TL;DR: We identify critical gaps in video token pruning and advance two predominant paradigms: saliency-based token selection (🔑 signal: attention weight) and diversity-oriented token merging (🔑 signal: cosine similarity).

✨ Highlights

Main contributions: We analyze the characteristics of pivotal visual signals and rethink how to utilize them more effectively (More details and findings in our paper).

(Left) Motivation of our method. (Right) Method overview.

• Attention Weight: The distribution is multi-modal and long-tailed, which a vanilla Top-k strategy fails to capture accurately.
  ➡️ Our approach: Expand the candidate set (cover the tail) and perform intra-cluster selection (cover diverse modes).

• Cosine Similarity: Direct similarity-based clustering often creates fragmented clusters, leading to noisy representations after average pooling.
  ➡️ Our approach: Inject a spatio-temporal locality prior (for smoothness) using our proposed ST-RoPE.

🛠️ Quick Setup

1. Create a conda virtual environment and install the required packages.

conda create -n Tango python=3.10
conda activate Tango
pip install -r requirements.txt

2. Install Flash Attention 2.

pip install -U flash-attn --no-build-isolation

3. Install evaluation frameworks.

# For main performance evaluation
pip install -e ./VLMEvalKit
# For efficiency analysis
pip install -e ./lmms-eval

💡 Evaluation

Performance Evaluation

We adopt the VLMEvalKit framework for performance evaluation, with retention ratios in {0.1, 0.15, 0.2}.

We currently support LLaVA-OneVision-7B, LLaVA-Video-7B, and Qwen2.5-VL-7B models.

cd VLMEvalKit/
# Evaluate on LLaVA-OneVision-7B
bash run_eval_ov.sh
# Evaluate on LLaVA-Video-7B (w/ intra-LLM pruning)
bash run_eval_video.sh
# Evaluate on Qwen2.5-VL-7B
bash run_eval_qwen.sh

Efficiency Profiling

We adopt the lmms-eval framework for efficiency profiling. Here's a sample script for evaluation under retention ratio 0.1.

WRAPPER=tango accelerate launch --num_processes=8 \
  -m lmms_eval \
  --model llava_onevision \
  --model_args pretrained=lmms-lab/llava-onevision-qwen2-7b-ov,conv_template=qwen_1_5,model_name=llava_qwen,attn_implementation=flash_attention_2 \
  --tasks videomme \
  --batch_size 1 \
  --log_samples \
  --log_samples_suffix llava_onevision \
  --output_path ./logs/

For reference, our results with 8 NVIDIA A800 GPUs are:

Metric	Value
Total_runtime (s)	1315.02
Total_GPU_runtime (s)	198.31
Peak_mem (GB)	18.61
Avg_ViT_Time (ms)	335.57
Avg_Other_Time (ms)	146.59
Avg_LLM_Prefill_Time (ms)	80.64
Avg_Total_TTFT (ms)	562.80
Avg_Decoding_Throughput (token/s)	83.66

🙌 Related Projects

• Sparrow: An efficient training scheme for video LLMs.
• Awesome-MLLM: A project keeping track of new papers and the latest developments in the field of MLLMs.

🌻 Acknowledgement

• Inspiring works with open-sourced implementation: VisionZip, FastVID, HoliTom.
• Our efficiency profiling implementation is extended upon VidCom².

🖋️ Citation

If you find our project useful, please consider citing our paper:

@article{yin2026tango,
  title={Tango: Taming Visual Signals for Efficient Video Large Language Models},
  author={Yin, Shukang and Zhao, Sirui and Wang, Hanchao and Jia, Baozhi and Wang, Xianquan and Fu, Chaoyou and Chen, Enhong},
  journal={arXiv preprint arXiv:2604.09547},
  year={2026}
}