---
license: apache-2.0
base_model: Qwen/Qwen3-VL-8B-Instruct
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - video-understanding
  - reinforcement-learning
  - eventmemagent
---

# EventMemAgent-8B

Model for **EventMemAgent: Hierarchical Event-Centric Memory for Online Video
Understanding with Adaptive Tool Use** (ECCV 2026).

- [Paper](https://arxiv.org/abs/2602.15329)
- [Code](https://github.com/lingcco/EventMemAgent)
- [Processed data](https://huggingface.co/datasets/lingcco/EventMemAgent)

## Model

EventMemAgent uses a Qwen3-VL-8B-Instruct policy with hierarchical event-centric
memory and adaptive memory-retrieval, OCR and object-detection tools. Event
captions are generated separately by a frozen Qwen3-VL-4B-Instruct model;
event/query embeddings use Qwen3-Embedding-0.6B.

This upload contains the merged actor from training run `1-30_8B`, step **150**,
in standard Hugging Face format. It includes the model, tokenizer, processor and
chat template. Optimizer states, training logs and evaluation predictions are
not included.

## Training

10K MovieChat samples annotated by VideoMarathon; multi-turn GRPO with group
size 8, global batch 64, PPO minibatch 32 and per-GPU microbatch 1. AdamW uses
a learning rate of 1e-6, five warmup steps, one epoch and no KL loss. Training
prompt/response limits are 8192 tokens each, with at most ten assistant turns.

## Usage

Download the checkpoint and configure `MODEL_PATH` in the accompanying code:

```bash
hf download lingcco/EventMemAgent-8B --local-dir checkpoints/released/EventMemAgent-8B
```

Use the code repository's inference and evaluation entry points. The checkpoint
alone does not implement streaming memory or provide OCR and detection services;
these are separate components of the pipeline.

## Paper results

| OVO-Bench | StreamingBench |
| ---: | ---: |
| 60.75 | 77.00 |

Results reported in the paper.

## License and limitations

The model is released under Apache 2.0, following its Qwen3-VL-8B-Instruct
base model. The research model can
produce incorrect answers; memory captions and perception tools can also make
errors. It is not intended for safety-critical use.