VideoJev Releases System One Model for Video Segment Finding and QA

Seven institutions, including Nanjing University MCG-NJU and JD.com, released the VideoJev model, which performs Temporal Grounding and multiple-choice QA by scoring candidates without text generation. The model adopts a System One approach, removing the LM head and final MLP and attaching a candidate scorer to make decisions in a single forward pass.

Model Architecture and Performance

VideoJev was retrained based on Qwen3.5-4B and Qwen3.6-35B-A3B. It outputs sigmoid probabilities per segment for temporal grounding and softmax probabilities per option for multiple-choice QA. Demo results showed the 4B model detecting segment [3.0, 10.0] and the correct answer A (Pasta, probability 0.9999), while the 35B-A3B model detected a shorter segment [3.0, 8.0].

Limitations and Execution Environment

The current architecture encodes the entire video first, so handling real-time streaming input is noted as a future task. For questions requiring detailed verification, such as the intensity of small objects or intent inference, the model may produce incorrect answers despite high probabilities. The code and weights are released under the Apache License 2.0, permitting research and commercial use, and can be run on Hugging Face with the trust_remote_code=True option.