Large language models (LLMs) are increasingly used in computational social science (CSS) for tasks like hate speech detection, analyzing public sentiment, and understanding social trends. But how do we best adapt these powerful AI tools for specific research needs? This post delves into two popular methods: instruction tuning and in-context learning. Instruction tuning is like giving an LLM a crash course. We fine-tune the model on a dataset with specific instructions, essentially teaching it the ropes for a particular task. This method typically improves accuracy and helps the LLM generalize its learning to unseen examples. In contrast, in-context learning is a more agile approach. Instead of explicitly training the model, we provide it with a few examples within the input prompt itself. The LLM then uses these examples as a guide to understand the task and generate the desired output. This method bypasses model updates, which drastically improves task adaptation speed. So, which method reigns supreme in the realm of social science? Recent research suggests that in-context learning often outperforms instruction tuning in few-shot CSS tasks. Why? One reason may be the rich contextual knowledge pre-trained LLMs possess. This vast knowledge base allows them to quickly grasp complex social contexts when provided with relevant examples. Instruction tuning, while effective, can be more resource-intensive and prone to overfitting when limited examples are available. Simply throwing more data at the problem isn’t always the solution either; both methods can see performance dips when quantity trumps quality. Carefully curated examples that represent diverse perspectives and social contexts are more beneficial than a large sample of homogeneous data. Additionally, prompting strategies significantly impact performance. In-context learning, with a few well-chosen examples in the prompt, generally outperforms zero-shot learning (no examples) or chain-of-thought prompting (which involves breaking down reasoning steps). The complexity of chain-of-thought prompting can sometimes hinder performance by turning a simple classification into a more complex language understanding task. The research highlights the dynamic interplay between data quality, example diversity, and prompting strategies in harnessing the power of LLMs for social science research. In-context learning, with its agility and reliance on contextual examples, presents an exciting avenue for future research in computational social science, offering a more rapid and efficient way to apply LLMs to complex social questions.