TruthFlow: Truthful LLM Generation via Representation Flow Correction
This addresses the issue of inconsistent truthfulness in LLM outputs for users relying on accurate information, though it is an incremental advancement over existing representation intervention techniques.
The paper tackles the problem of large language models (LLMs) generating untruthful responses by introducing TruthFlow, a method that uses Flow Matching for query-specific representation correction, resulting in significant improvements on TruthfulQA and strong transferability to other hallucination benchmarks.
Large language models (LLMs) are known to struggle with consistently generating truthful responses. While various representation intervention techniques have been proposed, these methods typically apply a universal representation correction vector to all input queries, limiting their effectiveness against diverse queries in practice. In this study, we introduce TruthFlow, a novel method that leverages the Flow Matching technique for query-specific truthful representation correction. Specifically, TruthFlow first uses a flow model to learn query-specific correction vectors that transition representations from hallucinated to truthful states. Then, during inference, the trained flow model generates these correction vectors to enhance the truthfulness of LLM outputs. Experimental results demonstrate that TruthFlow significantly improves performance on open-ended generation tasks across various advanced LLMs evaluated on TruthfulQA. Moreover, the trained TruthFlow model exhibits strong transferability, performing effectively on other unseen hallucination benchmarks.