A streaming model can receive a good overall score for the wrong sequence of decisions: an early useful answer may give later, redundant replies undeserved credit. The key insight here is to judge each turn by the return still available from that point, making response timing—not just the final trajectory—the unit of learning.
Key Findings
- Silence-completed prefix scoring estimates the reward earned before each decision by retaining earlier actions and making later turns silent. Subtracting this accrued reward gives later turns credit based on their remaining return.
- Turn-mean optimization averages token losses within each turn, then across turns. This prevents long responses from dominating and gives waiting decisions a meaningful place in the update.
- The combined method improves benchmark aggregates across ProactiveVQA, OVO-Timing, and StreamingBench using two Qwen vision-language backbones. On the 3B backbone, it reaches 43.56 PAUC, 21.59 average Timing F1, and 49.43 average StreamingBench accuracy.
- The components complement each other: on the 3B model, combining them reduces repetition from 15.53 to 5.98 compared with turn-level credit alone, while also improving timing and streaming accuracy.
Method and Fit
The approach uses the existing trajectory reward scorer to evaluate silence-completed prefixes, rather than training a critic or collecting additional rollouts. It is a good fit for research on proactive, multi-turn streaming video systems where deciding whether to speak is part of the task. Look elsewhere if you need evidence for static-image understanding or broad general-purpose video tasks: the reported results are specific to streaming benchmarks and the evaluated backbones.