# Conflict-Aware Soft Prompting for Retrieval-Augmented Generation

**Authors:** Eunseong Choi, June Park, Hyeri Lee, Jongwuk Lee  
**arXiv:** [2508.15253](https://arxiv.org/abs/2508.15253)  
**Date:** 2025-08-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.15253v1/

## Key points

- Introduction of Conflict-Aware Retrieval-Augmented Generation (CARE) approach
- Integration of context assessor and base LLM to handle context-memory conflicts
- Training of context assessor through grounded/adversarial soft prompting to identify unreliable context
- Mechanism to direct reasoning process towards correct parametric knowledge
- Outperformance of existing methods by CARE in QA and fact-checking benchmarks with 5.0% performance gain
- Effectiveness of CARE in discerning conflicting knowledge and guiding base LLM towards accurate responses shown through t-SNE visualization
- Limitations include focusing on top-1 retrieved passages and single-step decoding, but represents a significant advancement in developing trustworthy and adaptive RAG systems

## Summary

The paper "Conflict-Aware Soft Prompting for Retrieval-Augmented Generation" introduces a novel approach called Conflict-Aware Retrieval-Augmented Generation (CARE) to address the issue of context-memory conflicts in large language models (LLMs) enhanced with retrieval-augmented generation (RAG). The core idea behind CARE is the integration of a context assessor and a base LLM to effectively handle conflicting external knowledge retrieved during the generation process. The context assessor in CARE encodes memory token embeddings from raw context tokens and is trained through grounded/adversarial soft prompting to identify unreliable context and provide guidance signals for more reliable reasoning. This mechanism helps direct the reasoning process towards the correct parametric knowledge, mitigating context-memory conflicts effectively. Extensive experiments conducted on QA and fact-checking benchmarks demonstrate that CARE outperforms existing methods, leading to an average performance gain of 5.0%. The qualitative analysis using t-SNE visualization showcases the effectiveness of CARE in discerning conflicting knowledge and guiding the base LLM towards accurate responses. However, despite its limitations such as focusing solely on top-1 retrieved passages and single-step decoding, CARE represents a significant advancement in developing trustworthy and adaptive RAG systems by addressing context-memory conflicts effectively.