Less from More: Reinforcing Sparse Video Reasoning from Dense References
SAVER trains sparse-frame video reasoning against dense-frame references, then runs with fewer frames.
The paper proposes a dense-to-sparse post-training method for video-language models. Its reinforcement setup pairs dense and sparse video views, using grounding rewards and a reliability-gated reference reward to keep sparse predictions tied to task-relevant temporal evidence. The authors say it was trained on 1,250 temporal grounding examples, without video question-answering annotations. Across temporal grounding and video QA benchmarks, SAVER improved performance across frame budgets and could match or beat dense-frame Qwen3.5 baselines while using fewer frames. HF Daily Papers' note
The paper proposes a dense-to-sparse post-training method for video-language models. Its reinforcement setup pairs dense and sparse video views, using grounding rewards and a reliability-gated reference reward to keep sparse predictions tied to task-relevant temporal evidence. The authors say it was trained on 1,250 temporal grounding examples, without video question-answering annotations. Across temporal grounding and video QA benchmarks, SAVER improved performance across frame budgets and could match or beat dense-frame Qwen3.5 baselines while using fewer frames. HF Daily Papers' note
score 4