The Unreasonable Effectiveness of Eccentric Automatic Prompts

9 Feb 2024  ·  Rick Battle, Teja Gollapudi ·

Large Language Models (LLMs) have demonstrated remarkable problem-solving and basic mathematics abilities. However, their efficacy is highly contingent on the formulation of the prompt. This study endeavors to quantify the influence of incorporating "positive thinking" into the system message of the prompt, then compare that to systematic prompt optimization. We assess the performance of 60 combinations of system message snippets, tested with and without Chain of Thought prompting, across three models with parameters ranging from 7 to 70 billion on the GSM8K dataset. Our findings reveal that results do not universally generalize across models. In most instances, the inclusion of "positive thinking" prompts positively affected model performance. Notably, however, Llama2-70B exhibited an exception when not utilizing Chain of Thought, as the optimal system message was found to be none at all. Given the combinatorial complexity, and thus computation time, of experimenting with hand-tuning prompts for large black-box models, we then compared the performance of the best "positive thinking" prompt against the output of systematic prompt optimization. We show that employing an automated prompt optimizer emerges as the most effective method for enhancing performance, even when working with smaller open-source models. Additionally, our findings reveal that the highest-scoring, automatically-optimized prompt exhibits a degree of peculiarity far beyond expectations.

PDF Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Arithmetic Reasoning GSM8K Llama-2 70B (on 100 first questions, 4-shot, auto-optimized prompting) Accuracy 61 # 104
Parameters (Billion) 70 # 86
Arithmetic Reasoning GSM8K Llama-2 13B (on 100 first questions, 4-shot, auto-optimized prompting) Accuracy 43 # 125
Parameters (Billion) 13 # 53
Arithmetic Reasoning GSM8K Mistral 7B (on 100 first questions, 4-shot, auto-optimized prompting) Accuracy 41 # 127
Parameters (Billion) 7 # 10

Methods


No methods listed for this paper. Add relevant methods here