lingcco/EventMemAgent-8B

1---2license: apache-2.03base_model: Qwen/Qwen3-VL-8B-Instruct4library_name: transformers5pipeline_tag: image-text-to-text6tags:7  - video-understanding8  - reinforcement-learning9  - eventmemagent10---11 12# EventMemAgent-8B13 14Model for **EventMemAgent: Hierarchical Event-Centric Memory for Online Video15Understanding with Adaptive Tool Use** (ECCV 2026).16 17- [Paper](https://arxiv.org/abs/2602.15329)18- [Code](https://github.com/lingcco/EventMemAgent)19- [Processed data](https://huggingface.co/datasets/lingcco/EventMemAgent)20 21## Model22 23EventMemAgent uses a Qwen3-VL-8B-Instruct policy with hierarchical event-centric24memory and adaptive memory-retrieval, OCR and object-detection tools. Event25captions are generated separately by a frozen Qwen3-VL-4B-Instruct model;26event/query embeddings use Qwen3-Embedding-0.6B.27 28This upload contains the merged actor from training run `1-30_8B`, step **150**,29in standard Hugging Face format. It includes the model, tokenizer, processor and30chat template. Optimizer states, training logs and evaluation predictions are31not included.32 33## Training34 3510K MovieChat samples annotated by VideoMarathon; multi-turn GRPO with group36size 8, global batch 64, PPO minibatch 32 and per-GPU microbatch 1. AdamW uses37a learning rate of 1e-6, five warmup steps, one epoch and no KL loss. Training38prompt/response limits are 8192 tokens each, with at most ten assistant turns.39 40## Usage41 42Download the checkpoint and configure `MODEL_PATH` in the accompanying code:43 44```bash45hf download lingcco/EventMemAgent-8B --local-dir checkpoints/released/EventMemAgent-8B46```47 48Use the code repository's inference and evaluation entry points. The checkpoint49alone does not implement streaming memory or provide OCR and detection services;50these are separate components of the pipeline.51 52## Paper results53 54| OVO-Bench | StreamingBench |55| ---: | ---: |56| 60.75 | 77.00 |57 58Results reported in the paper.59 60## License and limitations61 62The model is released under Apache 2.0, following its Qwen3-VL-8B-Instruct63base model. The research model can64produce incorrect answers; memory captions and perception tools can also make65errors. It is not intended for safety-critical use.66