Physics of Language Models: Knowledge Storage, Extraction, and Manipulation

Summary

A CMU professor and MBZUAI affiliated faculty presented research on how LLMs store and use knowledge learned during pre-training. The study used a synthetic biography dataset to show that LLMs may not effectively use memorized knowledge at inference time, even with zero training loss. Data augmentation during pre-training can force the model to store knowledge in specific token embeddings. Why it matters: The research highlights limitations in LLM knowledge manipulation and extraction, with implications for improving model architectures and training strategies for more effective knowledge utilization in Arabic LLMs.