How Many Examples, and Which Ones? A Study of In-Context Learning for Natural Language Inference

원문 초록

Natural language inference (NLI) is a fundamental task of natural language processing and crucial for evaluating the strengths and weaknesses of AI methods. Research has mostly focused on encoder or encoder-decoder architectures. However, the field of decoder-only Large Language Models (LLMs) remains underexplored. While prior work has mainly investigated complex sampling and prompting techniques, the effectiveness of a minimalistic prompting approach is yet to be examined in decoder-only LLMs. In this work, we examine and provide a comprehensive benchmark of LLMs (Llama, Mistral, Gemma, and Phi 4) in this domain, across the MNLI, QNLI and SciTail datasets. The performance of these models was evaluated using zero-shot and few-shot prompting strategies with simple and direct prompts. The results of the article indicate that LLMs generally perform exceptionally well when they are provided with a simple, straightforward prompt, challenging the notion that complex prompting techniques are always required for NLI tasks. Additionally, we find that the choice of demonstration label in one-shot prompting significantly affects performance for most models, with entailment-primed examples performing best. Our findings also show that the use of a few representative examples, in the case of few-shot, produced performance comparable or superior to prompts with higher example counts, although the efficacy of the prompting techniques is model and dataset-dependent. Moreover, our simple prompting approach consistently matched or outperformed more elaborate methods such as ICS and MCQ-based prompting.