Source-linked AI summary
Knowledge Conflicts for LLMs: A Survey
Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, Wei Xu
TL;DR
LLMs face conflicts when contextual information interacts with parametric knowledge, but existing reviews do not comprehensively organize this problem. This survey categorizes conflict types, examines their causes and model behaviors, and reviews solutions. It concludes that knowledge conflicts are multifaceted, interact in complex ways, and are often studied through artificially constructed scenarios that sacrifice granularity and breadth.
Problem
Knowledge conflicts between contextual information and parametric knowledge affect LLM behavior, while prior reviews provide limited or scenario-specific coverage.
Method
The survey categorizes knowledge conflicts, analyzes their causes and effects on model behavior, and systematically reviews solutions organized by desired behaviors.
Results
Knowledge conflict is multifaceted: model behavior depends on the conflicting knowledge type, the three conflict types interact in complex ways, and existing solutions mainly address artificially constructed scenarios.
Takeaways & Limitations
The survey provides a framework for understanding conflict categorization, causes, model responses, and possible solutions as retrieval-augmented use increases conflict complexity.
Takeaways & Limitations
The survey may miss some recent or less relevant findings despite including the authors’ identified essential materials.
Abstract
from arXiv · showhide
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.
1 Introduction
Knowledge conflicts arise when LLMs combine contextual knowledge with parametric memory, producing discrepancies that must be resolved for accurate responses. This survey organizes these conflicts by type, examines their causes and behavioral effects, and reviews solutions.
- Knowledge sources: LLMs combine parametric knowledge stored in model parameters with contextual knowledge from prompts, dialogue history, and retrieved documents.This combination enables current and accurate responses but creates opportunities for discrepancies between knowledge sources.
- Conflict taxonomy: The survey distinguishes context-memory, inter-context, and intra-memory conflicts.Context-memory conflicts involve context versus model memory; inter-context conflicts involve multiple contextual sources; intra-memory conflicts reflect inconsistent parametric knowledge across differently phrased questions.
- Survey scope: The survey aims to provide a comprehensive review because prior reviews offered only brief or scenario-specific coverage of knowledge conflicts.Its scope includes categorization, cause and behavior analysis, and solutions.
- Causes and behaviors: Knowledge conflicts connect causal triggers to model behaviors and can contribute to factually incorrect information, including hallucinations.The survey emphasizes understanding how conflicts emerge and affect model behavior rather than treating them only as isolated phenomena.
- Research gap: Existing analyses often construct conflicts artificially, leaving the interconnectedness of knowledge conflicts insufficiently addressed.The survey positions its broader review as a response to this limitation in prior work.
- Solutions: The survey systematically reviews solutions organized as pre-hoc or post-hoc strategies according to whether adjustments occur before or after potential conflicts arise.It discusses causes, model behaviors, and available solutions for each conflict type.
2 Context-Memory Conflict
Context-memory conflict arises when external contextual knowledge disagrees with an LLM’s parametric knowledge, especially because model memory is static while external information changes or may contain misinformation. Research finds that models’ responses depend on the coherence, relevance, and credibility of conflicting information, motivating strategies that distinguish faithful context use from misinformation resistance.
- Causes: Context-memory conflict reflects discrepancies between external context and parametric knowledge caused chiefly by temporal misalignment or misinformation pollution.Temporal misalignment arises because models are trained on past data, whereas misinformation pollution introduces incorrect external information.
- Solutions: Mitigation strategies include updating model parameters, retrieving external documents, prompting for context faithfulness or misinformation vigilance, and contrastive decoding.Knowledge editing changes parametric knowledge, whereas retrieval supplements it without altering parameters; other approaches modify training, prompts, or output probabilities.
- Causes: External documents may be poisoned with false information, and prompt injection can exploit user conversations to induce misinformation propagation.These attacks threaten systems that incorporate retrieved documents or conversational context without adequately checking reliability.
- Model behaviors: LLMs can identify that knowledge conflicts exist but struggle to locate the conflicting segments and provide distinct answers for contradictory information.Misleading prompts particularly affect factual robustness, while desired behavior includes separating and presenting the competing answers.
- Model behaviors: LLMs may favor either parametric or contextual knowledge, but coherent, logical, and compelling external evidence is typically preferred over generic conflicting information.Studies report both receptivity to convincing external evidence and confirmation bias toward information consistent with internal memory.
