Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict

Researchers introduce Context-Driven Decomposition (CDD), a diagnostic tool that reveals how retrieval-augmented generation (RAG) systems blindly follow retrieved context even when it contradicts their underlying knowledge. Testing across multiple AI models shows CDD can improve accuracy to 64% on adversarial scenarios, though improvements don't consistently transfer across different model families, suggesting RAG systems resolve conflicts through fundamentally different mechanisms.

This research addresses a critical vulnerability in RAG systems—the tendency to prioritize retrieved external context over parametric knowledge, even when that context is demonstrably false. The problem matters because RAG has become foundational for enterprise AI applications, financial analysis tools, and knowledge-intensive systems where accuracy directly impacts business decisions. Standard RAG reached only 15% accuracy when fed intentionally wrong information, exposing systemic compliance with flawed context rather than truth-seeking behavior.

The Context-Driven Decomposition technique operates as an inference-time diagnostic, decomposing how models weigh conflicting information sources. Testing across Gemini-2.5-Flash and Claude variants reveals that while CDD improvements transfer between model families—accuracy gains appear real—the underlying causal mechanisms differ sharply. Gemini shows explicit conflict-resolution traces at 64% sensitivity, whereas Claude variants achieve similar accuracy gains through opaque mechanisms unrelated to rationale-answer coupling. This disconnect suggests models internalize different decision-making architectures despite achieving comparable performance.

For developers building production RAG systems, this research highlights architectural limitations that generic retrieval quality improvements cannot solve. The release of Epi-Scale benchmarks enables systematic evaluation of context-compliance behavior across retrieval pipelines and model families. Financial applications relying on RAG for market research, compliance analysis, or risk assessment should recognize that accuracy metrics alone mask susceptibility to adversarial or outdated context. Organizations should implement independent validation layers and conflict-detection mechanisms rather than assuming retrieved context automatically improves reliability. The research motivates investigating why different models resolve information conflicts through distinct pathways—understanding these pathways could yield more robust architectures.

• →RAG systems exhibit 'context compliance,' prioritizing retrieved information even when it contradicts accurate parametric knowledge, reaching only 15% accuracy under adversarial misconception injection.
• →Context-Driven Decomposition improves accuracy to 64% on Gemini-2.5-Flash but uses fundamentally different conflict-resolution mechanisms across Claude variants, indicating non-transferable improvements.
• →CDD demonstrates 71% robustness against temporal drift and 70% against noisy distractors, suggesting explicit conflict decomposition strengthens RAG resilience beyond standard retrieval quality.
• →Accuracy gains from conflict-aware RAG don't correlate with transparent causal reasoning, implying models achieve correctness through opaque mechanisms distinct from explicit reconciliation.
• →Epi-Scale benchmark release enables systematic evaluation of context-compliance behavior, critical for enterprises deploying RAG in high-stakes financial, legal, and compliance applications.