SAVER: 1,250 grounding samples train sparse-frame video reasoning to match dense baselines
AmsUni · hf · 2026-10-12
The SAVER paper proposes a dense-to-sparse post-training framework that questions the assumption that more temporal observations yield more reliable video reasoning.
Method
- Reinforcement post-training optimizes paired dense and sparse views with grounding rewards and a reliability-gated reference reward, encouraging sparse-view predictions to preserve task-relevant temporal evidence
- Trained on only 1,250 randomly sampled temporal grounding examples, with no video QA annotations
Results
- Consistent gains across 3 temporal grounding benchmarks and 6 video QA benchmarks at various frame budgets
- Matches or surpasses dense-frame Qwen3.5 baselines while using substantially fewer frames
Takeaway: temporal grounding serves as an effective evidence-localization proxy for sparse video reasoning that transfers to broader video understanding.
More from Research
- Nature: The century-old debate over whether adult human brains grow new neurons remains unresolved — AnnaCiaunica · 2026-10-12
- Half of a study's 1,168-row dataset are duplicated observations, replication watchdog finds — RexDouglass · 2026-10-12
- Replication watchdog admits failure after independent researcher flags data irregularities — RexDouglass · 2026-10-12
- Brazilian data looks randomly generated: report cites bizarre correlations, implausible effect sizes — RexDouglass · 2026-10-12
- Matrices are graphs: the most undervalued fact in linear algebra — TivadarDanka · 2026-10-12
- MetaKernelBench measures cross-DSL GPU kernel skill transfer, with lifts from -19% to +29% — encodelab · 2026-10-12