Research map / Video / Video-language understanding
Video-in-the-Loop: Span-Grounded Long Video QA with Interleaved Reasoning
Chendong Wang, Donglin Bai, Yifan Yang, Xiao Jin, Anlan Zhang, Rui Wang, Shiqi Jiang, Yuqing Yang, Hao Wu, Qi Dai, Chong Luo, Ting Cao and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add frame-level textual timestamp injection (ablation) | single-stage SFT (Qwen2.5-VL) · LVBench | accuracy (%): 40.2 -> 44.3 | Improved |
| add frame-level textual timestamp injection (ablation) | single-stage SFT (Qwen2.5-VL) · LongVideoBench | accuracy (%): 61.3 -> 62.1 | Improved |
| use Stage-2 answering prompt on the full video (ablation) | SFT + TI · LVBench | accuracy (%): 44.3 -> 45.4 | Improved |
| add learned localization and interleaved GRPO with coupled QA+TVG rewards (ablation) | ViTL · LVBench | accuracy (%): 45.4 -> 47.9 | Improved |
| add learned localization and interleaved GRPO (ablation) | ViTL · LongVideoBench | accuracy (%): 62.5 -> 63.5 | Improved |
| replace CLIP top-k frame selection with learned parametric selection (ablation) | ViTL · LongVideoBench | accuracy (%): 58.1 -> 63.3 | Improved |
| use predicted spans instead of random zoom (ablation) | ViTL · LongVideoBench | accuracy (%): 55.2 -> 63.3 | Improved |
| use ground-truth spans instead of predicted spans (oracle) (ablation) | ViTL · LongVideoBench | accuracy (%): 63.3 -> 70.5 | Improved |
| zoom to high fidelity instead of reusing low-fps frames (Refine-only) (ablation) | ViTL · LongVideoBench | accuracy (%): 56.2 -> 63.3 | Improved |
| add iterative reasoning without zoom (Refine-only) (ablation) | single-stage Qwen2.5-VL · LongVideoBench | accuracy (%): 56.0 -> 56.2 | No clear effect |
| allow multiple disjoint spans instead of a single span (ablation) | ViTL · LVBench | accuracy (%): 45.1 -> 47.4 | Improved |
| add two-stage skim-zoom pipeline | ViTL · latency, 128 frames | latency (s): 6.24 -> 7.32 | Hurt |
Closest papers by meaning
- VITED: Video Temporal Evidence Distillation (CVPR 2025 · similarity 0.89)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 2024 · similarity 0.89)
- Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA (NeurIPS 2026 · similarity 0.89)
- Modularized Self-Reflected Video Reasoner for Multimodal LLM with Application to Video Question Answering (ICML 2025 · similarity 0.88)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.88)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
