Recent research reveals that large language models (LLMs) can effectively simulate environments, addressing the training bottleneck faced by autonomous AI agents. These agents require extensive real-world experience to learn, which is often limited. Researchers from various institutions tested whether LLMs could serve as internal world models, simulating the outcomes of actions taken by agents. This study involved five text-based environments, including ALFWorld and SciWorld, to evaluate the accuracy and consistency of LLM predictions in simulating actions. Fine-tuning pre-trained models like Qwen2.5-7B and Llama3.1-8B achieved over 99% accuracy in specific structured tasks. Scalability depended on both data volume and model size, revealing that larger models perform better in complex environments. The findings support a shift towards experience-based AI training, aligning with calls for continuous learning in AI systems. This innovative approach could significantly enhance the training of autonomous agents, fostering advancements in AI capabilities.
Source link
Share
Read more