Research map / Large language models / Reasoning and chain of thought
ReaForest: Fostering Generative Video Reasoning for Spatial Planning
Kun Ouyang, Yuanxin Liu, Xinhao Li, Linli Yao, Xiangyu Zeng, Haoning Wu, Hao Zhou, Fandong Meng, Jie Zhou, Xu Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ReaForest, a method to foster generative video reasoning for spatial planning in video generation models, motivated by the gap between verbal CoT in LLMs and visual mental simulation in VGMs.
Problem
Video generation models struggle with visual reasoning tasks such as spatial planning, unlike LLMs with strong textual chain-of-thought reasoning.
Approach
Frames video generation as a generative visual reasoning process for spatial planning; the abstract only identifies two gaps (VGMs are pre-trained without reasoning-oriented supervision, as far as the visible text states) and the remainder is unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reasoning activation fine-tuning on ReaGen-27k (ablation) | ReaForest · VRBench | EM: 0.0 -> 88.4 | Improved |
| add text-image alignment to the training data (ablation) | ReaForest · VRBench | EM: 85.2 -> 88.4 | Improved |
| add ReaTTS test-time scaling (ablation) | ReaForest · VRBench | EM: 70.2 -> 88.4 | Improved |
| add entropy-aware budget reallocation (ablation) | ReaForest · VRBench | EM: 82.3 -> 88.4 | Improved |
| add temperature scaling in the softmax allocation (ablation) | ReaForest · VRBench | EM: 84.1 -> 88.4 | Improved |
| add temperature adaptation (ablation) | ReaForest · VRBench | EM: 85.3 -> 88.4 | Improved |
| add forest entropy modulation of temperature (ablation) | ReaForest · VRBench | EM: 86.0 -> 88.4 | Improved |
| add reflective rectification (ablation) | ReaForest · VRBench | EM: 81.5 -> 88.4 | Improved |
| use hierarchical verification instead of random verification (ablation) | ReaForest · VRBench | EM: 61.6 -> 88.4 | Improved |
| add hard constraints in verification (ablation) | ReaForest · VRBench | EM: 83.7 -> 88.4 | Improved |
| add soft constraints in verification (ablation) | ReaForest · VRBench | EM: 85.1 -> 88.4 | Improved |
| add terminal constraints in verification (ablation) | ReaForest · VRBench | EM: 85.7 -> 88.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Wan-R1 — Wan-R1: Verifiable-Reinforcement Learning for Generalizable Video Reasoning (ECCV 2026) · 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
- Visual Planning: Let's Think Only with Images (ICLR 2026 · similarity 0.90)
- CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models (NeurIPS 2026 · similarity 0.89)
- Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.89)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators (NeurIPS 2026 · similarity 0.88)
- MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.88)
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.
