StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding
Post-training alone pushes a memory-free streaming-video setup close to its larger teacher.
The paper fixes a recent-window streaming protocol and tests how far training changes can go without extra inference memory. StreamOPD combines verifiable streaming-video data, thinking-mode on-policy distillation, and instruct-mode deployment, lifting StreamingBench from 77.9% to 83.9%. Its ST-CueGate extension reweights distillation using teacher likelihood ratios and is the only variant reported above the base model on all four benchmarks. The authors also report that self-distillation keeps most gains and improves hallucination detection to 57.0%. HF Daily Papers' note
The paper fixes a recent-window streaming protocol and tests how far training changes can go without extra inference memory. StreamOPD combines verifiable streaming-video data, thinking-mode on-policy distillation, and instruct-mode deployment, lifting StreamingBench from 77.9% to 83.9%. Its ST-CueGate extension reweights distillation using teacher likelihood ratios and is the only variant reported above the base model on all four benchmarks. The authors also report that self-distillation keeps most gains and improves hallucination detection to 57.0%. HF Daily Papers' note
score 5