Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs

Researchers have proposed a simulation-based synthesis framework called Synapse, aimed at addressing the issues of knowledge insertion and updating in large language models (LLMs). This study introduced the ParallelEvents benchmark dataset, which contains fictional but realistic future worlds and coherent event trajectories to support controlled evaluation and avoid knowledge contamination. The Synapse training framework developed on this basis uses model-generated data to update model parameters through mid-training and instruction fine-tuning. Experimental results show that this method achieves scalable knowledge integration without the need for expensive manual data, with performance improved by 14.23% compared to existing methods, proving that simulation-based synthesis training can achieve robust and consistent knowledge insertion.