Less from More: Reinforcing Sparse Video Reasoning from Dense References
SAVER post-trains video models to reason from sparse frames using dense views as references.
Researchers propose SAVER, a dense-to-sparse reinforcement post-training method that uses dense video views as references so sparse-frame predictions keep task-relevant temporal evidence. It is trained on 1,250 randomly sampled temporal grounding examples, without video question-answering labels, using grounding rewards and a reliability-gated reference reward. Across three temporal grounding benchmarks and six video QA benchmarks, SAVER matches or surpasses dense-frame Qwen3.5 baselines while using substantially fewer frames.
- SAVER treats dense video views as training references for sparse-frame inference.
- Training uses 1,250 temporal grounding samples and no video QA labels.
- It matches or beats dense-frame Qwen3.5 baselines with far fewer frames.
- Results span three grounding and six video question-answering benchmarks.
- Tags
- savervideo-language-modelssparse-reasoningtemporal-groundingqwenreinforcement-learningpost-training
- AI models
- Qwen3.5
From the original article
1 min read at huggingface.co
Video-language models commonly assume that more temporal observations lead to more reliable reasoning. We question this assumption and argue that the key challenge is not merely processing more video frames efficiently, but learning to reason reliably under limited temporal evidence.
By Wenfang Sun, Yingjun Du, Cees G. M. Snoek · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.