LITO: Learnable Intervention for Truthfulness Optimization

arXiv:2405.00301v1 Announce Type: new
Abstract: Large language models (LLMs) can generate long-form and coherent text, but they still frequently hallucinate facts, thus limiting their reliability. To address this issue, inference-time methods that elicit truthful responses have been proposed by shifting LLM representations towards learned "truthful directions". However, applying the truthful directions with the same intensity fails to generalize across different question contexts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored …

abstract arxiv cs.cl facts form generate however inference issue language language models large language large language models llm llms optimization reliability responses text type