A proactive assistant can give a correct answer and still fail by interrupting too early—or by waiting until the useful moment has passed. The central challenge here is therefore not just video understanding: it is deciding when to act while continuously updating what the user wants.
Key Findings
- The benchmark spans explicit, implicit, and changing intents, with 85K training and 4.1K evaluation samples. This makes timing and intent inference part of the evaluation, rather than testing only answers to clearly stated prompts.
- The streaming model improves response timing and answer quality by 10.7% over the reported baselines while retaining general video-understanding ability. The result targets a practical trade-off: proactive behavior should not come at the expense of ordinary comprehension.
- A hierarchical-memory component represents intent as a persistent state that can evolve. In sustained interaction, it yields a reported 12.3% gain in proactive interaction, addressing the gap between isolated responses and coherent assistance over time.
- Implicit and dynamic intents remain difficult for existing systems. The reported improvements are progress, not evidence that an agent can reliably infer shifting goals in every long-running interaction.
Method and Trade-offs
The work combines a streaming benchmark, a model that infers intent from ongoing observations, and a memory mechanism for tracking the latest intent and environmental state. This makes it useful for studying both response quality and intervention policy. It is a good fit for researchers working on video agents and proactive interaction; look elsewhere if you need evidence of deployment readiness or robust performance across arbitrary long-horizon settings, which the reported results do not establish.