TATRA: Training-Free Instance-Adaptive Prompting Through Rephrasing and Aggregation

Researchers introduce TATRA, a training-free prompting method for Large Language Models that creates instance-specific few-shot prompts without requiring labeled training data. The method achieves state-of-the-art performance on mathematical reasoning benchmarks like GSM8K and DeepMath, matching or outperforming existing prompt optimization methods that rely on expensive training processes.

• →TATRA eliminates the need for task-specific training data and expensive optimization loops in prompt engineering.
• →The method constructs instance-adaptive prompts by synthesizing on-the-fly examples for each specific query.
• →TATRA achieves state-of-the-art performance on GSM8K and DeepMath mathematical reasoning benchmarks.
• →Results suggest per-instance prompt construction is more effective than single dataset-level prompt optimization.
• →The approach matches or improves upon strong baselines across standard text classification tasks.