On the Limitations of Large Language Models (LLMs): False Attribution

TL;DR

This paper introduces the Simple Hallucination Index (SHI) to measure false attribution in large language models, evaluates three open models on author prediction tasks, and analyzes their hallucination tendencies and correlations with data frequency.

Contribution

The work presents a new hallucination metric (SHI), empirically evaluates state-of-the-art LLMs on author attribution, and provides insights into their false attribution limitations.

Findings

Mixtral 8x7B achieves highest accuracy and lowest SHI among models.

Mixtral 8x7B exhibits high hallucination levels for some books.

Prediction accuracy correlates with Wikipedia mention frequencies.

Abstract

In this work, we introduce a new hallucination metric - Simple Hallucination Index (SHI) and provide insight into one important limitation of the parametric knowledge of large language models (LLMs), i.e. false attribution. The task of automatic author attribution for relatively small chunks of text is an important NLP task but can be challenging. We empirically evaluate the power of 3 open SotA LLMs in zero-shot setting (Gemma-7B, Mixtral 8x7B, and LLaMA-2-13B). We acquired the top 10 most popular books of a month, according to Project Gutenberg, divided each one into equal chunks of 400 words, and prompted each LLM to predict the author. We then randomly sampled 162 chunks per book for human evaluation, based on the error margin of 7% and a confidence level of 95%. The average results show that Mixtral 8x7B has the highest prediction accuracy, the lowest SHI, and a Pearson's correlation (r) of 0.724, 0.263, and -0.9996, respectively, followed by LLaMA-2-13B and Gemma-7B. However, Mixtral 8x7B suffers from high hallucinations for 3 books, rising as high as a SHI of 0.87 (in the range 0-1, where 1 is the worst). The strong negative correlation of accuracy and SHI, given by r, demonstrates the fidelity of the new hallucination metric, which may generalize to other tasks. We also show that prediction accuracies correlate positively with the frequencies of Wikipedia instances of the book titles instead of the downloads and we perform error analyses of predictions. We publicly release the annotated chunks of data and our codes to aid the reproducibility and evaluation of other models.