- Solutions: Current mitigation approaches can pursue contradictory goals because blindly privileging either context or memory is undesirable when the conflict’s cause is unspecified.Some proposals instead favor detecting conflicts and presenting distinct answers so users can make informed decisions.
3 Inter-Context Conflict
Inter-context conflict occurs when retrieved external documents disagree, including through misinformation or differences between updated and outdated facts. Such conflicts can impair reasoning and detection, while proposed solutions use contradiction modeling, linguistic knowledge, external evidence, and tools.
- Problem: Inter-context conflict arises when external documents incorporated through retrieval-augmented generation contain mutually inconsistent information.The conflicts may involve misinformation versus accurate information or updated versus outdated facts.
- Model behaviors: Conflicting contexts can leave model confidence largely unchanged while increasing susceptibility to misleading documents, query-irrelevant evidence, ordering effects, and reasoning difficulty.Models tend to favor query-relevant, parametric-consistent, and more widely corroborated answers; increasing conflicting hops makes reasoning harder.
- Model behaviors: LLMs perform poorly at detecting contradictory statements and documents, with additional difficulty for contradictions involving subjective emotions or perspectives.Document length and self-contradiction variety have only a minor influence on detection performance in the cited study.
- Model behaviors: Model responses to contextual conflict vary with training-data differences, and textual misinformation discernment differs substantially from human judgment.The survey identifies contextual nuance and model-specific behavior as areas requiring further exploration.
- Solutions: Solutions include specialized contradiction detectors, linguistic features, and approaches that combine model knowledge with retrieved external documents.Examples include Pairwise Contradiction Neural Network and methods designed to improve syntactic and semantic understanding for contradiction detection.
- Solutions: Unified and efficient methods for handling different knowledge-conflict types remain a formidable challenge, while external tools are identified as a promising direction.Existing strategies primarily rely on model knowledge or retrieved documents.
4 Intra-Memory Conflict
Intra-memory conflict is inconsistent behavior arising from conflicting knowledge within model parameters, with causes spanning training, decoding, and knowledge editing. Studies report inconsistency across paraphrases, layers, and languages, while proposed remedies target training, generation, and post-hoc processing.
- Definition and scope: Intra-memory conflict produces divergent responses to semantically equivalent prompts, undermining consistent outputs in knowledge-intensive systems.The survey attributes this conflict to inconsistent knowledge embedded in parameters and emphasizes its relevance to question-answering systems.
- Causes: Training corpus bias can encode incorrect knowledge and amplify inaccuracies, allowing conflicting knowledge to coexist within model parameters.The survey identifies training corpus bias as a fundamental catalyst for intra-memory conflict.
- Causes: Stochastic decoding introduces uncertainty and can yield entirely different content from the same context because early token choices affect later generations.The survey contrasts stochastic sampling with deterministic methods and links its randomness to intra-memory conflict.
- Causes: Knowledge editing can create inconsistent responses to the same knowledge in varying situations, making intra-memory conflict a side effect of post-hoc modification.Editing changes a limited scope of encoded knowledge, but consistency across situations remains challenging.
- Observed behaviors: GPT-4 shows a 13% inconsistency rate on Commonsense Question-Answering, while earlier encoder-based models report consistency accuracy of only 50% to 60%.The reported inconsistency is higher for uncommon knowledge, and multiple open-source LLMs also exhibit strong inconsistencies.
- Observed behaviors: A 40% disparity separates knowledge-probe accuracy from generation accuracy, indicating that correct knowledge may be stored in parameters but not expressed during generation.The comparison uses two experiments on the same LLM: generation accuracy and a knowledge probe measuring knowledge containment.
- Observed behaviors: LLMs maintain language-dependent knowledge sets, and RankC analyses find no improvement in cross-lingual consistency with increased model size.The survey reports that knowledge related to different languages is stored separately within model parameters.
- Solutions: Existing remedies span training, generation, and post-hoc processing, but efficiently resolving inconsistencies in parameter knowledge remains difficult.The survey describes training methods as targeting parameter inconsistency and generation or post-hoc methods as algorithmic interventions.
5 Challenges and Future Directions
The survey identifies unresolved challenges in evaluating and explaining knowledge conflicts, studying their interactions, and extending research beyond artificially constructed English text scenarios. It calls for finer-resolution and more realistic analyses across applications, languages, and modalities.
- Knowledge Conflicts in the Wild: Real-world retrieval-augmented systems contain conflicts in documents retrieved directly from the Web, but artificially generated conflicts may not transfer easily to practice.Recent work has begun curating conflicting documents from actual Google search results for open-ended questions.
- Solution at a Finer Resolution: No one-size-fits-all solution exists because knowledge conflicts are complex, and current approaches assume priors or target only conflict subclasses.The survey argues for finer-grained treatment that considers factors such as the nature of the user’s query.
- Evaluation on Downstream Tasks: Research mainly evaluates knowledge conflicts on QA datasets, leaving their effects on broader downstream applications insufficiently explored.The survey highlights legal document analysis and medical diagnosis as examples of accuracy- and consistency-sensitive applications.
- Interplay among the Conflicts: Most studies examine a single conflict type or a joint inter-context and context-memory setting, while interactions involving intra-memory conflict remain understudied.Studying these interactions could clarify relationships between internal inconsistency and responses to context, as well as knowledge representation and processing.
- Explainability: Output-level analyses dominate explainability research, with less attention to internal mechanisms such as attention heads and neuron activations during conflicts.The survey identifies microscopic examination of model decisions under conflict as a needed direction.
- Multilinguality: Knowledge-conflict research has primarily focused on English, motivating studies of non-English prompts and language-specific differences.The survey points to multilingual models and non-English models as resources for extending this work.
- Multimodality: Knowledge-conflict research mainly concerns text, leaving multimodal conflicts involving images, video, and audio as a promising area for exploration.The survey expects conflicts to become more complex as LLMs process information across formats.
6 Conclusion
The survey categorizes knowledge conflicts, analyzes their causes and behavioral manifestations, and reviews possible solutions. It concludes that conflict behavior depends on the type of conflicting knowledge and that existing work needs broader, more granular, and more realistic evaluation.
- Conclusion: Knowledge conflict is multifaceted: model behavior depends closely on the particular type of conflicting knowledge.The survey examines context-memory, inter-context, and intra-memory conflicts together with their causes, behaviors, and solutions.
- Conclusion: Existing solutions largely target artificially constructed scenarios and rely on assumed priors, sacrificing granularity and breadth.The survey connects this limitation to the need for more comprehensive research as retrieval-augmented language models become more widespread.
Limitations
The survey may omit some recent or less relevant findings despite including what the authors describe as essential materials.
- Limitations: The survey may have missed some recent or less relevant findings because of the rapid growth and large volume of research.The authors nevertheless state that they included all essential materials.
Ethics Statement
The survey searched post-2021 literature using targeted terms across Google Scholar and the ACL Anthology, then expanded coverage through citation tracking.
- The search mainly targeted papers published after 2021.
- Search terms included “knowledge conflict”, “knowledge inconsistency”, and “knowledge gap”.
- The authors classified initially identified papers by reading them and tracked related papers through citations.
A.1 Quantitative Analysis and Comparison
The survey acknowledges that quantitative synthesis is difficult because studies use disparate datasets, changing model knowledge, and different mitigation objectives.
- Quantitative experiments were considered complex and impractical because behavioral analyses use disparate datasets.
- Fair mitigation-strategy comparisons are difficult because methods pursue objectives shaped by different assumptions about context and inherent knowledge.
- Reported quantitative results should be interpreted cautiously because studies may use different datasets, LLM variants, or objectives.
A.2 Quantitative Results on the Impact of Knowledge Conflicts
The survey compares quantitative evidence on how the three knowledge-conflict types affect LLM behavior, using representative behavior-analysis studies.
- Table 2 compares quantitative results on the impact of the three knowledge-conflict types.
- The comparison concerns the impact of various knowledge-conflict types.
- Representative behavior-analysis literature provides the results selected for comparison.
A.3 Quantitative Results on the Effectiveness of Mitigation Strategies
The survey quantitatively compares mitigation-strategy effectiveness with respect to objectives, but only where enough literature supports meaningful cross-method comparison.
- Table 3 compares the effectiveness of various mitigation strategies with respect to their objectives.
- The analysis is limited to works addressing three predominant mitigating objectives for memory conflicts.
- Other mitigation objectives lack enough literature for meaningful cross-method comparison.