Data Party THU
Sep 28, 2026 · Artificial Intelligence
How Causal Inductive Biases Shape Structured World Models: VCD, COMET & SPARTAN
An Oxford PhD thesis introduces three methods—VCD, COMET, and SPARTAN—that embed causal inductive biases into world models, achieving better adaptability, robustness, and interpretability across simulation, reinforcement learning, and language modeling tasks.
COMETReinforcement LearningSPARTAN
0 likes · 19 min read