Summary of Mitigating Llm Hallucinations Via Conformal Abstention, by Yasin Abbasi Yadkori et al.
Mitigating LLM Hallucinations via Conformal Abstention
by Yasin Abbasi Yadkori, Ilja Kuzborskij, David Stutz, András György, Adam Fisch, Arnaud Doucet, Iuliya Beloshapka, Wei-Hung Weng, Yao-Yuan Yang, Csaba Szepesvári, Ali Taylan Cemgil, Nenad Tomasev
First submitted to arxiv on: 4 Apr 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A novel principled procedure is proposed to determine when a large language model (LLM) should abstain from responding in general domains. This approach builds upon earlier methods that utilize self-consistency as a reliable measure of model confidence. The LLM is employed to self-evaluate the similarity between its sampled responses for a given query, and conformal prediction techniques are leveraged to develop an abstention procedure with rigorous theoretical guarantees on hallucination rates (error rates). Experimental results demonstrate that the resulting conformal abstention method effectively bounds hallucination rates on various closed-book, open-domain generative question answering datasets while maintaining a less conservative abstention rate compared to baselines. This approach achieves comparable performance on a dataset with short answers (TriviaQA) and outperforms baselines on a dataset with long responses (Temporal Sequences). The procedure is evaluated automatically by determining if two responses are equivalent given a question, using a thresholded similarity function. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary A new way to help large language models decide when they don’t know the answer is introduced. This approach uses the model itself to check how similar its different answers are for the same question. The results show that this method can effectively limit the number of incorrect or nonsense answers on various datasets, while still giving accurate responses most of the time. |
Keywords
» Artificial intelligence » Hallucination » Large language model » Question answering
