Abstract: Large language models (LLMs) have demonstrated impressive capabilities across diverse downstream tasks. However, their output often includes factual errors (i.e., hallucinations), making it challenging to apply them to real-world systems and potentially resulting in catastrophic failures.
In this talk, I will begin by presenting our recent analysis of the limitations of memorizing factual knowledge in LM parameters (parametric knowledge), which frequently leads to factual inaccuracies. We found that LLMs, even the largest state-of-the-art models, often struggle with long-tail factual knowledge that is less represented on the web, and scaling up may not necessarily resolve hallucinations in long-tail contexts. Incorporating retrieved non-parametric knowledge, often referred to as Retrieval-Augmented Generation (RAG), can significantly mitigate this issue, albeit at the cost of efficiency, versatility, and robustness to irrelevant context. Can we build a reliable yet versatile LLM that effectively leverages both parametric and non-parametric memories?
As a first step, I will discuss our new framework, Self-RAG, which trains any LM to learn to retrieve, generate, and criticize. Self-RAG generates output and includes special reflection tokens to invoke a retriever on demand and criticize its own output. It can also retrieve passages from multiple fine-grained aspects. Generating reflection tokens enables the LM to be controlled during the inference phase, allowing it to adapt its behavior to diverse task requirements. Self-RAG enhances the factuality of generation and provides more reliable citations without compromising the fluency of LLMs, outperforming ChatGPT in five tasks.
I will conclude this talk by discussing our ongoing efforts to further enhance the factuality and verifiability of LLMs. In particular, we are proposing a more fine-grained taxonomy and definition of "hallucinations" and building a system that can detect and suggest edits based on our taxonomy.
Bio: Akari Asai is a 5th year PhD student in the Paul G. Allen School of Computer Science & Engineering at the University of Washington, advised by Prof. Hannaneh Hajishirzi. She develops methods in Natural language processing and Machine learning to develop reliable systems that can help humans access the information they need. Her recent work focuses on retrieval-augmented language models, which incorporate external knowledge at inference time to enhance the capabilities of large language models. Her work has been published in venues including ACL, EMNLP, NeurIPS, and ICLR, and featured in media outlets such as MIT Technology Review. Her work is also recognized by the IBM fellowship, the Nakajima Fellowship, and EECS Rising Stars.