Research map / Video / Video-language understanding
ReMoT: Reinforcement Learning with Motion Contrast Triplets
Cong Wan, Zeyu Guo, Jiangyang Li, Songlin Dong, Yifan Bai, Lin Peng, Zhiheng Ma, Yihong Gong
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 alternating SFT and GRPO training instead of pure GRPO (ablation) | Qwen3-VL-4B-Thinking · ReMoT-16k-Test (avg, overall) | overall accuracy (%): 33.6 -> 38.0 | Improved |
| add GRPO training on motion-contrast data to the base model (ablation) | Qwen3-VL-4B-CoT · ReMoT-16k-Test (avg, overall) | overall accuracy (%): 20.7 -> 33.6 | Improved |
| add alternating SFT instead of sequential SFT then GRPO (ablation) | Qwen3-VL-4B-Thinking · ReMoT-16k-Test Camera (navigation, overall) | overall accuracy (%): 27.0 -> 21.4 | Hurt |
| add navigation data to manipulation-only training (ablation) | ReMoT training · ReMoT-16k-Test (overall) | overall accuracy (%): 23.9 -> 32.4 | Improved |
| add simulation data (ablation) | ReMoT training · ReMoT-16k-Test (overall) | overall accuracy (%): 32.4 -> 38.0 | Improved |
| add logic consistency reward to GRPO (ablation) | Qwen3-VL-4B-CoT · Manipulation subset (logic test) | overall accuracy (%): 68.6 -> 78.0 | Improved |
| add logic consistency reward to GRPO (ablation) | Qwen3-VL-4B-CoT · Manipulation subset (logic test) | logic consistency (%): 98.6 -> 99.3 | Improved |
| build training data with multi-expert rule-based pipeline instead of VLM-generated data (ablation) | ReMoT data construction · ReMoT-16k-Test scaling | accuracy: 0.49 -> 0.66 | Improved |
| apply ReMoT training to Qwen3-VL-4B on spatial-temporal benchmarks | Qwen3-VL-4B-CoT · VLM2 | accuracy (%): 66.4 -> 70.0 | Improved |
| apply ReMoT training to Qwen3-VL-4B on general benchmarks | Qwen3-VL-4B-CoT · MUIR | accuracy (%): 73.8 -> 72.8 | Hurt |
| apply ReMoT training to Qwen3-VL-4B on general benchmarks | Qwen3-VL-4B-CoT · MMMU | accuracy (%): 70.8 -> 71.4 | No clear effect |
Closest papers by meaning
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.88)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.87)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- PeRL: Permutation-Enhanced Reinforcement Learning for Interleaved Vision-Language Reasoning (NeurIPS 2025 · similarity 0.86)
- METRA: Scalable Unsupervised RL with Metric-Aware Abstraction (ICLR 2024 · similarity 0.86)
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.
