lingcco/EventMemAgent-8B
EventMemAgent-8B by lingcco is an 8 billion parameter model designed for online video understanding, utilizing a Qwen3-VL-8B-Instruct policy. It integrates hierarchical event-centric memory and adaptive memory-retrieval, OCR, and object-detection tools. This model excels in processing video data by leveraging event captions generated by a separate Qwen3-VL-4B-Instruct model and embeddings from Qwen3-Embedding-0.6B, making it specialized for complex video analysis tasks.
Loading preview...
Overview
EventMemAgent-8B is a specialized 8 billion parameter model developed by lingcco for online video understanding, detailed in the paper "EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use" (ECCV 2026). It is built upon a Qwen3-VL-8B-Instruct policy and incorporates a sophisticated architecture that includes hierarchical event-centric memory and adaptive tools for memory retrieval, Optical Character Recognition (OCR), and object detection. The model processes video data by integrating event captions generated by a frozen Qwen3-VL-4B-Instruct model and uses Qwen3-Embedding-0.6B for event and query embeddings.
Key Capabilities
- Hierarchical Event-Centric Memory: Manages video information through a structured, event-focused memory system.
- Adaptive Tool Use: Dynamically employs memory-retrieval, OCR, and object-detection tools to enhance video understanding.
- Online Video Understanding: Designed for real-time or near real-time analysis of video content.
- Integrated Architecture: Combines multiple specialized models (Qwen3-VL-8B-Instruct, Qwen3-VL-4B-Instruct, Qwen3-Embedding-0.6B) for comprehensive processing.
Training Details
The model was trained on 10,000 MovieChat samples, annotated by VideoMarathon, using multi-turn GRPO with specific group, global batch, PPO minibatch, and microbatch sizes. It utilized AdamW with a learning rate of 1e-6, five warmup steps, and one epoch, without KL loss. Training prompt/response limits were set at 8192 tokens each, allowing up to ten assistant turns.
Performance
EventMemAgent-8B achieved notable results on video understanding benchmarks, scoring 60.75 on OVO-Bench and 77.00 on StreamingBench, as reported in the associated research paper.
Good for
- Researchers and developers working on advanced online video understanding systems.
- Applications requiring detailed event-based analysis of video content.
- Projects that can leverage adaptive tool use for perception tasks like OCR and object detection within video streams.