CDD Probe Diagnoses RAG Context Compliance Under Knowledge Conflict
A novel approach known as Context-Driven Decomposition (CDD) has been developed to identify instances when retrieval-augmented generation (RAG) systems depend on contradictory retrieved information rather than their own understanding. This method, detailed in an arXiv paper (2605.14473), functions as a belief-decomposition probe during inference and acts as an intervention tool for managing retrieval conflicts. In evaluations such as Epi-Scale stress tests, TruthfulQA misconception injection (N=500), and cross-model reruns, CDD uncovers three distinct patterns: in adversarial scenarios, Standard RAG only achieves 15.0% accuracy; accuracy improvements in adversarial settings are transferable across model families (Gemini-2.5-Flash, Claude Haiku/Sonnet/Opus), while causal coupling between rationale and answers does not transfer. This study explores the Context-Compliance Regime, where retrieved context influences the final output despite conflicting with the model's inherent knowledge.
Key facts
- Context-Driven Decomposition (CDD) is a belief-decomposition probe for RAG systems.
- CDD operates at inference time and serves as an intervention mechanism.
- Standard RAG achieves 15.0% accuracy on TruthfulQA misconception injection (N=500).
- CDD improves accuracy on Gemini-2.5-Flash and Claude Haiku/Sonnet/Opus.
- Rationale-answer causal coupling does not transfer across model families.
- The Context-Compliance Regime occurs when retrieved context dominates despite conflict with parametric knowledge.
- Tests include Epi-Scale stress tests and cross-model reruns.
- The paper is available on arXiv with ID 2605.14473.
Entities
Institutions
- arXiv