Enhancing Input-Label Mapping in In-Context Learning with Contrastive Decoding
This addresses a specific bottleneck in in-context learning for natural language processing researchers and practitioners, offering an incremental enhancement to existing methods.
The paper tackles the problem of large language models overlooking input-label mapping in in-context learning by introducing In-Context Contrastive Decoding (ICCD), which contrasts output distributions to emphasize this mapping, resulting in up to a 1.8-point average improvement on 7 natural language understanding tasks across 6 LLM scales without extra training.
Large language models (LLMs) excel at a range of tasks through in-context learning (ICL), where only a few task examples guide their predictions. However, prior research highlights that LLMs often overlook input-label mapping information in ICL, relying more on their pre-trained knowledge. To address this issue, we introduce In-Context Contrastive Decoding (ICCD), a novel method that emphasizes input-label mapping by contrasting the output distributions between positive and negative in-context examples. Experiments on 7 natural language understanding (NLU) tasks show that our ICCD method brings consistent and significant improvement (up to +1.8 improvement on average) upon 6 different scales of LLMs without requiring additional training. Our approach is versatile, enhancing performance with various demonstration selection methods, demonstrating its broad applicability and effectiveness. The code and scripts are released at https://github.com/Romainpkq/CD_ICL.