CogRAG+: Cognitive-Level Guided Diagnosis and Remediation of Memory and Reasoning Deficiencies in Professional Exam QA
New training-free framework decouples retrieval and reasoning, slashing unanswered rates by over 80%.
A new paper from researchers introduces CogRAG+, a training-free framework designed to address the opaque inference processes of large language models in professional domains. The method decouples retrieval-augmented generation (RAG) into components that align with human cognitive hierarchies, tackling knowledge gaps and reasoning inconsistencies.
CogRAG+ features Reinforced Retrieval, a judge-driven dual-path strategy with fact-centric and option-centric paths to strengthen retrieval and prevent cascading failures from missing foundational knowledge. It then applies cognition-stratified Constrained Reasoning, replacing unconstrained chain-of-thought generation with structured templates to reduce logical inconsistency and generative redundancy. On the Registered Dietitian qualification exam, CogRAG+ raised accuracy to 85.8% for Qwen3-8B and 60.3% for Llama3.1-8B, with the unanswered rate dropping from 7.6% to 1.4%.
- CogRAG+ is a training-free framework that decouples RAG into human cognitive hierarchies.
- Reinforced Retrieval uses dual paths (fact-centric and option-centric) to strengthen retrieval.
- Constrained Reasoning reduces unanswered rate from 7.6% to 1.4% using structured templates.
Why It Matters
CogRAG+ offers a model-agnostic path to expert-level performance in specialized domains without costly training.