Abstract
This survey provides an in-depth analysis of knowledge conflicts for large language models (LLMs), highlighting the complex challenges they encounter when blending contextual and parametric knowledge. Our focus is on three categories of knowledge conflicts: context-memory, inter-context, and intra-memory conflict. These conflicts can significantly impact the trustworthiness and performance of LLMs, especially in real-world applications where noise and misinformation are common. By categorizing these conflicts, exploring the causes, examining the behaviors of LLMs under such conflicts, and reviewing available solutions, this survey aims to shed light on strategies for improving the robustness of LLMs, thereby serving as a valuable resource for advancing research in this evolving area.
Introduction
Large Language Models (LLMs) act as massive repositories of "parametric knowledge," which is essentially frozen at the time of their training. However, real-world applications often require these models to integrate "contextual knowledge"—up-to-date information provided via Retrieval-Augmented Generation (RAG) or user prompts. A critical challenge arises when these sources disagree. For instance, if a model was trained to believe a specific fact that has since changed, or if it is presented with conflicting search results, how does it resolve the discrepancy? These "knowledge conflicts" lead to hallucinations, sycophancy, and a general breakdown in model trustworthiness. This survey systematically maps this landscape to help researchers build more robust, fact-aware AI systems.
A Taxonomy of Knowledge Conflicts
The core contribution of this work is a comprehensive taxonomy that divides knowledge conflicts into three distinct pillars based on the source of the contradiction.
- Context-Memory Conflict (CM): These occur when the model’s internal weights (parametric memory) contradict the provided input context. A classic example is outdated training data versus a new, retrieved fact.
- Inter-Context Conflict (IC): These arise from discrepancies between two or more external sources provided in the prompt, such as two different search results providing conflicting dates for the same event.
- Intra-Memory Conflict (IM): These are internal inconsistencies where the model provides different answers to the same question depending on how the prompt is phrased or the language used, revealing a lack of robust internal belief.
Analyzing Causes and Solutions
The survey further categorizes the causes of these conflicts—such as temporal misalignment, misinformation, and training corpus bias—and reviews the strategies developed to mitigate them. Solutions are broadly divided into pre-hoc adjustments (like specialized fine-tuning) and post-hoc strategies applied during inference (like contrastive decoding).
Key Insights
- Confirmation Bias: LLMs exhibit a strong bias toward information that aligns with their internal memory. However, they are also surprisingly susceptible to "persuasive" misinformation if it is presented in a coherent, logical style.
- The Popularity Problem: Models are more likely to trust their internal memory for "famous" entities but are easily swayed by context for "long-tail" or obscure entities.
- Source Disentanglement: A promising solution path involves training models to provide "disentangled" answers (e.g., "According to Source A, X is true; but according to my training, Y is true") rather than forcing a single choice.
Benchmarks and Quantitative Results
The survey synthesizes results from several key benchmarks:
- CM Benchmarks: Xie et al. (2023) and KRE (2023) use entity substitution to create "counter-memory" facts.
- IC Benchmarks: WikiContradiction and ClaimDiff focus on conflicting claims within documents.
- IM Benchmarks: PARAREL tests consistency across different syntactic phrasings.
Quantitative findings highlight the severity of the issue: misinformation in the context can lead to an 87% drop in performance on certain QA tasks. While advanced models like GPT-4 can identify contradictions over 70% of the time, they still struggle to resolve them correctly. Mitigation techniques like Knowledge Aware Fine-Tuning (KAFT) have shown the ability to improve model faithfulness to context by over 50% to 120% on specific benchmarks.
Future Directions
Despite progress, several gaps remain. Most research is currently focused on English, leaving "Cross-lingual Inconsistency" under-explored. Additionally, the field lacks a deep understanding of multimodal conflicts (e.g., text vs. image) and needs more "in the wild" datasets that reflect the subtle noise found in real-world RAG scenarios rather than artificial name-swapping. Understanding the internal "knowledge circuits" that prioritize information sources remains a critical frontier for explainable AI.
Summary updated