MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding

MARC (Memory-Augmented RL Token Compression), accepted at ICLR 2026.

Quick Start

Inference Example

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
import torch

# Load model
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "path/to/model",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("path/to/model")

# Prepare video input
messages = [{
    "role": "user",
    "content": [
        {"type": "video", "video": "path/to/video.mp4"},
        {"type": "text", "text": "What is happening in this video?"}
    ]
}]

# Generate with compression
inputs = processor(
    messages=messages,
    videos=videos,
    compress=True,  # Enable compression
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, compress=True, max_new_tokens=512)
response = processor.decode(outputs[0], skip_special_tokens=True)

See inference_script/inference_example.py for a complete example.