GrIPS: Gradient-free, Edit-based Instruction Search for Prompting Large Language Models

Introduction

• What is the name of the GrIPS paper?

  **GrIPS: Gradient-free, Edit-based Instruction Search for Prompting Large Language Models**

  Archiki Prasad Peter Hase Xiang Zhou Mohit Bansal, UNC Chapel Hill

• What are the main contributions of the GrIPS paper?
  • Gradient Free method to improve performance of GPT models on natural instructions
  • Improved Performance for instruct GPT and manual rewriting
  • Can improve instructions with as few as 20 data points
  • Show that models can benefit from semantically incoherent instructions
• One of the main motivations for gradient free optimization of prompts for generative models is models are behind openai paywall

Method

• GrIPS paper follows the prompt format from the Natural instruction paper
  • Format is Instructions or Instructions + Examples
  • note: I contributed to the Natural Instructions v2 paper
• GrIPS optimizes instructional prompt by editing instructions and greedily selecting best edit
  • search guided by performance on a small score set (100 examples)
  • generate mxl new candidates at each step

  

• GrIPS scores is a combination of balanced accuracy and entropy of class predictions
  • entropy term is to encourage diverse label output

• GrIPS instructions are broken into phrases using a SOTA CRF based constituency parser

• GrIPS generates new prompting candidates by applying 4 primary edit operations
  • delete
  • swap
  • paraphrase (use pegasus model)
  • addition: add back in a previously deleted phrase

  This reminds me of the operations performs in the Monte Carlo search in Bayesian Rule Lists

Results

• GrIPS on average gives several points of performance increase over GPT models

  

• GrIPS generates better performing instructions even if they are semantically incoherent

  

Conclusions

Obviously would be interesting to compare to actual finetuning on a similarly sized training set (ie finetune on a number of examples equal to the size of the scoring set. However this paper shows a promising gradient free approach to optimizing generative models.

Reference

@article{Prasad2022GrIPS,
  title         = {GrIPS: Gradient-free, Edit-based Instruction Search for Prompting Large Language Models},
  author        = {Archiki Prasad and Peter Hase and Xiang Zhou and Mohit Bansal},
  year          = {2022},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  eprint        = {2203.07281}
}