Constructing Benchmarks and Interventions for Combating Hallucinations in LLMs
This work addresses hallucinations in LLMs, which is a critical issue for AI reliability, but it is incremental as it builds on existing intervention methods with new benchmarks and analysis.
The authors tackled the problem of hallucinations in large language models by introducing a method to categorize examples based on prior knowledge and constructing benchmarks for interventions in open-book and closed-book question answering, finding that intervention success varies by component and that dynamic interventions are more robust than static ones.
Large language models (LLMs) are prone to hallucinations, which sparked a widespread effort to detect and prevent them. Recent work attempts to mitigate hallucinations by intervening in the model's generation, typically computing representative vectors of hallucinations vs. grounded generations, for steering the model's hidden states away from a hallucinatory state. However, common studies employ different setups and do not properly separate different possible causes of hallucinations, making interventions misguided. In this work, we introduce a method for categorizing examples based on the model's prior knowledge, named WACK. We construct WACK benchmarks that support interventions in two settings: open-book and closed-book question answering. Using the benchmarks, we perform an extensive investigation of the effect of different choices for intervention, such as the intervened components, and how often and how strongly to intervene. We find that intervention success varies depending on the component, with the attention blocks performing well and the residual stream proving detrimental to language modeling capabilities. We also show that interventions can benefit from representative vectors collected before, rather than after, a hallucination occurs. Finally, we introduce a new dynamic intervention, which intervenes only if needed, and thus is more robust than standard static interventions. The code is available at https://github.com/technion-cs-nlp/hallucination-mitigation .