Trusting Your Evidence: Hallucinate Less with Context-aware Decoding
This addresses the issue of hallucinations in language models for users relying on accurate text generation, representing an incremental improvement through a novel decoding method.
The paper tackles the problem of language models generating unfaithful or hallucinated text by introducing context-aware decoding (CAD), which amplifies the difference in output probabilities with and without context, resulting in significant improvements such as a 14.3% gain in factuality metrics for LLaMA in summarization tasks.
Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD), which follows a contrastive output distribution that amplifies the difference between the output probabilities when a model is used with and without context. Our experiments show that CAD, without additional training, significantly improves the faithfulness of different LM families, including OPT, GPT, LLaMA and FLAN-T5 for summarization tasks (e.g., 14.3% gain for LLaMA in factuality metrics). Furthermore, CAD is particularly effective in overriding a model's prior knowledge when it contradicts the provided context, leading to substantial improvements in tasks where resolving the knowledge conflict is essential.