# Memory-assisted prompt editing to improve GPT-3 after deployment

**Authors:** Aman Madaan, Niket Tandon, Peter Clark, Yiming Yang  
**arXiv:** [2201.06009](https://arxiv.org/abs/2201.06009)  
**Date:** 2022-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.06009v1/

## Key points

- Large language models (LMs) like GPT-3 are powerful but can make mistakes and be expensive to retrain
- Misinterpreting user instructions is a common failure mode for these models
- Authors propose an approach that allows users to correct errors directly through interaction without requiring retraining
- Approach involves pairing GPT-3 with a growing memory of cases where the model misunderstood the user's intent and was provided with feedback clarifying the instruction
- Memory-enhanced GPT-3 uses feedback from similar prior queries to enrich the prompt when presented with a new query
- Simple proof-of-concept experiments demonstrate how a simulated user can interactively teach a deployed GPT-3, doubling its accuracy on basic lexical tasks such as generating synonyms
- Memory helps avoid repeating similar past mistakes and is a first step towards strengthening deployed models and potentially broadening their utility
- All code and data available at https://github.com/madaan/memprompt for further exploration of this concept.

## Summary

In their paper titled "Memory-assisted prompt editing to improve GPT-3 after deployment," authors Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang discuss the limitations of large language models (LMs) such as GPT-3. While these models are powerful, they are not immune to mistakes and can be prohibitively expensive to retrain. One common failure mode is misinterpreting a user's instruction, which can result in incorrect outputs. The authors propose an approach that allows users to correct errors directly through interaction without requiring retraining. Their approach involves pairing GPT-3 with a growing memory of cases where the model misunderstood the user's intent and was provided with feedback clarifying the instruction. When presented with a new query, their memory-enhanced GPT-3 uses feedback from similar prior queries to enrich the prompt. Through simple proof-of-concept experiments, they demonstrate how a simulated user can interactively teach a deployed GPT-3, doubling its accuracy on basic lexical tasks such as generating synonyms. The authors note that memory helps avoid repeating similar past mistakes and that their simple idea is a first step towards strengthening deployed models and potentially broadening their utility. They make all code and data available at https://github.com/madaan/memprompt for further exploration of this concept. Overall, this paper highlights the potential for improving large language models through interactive feedback mechanisms rather than costly retraining efforts.