TANDREWYANG/optimism_train

Optimism Bias World Model Benchmark — Training Data

Training data for VLM-as-judge models evaluating world model predictions.

Structure

├── part1_physconscore/      # Physical Consistency Scoring
├── part2_action_following/  # Action Following Evaluation  
└── part3_optimism_bias/     # Optimism Bias Detection

Part 1: Physical Consistency Score (PhysConsScore)

• Task: Binary classification (Good/Bad) on 16 physics indicators
• Model: Qwen3-VL-8B-Instruct (LoRA fine-tuning)
• Input: 8-frame storyboard image (1280×1440)
• Data: 335 train / 51 val / 34 test
• Source: Human annotations on DreamDojo/Wan2.1 generated videos

Part 2: Action Following

• Task: 3 binary metrics (TCR/OPS/MQ) → OS score (0-100)
• Model: Qwen3-VL-8B / InternVL3-78B (LoRA fine-tuning)
• Input: Side-by-side video (predicted | ground truth)
• Data: 109 train / 42 val / 50 test
• Source: Human annotations + GPT-4o labels on GR1 robot episodes

Part 3: Optimism Bias Detection

• Task: Binary classification (Y=optimism bias / N=no bias)
• Model: Qwen3-VL-8B-Instruct / InternVL3-78B
• Input: 2-column video (baseline | perturbed), 32 frames
• Data: 420 annotations (286 Y+N usable), 460 videos
• Best AUC: 0.598 (InternVL3-78B + perturbation hints)
• Source: Human annotations on DreamDojo 14B/2B, HappyHorse, Wan2.1

Models

Part	Model	Best Result
Part 1	Qwen3-VL-8B	Per-indicator binary accuracy
Part 2	Qwen3-VL-8B / InternVL3-78B	OS Spearman ρ + Pairwise Acc
Part 3	InternVL3-78B (16 frames)	AUC=0.598

Citation

@misc{optimism_bias_benchmark,
  title={Optimism Bias in Robot Manipulation World Models},
  year={2026}
}