Self-Reflective Retrieval-Augmented Generation: Learning to Retrieve, Generate, and Critique скачать в хорошем качестве

Self-Reflective Retrieval-Augmented Generation: Learning to Retrieve, Generate, and Critique

Paper: https://arxiv.org/abs/2310.11511 The provided text introduces Self-Reflective Retrieval-Augmented Generation (SELF-RAG), a framework designed to improve the factual accuracy and quality of large language models. Unlike traditional methods that retrieve information indiscriminately, this system adaptively decides when to seek external knowledge and uses special "reflection tokens" to critique its own output. By training a model to evaluate the relevance and support of retrieved passages, the researchers created a more controllable and verifiable generation process. Their findings demonstrate that SELF-RAG significantly outperforms standard models like ChatGPT and Llama2 on complex tasks involving reasoning and long-form writing. Ultimately, this approach balances creativity with factual grounding by allowing the model to self-correct and cite its sources effectively.