Investigating Hallucination Cascades in Autoregressive Large Language Models

Investigating Hallucination Cascades in Autoregressive Large Language Models

Computer Science

Luc Mekouar

2024

Prof Bo Chen

As the adoption of Large Language Models (LLMs) is becoming ubiquitous, understanding their pitfalls is becoming increasingly urgent. LLMs are known to hallucinate and may be subject to cascading hallucinatory generations due to their autoregressive structure and exposure bias. This paper first introduces the topic of hallucination cascade, covering relevant literature. Then an experiment is presented on the propagation of hallucination through long answer text generation, for all combinations of five LLMs - Qwen2 (0.5B), Qwen2 (7B), Llama3.1 (8B), Gemma2 (9B) and Mistral-NeMo (12B) - at five levels of temperature: 0, 0.75, 1, 1.25, 2. In a novel methodology, responses are generated by the five LLMs, then labelled by two more LLMs as hallucinatory or not on both a token and sentence level. The labelling is done with clear instructions and additional context to ensure its quality. Correlational evidence was found on both granularity levels that hallucinatory tokens are positively autocorrelated at lag 1, and that autocorrelation monotonically decreases as lag increases (token level only), supporting the cascading hallucination hypothesis. Finally, some hallucination mitigating techniques are presented, with both research and application in mind. All the code used for the empirical part of this paper (Section 3) can be found at https://github.com/LucMekouar/URSS_2024.

Large Language Model, LLM, Machine Learning, Natural Language Processing, Hallucination