Panini：通过结构化记忆在标记空间中的持续学习

Language models are increasingly applied to reason over content they were not originally trained on, including new documents, evolving knowledge, and user-specific data. A prevalent method is retrieval-augmented generation (RAG), which stores documents externally as chunks and retrieves only relevant subsets at inference time for large language models (LLMs) to reason over. However, this approach leads to inefficient use of test-time compute, as LLMs repeatedly reason over the same documents. Additionally, chunk retrieval can introduce irrelevant context, increasing the risk of unsupported generation. To address these issues, a human-like non-parametric continual learning framework is proposed, where the base model remains fixed and learning occurs by integrating each new experience into an external semantic memory. This framework aims to enhance the model's adaptability to new information and improve reasoning efficiency.