Researchers have investigated the optimal strategy for in-context learning (ICL) within fixed token budgets, focusing on the trade-off between the number and length of examples. Their experiments on question-answering benchmarks like "Beat the AI" and "Natural Questions," using six instruction-tuned large language models, indicate that using more, shorter examples is generally more effective than fewer, longer ones. This finding was particularly pronounced on the "Beat the AI" dataset and also observed, though to a lesser extent, on NQ. An additional experiment on natural language inference using ANLI R3 also suggested that shorter examples often yield better results, highlighting demonstration granularity as a key factor in prompt design. AI
IMPACT Optimizing prompt construction for in-context learning can reduce inference costs and latency for LLM applications.
RANK_REASON Academic paper detailing research findings on LLM in-context learning. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →