Research map / Video / Video-language understanding
Incentivizing Vision Language Models to Search for Long Video Question Answering
Harsh Goel, S P Sharan, Sahil Shah, Minkyu Choi, Joungbin An, Kristen Grauman, Sandeep Chinchali
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 VETL temporal-logic reward to exact-match reward (ablation) | VSeek (GRPO, Qwen3-VL-4B-Thinking) · MLVU | Pass@1: 66.7 -> 67.9 | Improved |
| add VETL reward to exact-match reward (ablation) | VSeek · Video-MME | Pass@1: 59.3 -> 61.4 | Improved |
| add VETL reward to exact-match reward (ablation) | VSeek · LongVideoBench | Pass@1: 60.5 -> 60.9 | Improved |
| add VETL reward to exact-match reward (ablation) | VSeek · LVBench and CGBench-mini (OOD) | Pass@1: 40.0 -> 38.8; 33.2 -> 32.3 | Hurt |
| post-train with search and RL (VSeek-VETL) vs direct base inference | Qwen3-VL-4B-Thinking · LongVideoBench | Pass@1: 51.7 -> 60.9 | Improved |
| use temporal-logic decomposition instead of direct decomposition for rewards (ablation) | VSeek-VETL · MLVU | Pass@1: 66.3 -> 67.9 | Improved |
| use temporal-logic decomposition instead of direct decomposition (ablation) | VSeek-VETL · Video-MME | Pass@1: 59.3 -> 61.4 | Improved |
| use VETL specification instead of direct decomposition (judge score) (ablation) | VSeek reward pipeline · GPT-5.2 judge mean score (0-6) | score: 3.8 -> 5.3 | Improved |
| add VETL reward to RLOO training (ablation) | VSeek with RLOO · LongVideoBench | Pass@1: 58.0 -> 60.4 | Improved |
| add VETL reward to GSPO training (ablation) | VSeek with GSPO · MLVU | Pass@1: 67.7 -> 62.7 | Hurt |
| add VETL reward to GSPO training (ablation) | VSeek with GSPO · Video-MME | Pass@1: 62.9 -> 58.1 | Hurt |
| add VETL reward on Instruct base model (ablation) | VSeek-Ins · MLVU | Pass@1: 60.3 -> 66.1 | Improved |
Closest papers by meaning
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.92)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.91)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.91)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.91)
- Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA (NeurIPS 2026 · similarity 0.91)
- Koala: Key Frame-Conditioned Long Video-LLM (CVPR 2024 · similarity 0.90)
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.
