Researchers have introduced VideoJev, a family of models that reframes video understanding and interaction as structured decision tasks rather than open-ended text generation. The work builds on the emerging System One paradigm, in which visual agents must convert observations into timely decisions about what to do next.
Instead of decoding answers autoregressively, VideoJev adapts pretrained multimodal large language models (MLLMs) by attaching a shared head that scores temporal intervals and textual options directly against video evidence. Those scores produce either event spans or answer choices, unifying temporal grounding with multiple-choice video question answering. The approach aims to combine the semantic flexibility of MLLMs with decisions that agents can act on directly.
Because an interaction generates new goals and intermediate results that prompt further queries about the same visual history, VideoJev reuses visual computation in two ways: Cached inference handles queries as they arrive, while Joint inference answers multiple queries in a single forward pass.
On three temporal grounding benchmarks, VideoJev-4B and VideoJev-35B-A3B improved mean mIoU over their backbones by 6.7 and 10.8 points respectively, with measured inference speedups of 2.8–3.1×. On four video QA benchmarks, both models matched their backbones in Thinking mode while exceeding 100× faster model forward passes.
The authors describe the result as a unified, reusable decision interface for composing structured visual judgments into complex, open-ended workflows. Training and evaluation code and both checkpoints have been released.