Title: CogRAG: Addressing Diverse Cognitive Requirements in RAG Through Layered Retrieval and Reasoning

Abstract: Standard Retrieval-Augmented Generation (RAG) systems typically funnel every query through a uniform pipeline, neglecting the varied cognitive loads inherent to different tasks. This "cognitive-blind" methodology leads to two primary failure modes: hallucinated reasoning triggered by minor factual gaps at lower levels, and inconsistencies between reasoning processes and final answers in complex analytical scenarios. To address these challenges, we present CogRAG, a domain-agnostic framework that requires no training and employs stratified retrieval and reasoning mechanisms. Drawing inspiration from Bloom's Taxonomy, CogRAG utilizes the estimated cognitive load of a query as a central coordinating signal for two distinct modules. First, Cognition-Adaptive Evidence Refinement addresses missing context by selecting between fact-centric or option-centric retrieval paths. Second, Cognition-Stratified Structured Reasoning substitutes unstructured chain-of-thought processes with reasoning templates tailored to the specific cognitive level. We assessed CogRAG using the Registered Dietitian qualification examination as a rigorous professional testbed. The framework successfully mitigated early factual errors and resolved reasoning-answer inconsistencies, boosting the accuracy of Qwen3-8B from 73.4% to 85.8% in single-choice tasks and from 63.3% to 80.5% in scenario-based tasks. These findings demonstrate that cognitive-stratified control serves as a robust and generalizable paradigm for enhancing the reliability of complex reasoning in large language models.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC