# RE-Adapt: Reverse Engineered Adaptation of Large Language Models

**Authors:** William Fleshman, Benjamin Van Durme  
**arXiv:** [2405.15007](https://arxiv.org/abs/2405.15007)  
**Date:** 2024-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.15007v1/

## Key points

- RE-Adapt and LoRE-Adapt are innovative approaches for fine-tuning large language models (LLMs) on new domains without compromising pre-existing instruction-tuning.
- Reverse engineering an adapter isolates additional knowledge acquired by an instruction-tuned model, allowing the base model to be fine-tuned on a new domain and readapted to instruction following.
- Experiments conducted on StreamingQA and RetrievalQA datasets show that RE-Adapt and LoRE-Adapt consistently outperform other fine-tuning methods across various LLMs, even when combined with retrieval-augmented generation (RAG).
- Incorporating new knowledge through RE-Adapt significantly enhances question answering performance compared to traditional fine-tuning strategies, including improvements in RAG-based systems under ideal conditions of perfect retrieval.
- Limitations include focusing solely on question answering tasks due to resource constraints and not exploring different prompting strategies that could impact results.

## Summary

, , , , 

In this study, the researchers introduce RE-Adapt, an innovative approach to fine-tuning large language models (LLMs) on new domains without compromising any pre-existing instruction-tuning. They achieve this by reverse engineering an adapter that isolates the additional knowledge acquired by an instruction-tuned model beyond its original pretrained base model, without requiring extra data or training. The base model is then fine-tuned on a new domain and readapted to instruction following using the reverse engineered adapter. The study also introduces a low-rank variant called LoRE-Adapt. To validate their approach, the researchers conduct experiments on StreamingQA and RetrievalQA datasets, utilizing a BM-25 index for passage retrieval as context for the models. They compare their results with those obtained using an oracle retriever to eliminate any biases introduced by imperfect retrieval. The results demonstrate that RE-Adapt and LoRE-Adapt consistently outperform other fine-tuning methods across various LLMs and datasets, even when combined with retrieval-augmented generation (RAG). Furthermore, the study discusses how incorporating new knowledge into existing LLMs through RE-Adapt enhances question answering performance significantly compared to traditional fine-tuning strategies. The researchers also observe improvements in RAG-based systems, even under ideal conditions of perfect retrieval. Additionally, they highlight the potential of recovering additional pretraining knowledge by reducing the strength of instruction-tuning through partial adaptation. The limitations of the study include focusing solely on question answering tasks due to resource constraints and not exploring different prompting strategies that could impact results. However, overall, the findings suggest promising implications for future research in balancing knowledge acquisition and problem-solving capabilities in LLMs. In conclusion, this research contributes a valuable method for enhancing LLM performance in new domains while preserving previous instruction-tuning efforts. By enabling others to leverage existing instruction-tuning through open-source models, the study aims to reduce energy consumption and environmental impacts associated with LLM customization.