Research map / Video / Video-language understanding
SAGE: Training Smart Any-Horizon Agents for Long Video Reasoning with Reinforcement Learning
Jitesh Jain, Allen AI blank, Jialuo Li, Zixian Ma, Jieyu Zhang, Chris Dongjoo Kim, Sangho Lee, Rohun Tripathi, Tanmay Gupta, Christopher Clark, Humphrey Shi
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 RL after SFT (ablation) | SAGE (Qwen3-VL-4B) · SAGE-Bench | overall accuracy (%): 64.6 -> 68.4 | Improved |
| add RL after SFT (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 63.9 -> 68.0 | Improved |
| add the ground-event tool (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 67.3 -> 68.0 | Improved |
| add the web-search/parse-website tool (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 65.5 -> 68.0 | Improved |
| add the analyze tool (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 63.4 -> 68.0 | Improved |
| add the extract-video-parts tool (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 63.0 -> 68.0 | Improved |
| add the transcribe-speech tool (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 62.5 -> 68.0 | Improved |
| give access to the video (ablation) | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 41.0 -> 68.0 | Improved |
| use agentic mode instead of direct mode with SFT+RL (ablation) | SAGE (Qwen3-VL-4B) · SAGE-Bench | overall accuracy (%): 66.3 -> 68.4 | Improved |
| wrap Gemini-2.5-Flash in the SAGE-MM system | Gemini-2.5-Flash · SAGE-Bench | overall accuracy (%): 68.1 -> 71.3 | Improved |
| wrap GPT-4o in the SAGE-MM system | GPT-4o · SAGE-Bench | overall accuracy (%): 71.6 -> 73.4 | Improved |
| SFT only on Qwen3-VL-8B vs base model | SAGE (Qwen3-VL-8B) · SAGE-Bench | overall accuracy (%): 64.9 -> 63.9 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- VideoMind — VideoMind: A Chain-of-LoRA Agent for Temporal-Grounded Video Reasoning (ICLR 2026) · better on 1 of 1 shared result
- LVAgent — LVAgent: Long Video Understanding by Multi-Round Dynamical Collaboration of MLLM Agents (ICCV 2025) · better on 1 of 1 shared result
- VideoAgent — VideoAgent: A Memory-augmented Multimodal Agent for Video Understanding (ECCV 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.88)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.87)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.87)
- Scaling RL to Long Videos (NeurIPS 2025 · similarity 0.87)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.87)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · 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.
