LLM research suggests input diversity boosts accuracy more than output diversity

A new research paper explores Test-Time Augmentation (TTA) for Large Language Models (LLMs), proposing that diversifying input data is more compute-efficient for accuracy gains than diversifying output reasoning paths. The study found that semantic rephrasing, a form of TTA, outperformed standard self-consistency methods by approximately 1.8 times in accuracy per dollar across various tasks. This approach is particularly effective for mid-tier models where acquiring a more powerful model is not feasible or cost-effective. AI

IMPACT This research suggests a more cost-effective method for improving LLM accuracy, potentially influencing how developers optimize inference budgets.

RANK_REASON The cluster contains an academic paper detailing a novel method for improving LLM performance. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• Large Language Models
• Lexical perturbations
• Self Consistency In Llms
• Semantic rephrasing
• Visual transformations underlying apparent movement

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →