SyncPro-SFT

SyncPro-SFT is the official supervised checkpoint for SyncPro: Synchronous Reasoning and Responding for Proactive Streaming Video Understanding.

SyncPro synchronously decodes an internal reasoning stream and an external response stream under a causal streaming attention map.

Model

SyncPro forks the decoder into synchronous internal-reasoning and external-response branches. At each streaming step, it learns whether to remain silent, reason internally, or speak, while preserving causal processing.

This checkpoint was initialized from Qwen/Qwen3.5-4B and trained on 189,403 streaming trajectories using:

• split action loss over the dual-stream grammar;
• Action Temporal Ranking (ATR) against nearby hard negatives;
• dual LoRA with rank 64, alpha 128, and dropout 0.05;
• fixed additive branch merging with gamma = 0.5.

The release contains merged BF16 inference weights. Optimizer shards, scheduler state, RNG state, and other resume-only training artifacts are intentionally excluded. Inference code and benchmark adapters are documented in the SyncPro repository.

Data and intended use

The model is intended for research on causal streaming-video understanding, proactive response timing, and separated internal/external generation. Training annotations are available in SyncPro-189k. Source videos are not redistributed and must be obtained from their original publishers under the applicable terms.

The checkpoint may inherit limitations and biases from Qwen3.5-4B and the public research datasets used for training. It should not be treated as a safety-critical monitoring system.

License

The model weights are released under the Apache License 2.0. The annotation dataset has separate licensing and source-specific terms described on its dataset card.

Citation

@misc{tian2026syncprosynchronousreasoningresponding,
      title={SyncPro: Synchronous Reasoning and Responding for Proactive Streaming Video Understanding}, 
      author={Xueyun Tian and Bingbing Xu and Jiaming Liu and Jinpeng Yu and Yang Shi and Yu Zhao and Ruihua Huang and Yuanzhuo Wang and Huawei Shen},
      year={2026},
      eprint={2610.32132},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.32132}, 
}