaly grounding, which requires structured reasoning and reflection. Finally, SRVAU-R1 includes a novel (b) learning paradigm consisting of reflection-enhanced SFT and RFT. During the RL process, model parameters are optimized via task, reflection, and temporal IoU rewards.
• Figure 3: Qualitative results of (a) QA and (b) TAG tasks. All ground truth answers and anomaly intervals are highlighted in red. VAU-R1 and SRVAU-R1 reason under the same CoT prompting. SRVAU-R1 leverages explicit self-reflection to produce more accurate and interpretable answer selection and temporal localization, whereas VAU-R1 relies on partial cues and yields less precise predictions.
• Figure 4: Overview of Dataset Composition and Semantic Distribution. (a) Dataset split statistics for the QA and Temporal Grounding tasks, including train, validation, and test splits. (b) Word cloud visualization constructed from the reflection texts in the training data, highlighting dominant semantic cues learned during reflection-based anomaly reasoning.