Thinking in Video: Can Video Generators Really Reason About the Real World?
This repository contains the official implementation of **Causal-Generative Dual-Judge (CGDJ)** for auditing world-model consistency of video generative models β the official codebase of the **Thinking in Video** paradigm.
## π Overview
**Thinking in Video** is a reasoning paradigm in which a video generative model is used not merely to synthesize pixels, but to **simulate, predict, and verify causal thought** over time. Convincing rollouts, however, do not necessarily imply causal understanding β they may reflect memorized visual appearance. Existing metrics also separate perceptual fidelity from semantic logic, leaving the reasoning ability of video generators unverified.
To close this gap, we propose the **Causal-Generative Dual-Judge (CGDJ)** framework, which audits **World Model Consistency** from two complementary perspectives:
- **Explicit Causal Perception** β does the generator *read* a video scenario as a reasoning problem? Tested via spatio-temporal flattened visual question answering.
- **Implicit Generative Prediction** β does the generator *render* the causal consequence as a consistent future video? Tested via reference-based generative evaluation.
The framework is built around two key ideas:
- **Flatten Temporal Video**: a static-image protocol that packs uniformly sampled frames into a single composite image, allowing video generative models to be probed for causal perception with image-only inputs.
- **Dual-track judging**: an audio+text aligned MLLM judge (Gemini-3-Pro) for explicit causal answers, and a reference-based video quality auditor (also Gemini-3-Pro) for implicit generative consistency.
## π Paper
**Title:** `Thinking in Video: Can Video Generators Really Reason About the Real World?`
### π Abstract
Recent advances in world models and video generation have given rise to an emerging reasoning paradigm that leverages video generative models to simulate, predict, and reason about real-world dynamics. We redefine this paradigm as **Thinking in Video**, where video is not merely an output artifact but a medium for constructing, extending, and verifying causal thought. However, this promise remains unverified: convincing rollouts may reflect memorized appearances rather than causal understanding, while existing metrics separate perceptual fidelity from semantic logic. To evaluate whether video generators support such reasoning, we introduce the **Causal-Generative Dual-Judge (CGDJ)**, auditing World Model Consistency from two perspectives. **Explicit Causal Perception** tests whether a generator reads a video scenario as a reasoning problem through spatio-temporal flattened visual question answering, while **Implicit Generative Perception-Prediction Gap** evaluates whether it renders the causal consequence as a consistent future video. Applying CGDJ to representative open- and closed-source generators reveals a clear **Perception-Prediction Gap**: open-source models produce plausible dynamics despite near-zero explicit causal perception, whereas advanced closed-source systems show stronger but still limited alignment between reasoning and generation. Further analysis exposes **audio-visual misalignment**, where models verbalize correct causal logic more reliably than they render it, challenging the "world simulator" narrative.
## π§ Method
CGDJ is composed of three tightly coupled pieces:
- **CausalβGenerative Dual-Judge Benchmark** β a 1,500-video benchmark (900 from Video-MME + 600 with paired input/gold future videos) covering both explicit and implicit evaluation.
- **Flatten Temporal Video** β converts a video (and an optional rasterized query) into a single composite image so that image-only or text-to-image-conditioned video generators can be probed on causal understanding.
- **CausalβGenerative Dual-Judge Evaluation** β an automated Gemini-3-Pro-based pipeline that scores both the explicit causal answer and the implicit generative rollout.
### 1. πΌοΈ Flatten Temporal Video
Because most current video generative models only accept static image inputs, we compress each video into a single composite image `I_composite β β^(1280Γ720Γ3)`:
- **Temporal Grid Construction** β uniformly sample `N = 70` frames and tile them into a `7 Γ 10` spatial matrix.
- **Semantic Rendering** β rasterize the textual query into a dense pixel-space strip and place it in the upper visual field.
- **Vertical Integration** β concatenate the semantic strip and the temporal grid, then resize to `1280 Γ 720`.
For implicit generative evaluation, a lighter **motion-anchor variant** extracts `N = 7` keyframes from the input video and concatenates them horizontally into a "visual arrow of time" `I_motion` that encodes the directional motion prior.
### 2. π§ͺ Explicit Causal Perception
For each Video-MME video, a multimodal judge receives:
1. The generator's video output,
2. Whisper-large-v2 transcription of the audio track,
3. The original question, options, and ground-truth answer,
and classifies the response as **Correct / Incorrect** for spatio-temporal causal reasoning.
### 3. π¬ Implicit Generative Prediction
For each of the 600 causal videos (300 Natural Sciences + 300 Sociology & Humanities), the generator is asked to roll out the post-event future, and a Gemini-3-Pro **Video Quality Auditor** scores the result in `[0.0, 1.0]` along three axes:
- **Semantic Alignment** β object / action correctness
- **Reference Consistency** β logical flow vs. gold video
- **Physical Validity** β absence of flickering or physics violations
## π Experimental Results
We audit both open- and closed-source generators on the CGDJ benchmark. The headline finding is a clear **Perception-Prediction Gap**:
- **Open-source models** (Wan-2.2-14B, HunyuanVideo-1.5) collapse under flattened causal reasoning (near-zero explicit perception) yet still produce moderately plausible causal continuations.
- **Closed-source models** (Sora-2, Veo-3.1) show stronger but still limited alignment between causal reasoning and video rollout.
- **Audio-visual misalignment** β models articulate correct causal logic more reliably than they render it, challenging the "world simulator" narrative.
## π Repository Structure
```text
Thinking-in-Video/
βββ Perception/ # Explicit Causal Perception track
β βββ perception_data.jsonl # Input data (imageID/image_path/question/options/answer)
β βββ 01_generate_video_veo3.py # Stage 1: Veo 3.1 image-to-video
β βββ 02_transcribe_audio_whisper.py # Stage 2: Whisper-large-v2 audio transcription
β βββ 03_judge_answer_gemini3.py # Stage 3: Gemini-3-Pro answer judging
β βββ outputs/ # Runtime artifacts
β βββ README.md
βββ Prediction/ # Implicit Generative Prediction track
β βββ prediction_data.jsonl # Input data (data_id/domain/sub_category/image_path/input_video_path/output_video_path)
β βββ 01_generate_video_veo3.py # Stage 1: Veo 3.1 image-to-video (predict the second half)
β βββ 02_judge_video_gemini3.py # Stage 2: Gemini-3-Pro video quality scoring
β βββ outputs/ # Runtime artifacts
β βββ README.md
βββ figures/ # Figures, icons, framework/teaser images
βββ README.md
```
## π Getting Started
### 1. π Install the Environment
```bash
conda create -n thinkvideo python=3.11 -y
conda activate thinkvideo
pip install -r requirements.txt
```
The judging pipeline uses **Gemini-3-Pro** as the multimodal judge and **Whisper-large-v2** for audio transcription. Configure the corresponding API credentials in the scripts before running.
## ποΈ Running the CGDJ Evaluation
The judging pipeline is split into two tracks that mirror the two CGDJ modules: `Perception/` for Explicit Causal Perception and `Prediction/` for Implicit Generative Prediction. Each track's scripts follow the same `NN__
