D-SAT：教AI理解"为什么"而非仅仅"是什么"的因果世界模型

Project Source

• Author/Maintainer: engineer-nithura
• Source Platform: GitHub
• Original Title: D-SAT-Phases-1-3-Data-Pipeline-Causal-Model-Training-Counterfactual-Fine-tuning
• Link: https://github.com/engineer-nithura/D-SAT-Phases-1-3-Data-Pipeline-Causal-Model-Training-Counterfactual-Fine-tuning
• Release/Update Time: 2026-06-01T17:12:09Z

Core Idea

D-SAT (Dynamic Scene-Action Transformer) aims to teach AI to understand 'why' (causal relationships) instead of just 'what' in videos. It builds a causal world model via three phases, using Gemma 3 and LoRA for scene graph-to-scene graph causal reasoning, plus counterfactual training to enhance causal understanding.

Current video understanding models have critical limitations:

• Action recognition models identify action types (e.g., 'cutting') but ignore执行者, objects, and changes.
• Scene graph generators capture static spatial relationships but not temporal evolution.
• Visual-Language Models (VLMs) generate descriptions but lack explicit causal reasoning (can't answer 'what if' questions).

D-SAT's goal is to learn a state transition function: given current scene graph Gₜ and an action, predict next scene graph Gₜ₊₁.

D-SAT has three core components:

1. Perception Module (Frozen)
   • Uses pre-trained DINOv2 ViT backbone + graph generation head to convert video frames into structured JSON scene graphs (no training here).
2. Causal Transition Model (Trainable)
   • Core component: Gemma 3 model fine-tuned with LoRA (parameter-efficient). Inputs current scene graph + action text, outputs predicted next scene graph. Trained with cross-entropy loss.
3. Counterfactual Reasoning Layer
   • Post-basic training: fine-tune on curated counterfactual examples to shift from pattern matching to true causal understanding.

Phase1: Automated Causal Dataset Generation

• Source: YouCook2 dataset (414 videos, 3180 subtitled clips).
• Steps: Load annotations → download video clips (yt-dlp) → extract start/end frames (ffmpeg) → Gemini 2.0 Flash generate Gₜ/Gₜ₊₁ → filter inconsistent triplets → output triplets.jsonl.

Phase2: Causal Model Training

• Base model: Gemma3 (2B instruction version).
• Training: Use peft library for LoRA fine-tuning on A100 GPU, cross-entropy loss for sequence prediction.
• Evaluation: Graph Edit Distance (GED) on holdout set.
• Output: lora_adapter/ (model checkpoints).

This phase tests if the model truly understands causality:

• Load Phase2's best checkpoint.
• Fine-tune on curated counterfactual examples (e.g., same start scene but 'add salt' vs 'add sugar' should yield different results).
• Evaluation: Check both counterfactual accuracy and original GED to avoid performance degradation.
• Output: lora_adapter_cf/ (causal-aware model checkpoints).

Four more phases to complete the end-to-end system:

1. Phase4: Expand dataset (full YouCook2 + other video datasets).
2. Phase5: Full training & comprehensive evaluation with expanded data.
3. Phase6: Connect frozen perception module to causal model for end-to-end video inference.
4. Phase7: Build interactive demo & write final report.

Key Highlights

• Combines LLM reasoning (Gemma3), parameter-efficient fine-tuning (LoRA), and counterfactual training.
• Shifts AI from pattern recognition to causal understanding.

Significance

• Addresses a core gap in AI: moving beyond correlation to causation.
• Paves the way for more reliable, explainable AI systems.
• Raises critical questions: What does it mean for AI to 'understand' the world? (Deep原理 vs surface patterns).