Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability
This addresses the problem of costly and hard-to-detect hallucinations in large language models for AI researchers and practitioners, though it is incremental in exploring scale effects.
The study investigated how scaling language models (LMs) affects hallucinations, focusing on cases where correct answers are in the training data, and found that larger, longer-trained LMs hallucinate less but require significantly more compute to reduce hallucinations to ≤5%. It also revealed that hallucination detectors improve with size but become less effective as LMs scale up.
While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations.