Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict
CDD method enhances RAG model accuracy under knowledge conflict, especially in entity swap and logical contradiction scenarios.
Key Findings
Methodology
The paper introduces Context-Driven Decomposition (CDD) to diagnose context compliance in RAG models under knowledge conflict. CDD decomposes the process into five steps: extracting contextual answers, extracting parametric answers, comparing them, isolating conflicting premises, and resolving conflicts. This method makes model behavior observable and testable during conflict resolution.
Key Results
- In the TruthfulQA misconception injection test, standard RAG model accuracy was only 15.0%, while CDD improved it to 62.0%.
- In the Epi-Scale test, CDD achieved 88.0% and 83.2% accuracy in entity swap and logical contradiction scenarios, respectively, significantly outperforming baseline models.
- CDD showed adversarial accuracy improvements on the Gemini-2.5-Flash model, with directional gains across Claude variants.
Significance
This study provides a novel observation and diagnostic method by revealing RAG model behavior under knowledge conflict. The CDD method not only improves model accuracy in specific conflict types but also offers new insights for future model improvements. Its introduction helps enhance model robustness in handling complex information, holding significant academic and practical value.
Technical Contribution
The CDD method makes RAG model behavior observable under knowledge conflict through explicit conflict decomposition steps. Compared to existing methods, CDD not only improves model accuracy but also uncovers potential mechanisms in conflict handling. This offers new theoretical and engineering possibilities for future model design.
Novelty
CDD is the first to treat conflict handling as an observability problem, making model behavior testable under conflict through decomposition of contextual and parametric answers. This approach differs from traditional RAG models, offering new perspectives and tools.
Limitations
- CDD is less effective in handling non-localized conflicts such as temporal shifts and distractor evidence.
- The method requires high computational resources, especially for large datasets.
Future Work
Future research could explore the applicability of CDD across different model architectures and how to reduce its computational cost. Additionally, integrating CDD with other conflict detection and resolution methods could enhance overall model performance.
AI Executive Summary
When handling complex information, RAG models typically rely on retrieved evidence. However, when this evidence conflicts with the model's prior knowledge, performance can be severely impacted. Existing methods often overlook this issue, focusing only on the correctness of the final answer.
This paper introduces a new method—Context-Driven Decomposition (CDD)—which makes model behavior observable and testable during conflict resolution by decomposing conflicts into multiple steps. Experimental results show that CDD significantly improves model accuracy in handling localized conflicts such as entity swaps and logical contradictions.
While the CDD method excels in specific conflict types, its effectiveness in handling non-localized conflicts is less pronounced. Future research could explore how to integrate CDD with other methods to enhance overall model performance and reduce computational costs.
Deep Analysis
Background
With the advancement of natural language processing, Retrieval-Augmented Generation (RAG) models have excelled in handling complex information. However, when retrieved evidence conflicts with the model's prior knowledge, performance can be affected. Existing methods often overlook this issue, focusing only on the correctness of the final answer.
Core Problem
RAG model performance is often unstable under knowledge conflict, especially when retrieved evidence contradicts the model's prior knowledge. Effectively detecting and handling these conflicts is a pressing issue.
Innovation
The paper introduces Context-Driven Decomposition (CDD), which makes model behavior observable under conflict through explicit decomposition steps. This method not only improves model accuracy but also uncovers potential mechanisms in conflict handling.
Methodology
- �� Extract contextual answers: Derive answers from retrieved evidence.
- �� Extract parametric answers: Derive answers from the model's prior knowledge.
- �� Compare them: Detect conflicts between the two.
- �� Isolate conflicting premises: Identify specific premises causing the conflict.
- �� Resolve conflicts: Analyze and adjust to resolve conflicts and generate the final answer.
Experiments
Experiments were conducted on the Epi-Scale dataset, comprising 2,250 base samples and 2,250 adversarial samples. Standard RAG, CDD, and other baseline models were compared to evaluate CDD's performance under different conflict types.
Results
CDD achieved 88.0% and 83.2% accuracy in entity swap and logical contradiction scenarios, respectively, significantly outperforming baseline models. This indicates CDD's significant advantage in handling localized conflicts.
Applications
The CDD method can be applied in fields requiring complex information handling, such as law, medicine, and finance. By improving model accuracy under conflict, CDD enhances automated decision-making capabilities in these fields.
Limitations & Outlook
CDD is less effective in handling non-localized conflicts. Additionally, it requires high computational resources, especially for large datasets. Future research could explore how to reduce its computational cost.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. You have a recipe book (the model's prior knowledge), but a friend texts you saying an ingredient is poisonous (retrieved evidence). You need to decide whether to trust this information. The CDD method acts like an assistant, helping you analyze whether the information is trustworthy and guiding you to make the right decision. By breaking down the problem, CDD helps you better understand and handle these conflicts.
ELI14 Explained like you're 14
Imagine you're playing a game, and there's a question you need to answer. You have a game guide (the model's prior knowledge), but the game suddenly gives you a hint saying the guide is wrong. You need to decide which to trust. CDD is like a smart assistant, helping you analyze the information and figure out which is correct. It guides you step by step to solve the problem and do better in the game!
Glossary
RAG (Retrieval-Augmented Generation)
A model combining retrieval and generation to enhance answer accuracy by retrieving external evidence.
Used to improve answer accuracy when handling complex information.
CDD (Context-Driven Decomposition)
A diagnostic method that decomposes conflicts into five steps, making model behavior observable under conflict.
Used to diagnose RAG model performance under knowledge conflict.
Context Compliance
Refers to whether the model follows retrieved evidence when answering, even if it conflicts with prior knowledge.
Used to evaluate model performance under knowledge conflict.
Entity Swap
An adversarial sample generation method that creates conflicts by swapping entities in text.
Used in experiments to test model performance under localized conflicts.
Logical Contradiction
Logical inconsistencies in text that may lead to incorrect model answers.
Used in experiments to test model performance under logical conflicts.
Open Questions Unanswered questions from this research
- 1 How to improve the applicability of CDD without increasing computational cost?
- 2 How to integrate CDD with other conflict detection methods to enhance overall model performance?
Applications
Immediate Applications
Legal Document Analysis
CDD can be used to analyze conflicting information in legal documents, helping lawyers quickly find key evidence.
Medical Diagnosis
In medicine, CDD can assist doctors in identifying and resolving conflicts in diagnostic information, improving diagnostic accuracy.
Long-term Vision
Intelligent Decision Systems
By enhancing model robustness under complex information, CDD is expected to drive the development of intelligent decision systems across industries.
Abstract
Retrieval-Augmented Generation (RAG) is usually evaluated by whether the final answer is correct. Under knowledge conflict, this hides a key question: did the model follow retrieved evidence, rely on its parametric prior, or produce a post-hoc rationale? We study this as context compliance, the regime in which retrieved context controls the answer even when it conflicts with the model's prior knowledge. We introduce Context-Driven Decomposition (CDD), an inference-time diagnostic intervention that elicits contextual and prior answers, isolates the conflicting premise, and records a resolution trace that can be perturbed. Across Epi-Scale stress tests, TruthfulQA misconception injection, and cross-model reruns, CDD makes three behaviors visible. First, misleading retrieval can severely degrade accuracy: under a worst-case TruthfulQA misconception-injection probe, Standard RAG reaches only 15.0%. Second, better answers need not share the same mechanism: CDD improves adversarial accuracy on Gemini-2.5-Flash and shows directional gains across Claude variants, yet trace-perturbation sensitivity is high only on Gemini. Third, explicit decomposition improves controlled-conflict robustness over a conflict-aware instruction baseline on localized factual conflicts, with the clearest margins on Entity Swap (88.0% vs 79.3%) and Logical Contradiction (83.2% vs 75.4%). We frame RAG conflict handling as an observability problem.