Toward Consistent World Models with Multi-Token Prediction and Latent Semantic Enhancement — ThinkLLM