This repository includes code implementations for two research papers that explore hallucination detection and mitigation in large language models (LLMs), focusing on different aspects of the problem. "Constructing Benchmarks and Interventions for Combating Hallucinations in LLMs" and "Distinguishing Ignorance from Error in LLM Hallucinations"
The first paper contain an overview of intervention for Hallucination Mitigation: develop WACK benchmarks to test different intervention methods, focusing on factors like the components being adjusted, and the frequency and intensity of intervention. Revealing that attention-based interventions outperform the residual stream. And introduce Dynamic interventions to apply mitigation only when necessary, offering a more robust approach.
The second paper addresses two types of hallucinations in LLMs during closed-book question answering (CBQA): (1) when the model lacks the correct answer in its parameters, and (2) when the model answers incorrectly despite possessing the necessary knowledge. We argue that distinguishing these cases is essential, as each requires different mitigation strategies. To assist in differentiating these cases, we introduce the WACK dataset (Wrong Answer despite having Correct Knowledge), which is model-specific. Probing experiments reveal distinct internal representations for each hallucination type, with WACK datasets showing variation in the model-specific-datasets between models. This model-specific approach enhances detection of case (2) hallucinations compared to using generic datasets.