CogRAG: Tackling Heterogeneous Cognitive Demands in RAG via Stratified Retrieval and Reasoning
Authors: Xudong Wang, Zilong Wang, Kui Su, Zhaoyan Ming
Organizations: School of Computer and Computing Science, Hangzhou City University, Hangzhou, Zhejiang 310015, China · Institute of Digital Twin, Eastern Institute of Technology, Ningbo, Zhejiang 315200, China · Innovation Center of Yangtze River Delta, Zhejiang University, Jiaxing, Zhejiang 314100, China
Retrieval-Augmented Generation (RAG) frameworks typically process all queries through a one-size-fits-all pipeline, ignoring the heterogeneous cognitive demands of different tasks. This cognitive-blind approach causes two failure modes: cascading errors when low-level factual gaps trigger hallucinated reasoning, and reasoning-answer inconsistency in higher-order analytical tasks. We introduce CogRAG, a training-free, domain-agnostic framework that tackles these heterogeneous cognitive demands via stratified retrieval and reasoning. Inspired by Bloom's Taxonomy, CogRAG uses the predicted cognitive load of a query as a central control signal that coordinates two modules: Cognition-Adaptive Evidence Refinement supplements missing context via fact-centric or option-centric paths, and Cognition-Stratified Structured Reasoning replaces unconstrained chain-of-thought with cognition-aligned reasoning templates. We evaluate CogRAG on a demanding professional testbed, the Registered Dietitian qualification examination. CogRAG effectively reduces early-stage factual errors and eliminates reasoning-answer inconsistency, raising Qwen3-8B accuracy from 73.4% to 85.8% in single-choice mode and from 63.3% to 80.5% in scenario mode. These results highlight cognitive-stratified control as an effective, generalizable paradigm for reliable complex reasoning in large language models.
Figures & tables
Figure 1: Representative bottlenecks in professional nutrition question answering. Low-level factual errors can cascade into high-level reasoning failures, while reasoning-answer inconsistency may cause the model to derive correct evidence but select the wrong option.
Figure 2: Illustration of Bloom’s Taxonomy, which organizes cognitive activities from basic knowledge recall to higher-order reasoning and creative synthesis.
Figure 3: Architecture of CogRAG, a cognition-driven control framework. A central Cognitive Demand Profiler (top) estimates the cognitive load of the input query based on Bloom’s Taxonomy and emits a binary route z*. This signal acts as a global control input (dashed arrows) that simultaneously governs two downstream modules: Cognition-Adaptive Evidence Refinement (left) and Cognition-Stratified Structured Reasoning (right). Solid arrows denote the data flow (query → evidence → answer), while dashed arrows denote the cognitive control flow.
Mode
Field
Constraint
LOW
key_fact
One-sentence core definition or classification.
evidence
Concise points directly extracted from retrieved evidence.
elimination
Brief rationale for excluding distractors, within three sentences.
answer
Final option letter.
HIGH
assumptions
Known conditions provided in the question.
rules
2–5 applicable professional guidelines or mechanisms.
Table 1: Structured proof templates for Cognition-Stratified Structured Reasoning (CSSR).
Source
# QA
Method
T1
T2
T3
T4
T5
T6
FoodEarth
25,629
Curated
24817
0
812
0
0
0
MedQA
12,652
Curated
0
12652
0
0
0
0
Nutri7Base
24,197
Synthetic †
0
0
14013
15444
10676
1219
Total
62,478
–
24817
12652
14825
15444
10676
1219
† Generated via Qwen3-Max based on raw nutrition corpora.
Table 2: Summary of the RAG Knowledge Base. The corpus comprises 62,478 nutrition-specific QA pairs. Tags (T1-T6) represent our 6-category taxonomy for topic-aware retrieval.
General Acc. (%)
Cognitive Level Acc. (%)
Model
Mode
Method
Overall
Macro
Rem.
Und.
App.
Ana.
Eva.
Qwen
Single
Baseline
73.4
72.0
72.7
73.5
65.3
77.0
71.8
BM25
72.1
71.0
69.8
72.7
66.3
74.6
71.8
Dense
78.2
77.8
80.6
77.1
74.5
79.8
76.9
Hybrid
71.9
70.0
69.8
73.4
62.2
75.4
69.2
Base Ret.
79.5
80.3
81.4
77.8
76.5
81.0
84.6
Table 3: Performance comparison on the RD exam benchmark. Bold and underlined values indicate the best and second-best results, respectively.
Components
General Acc.(%)
Cognitive Level Acc.(%)
Base Retriever
CAER
CSSR
Overall
Macro
Rem.
Und.
App.
Ana.
Eva.
73.4
72.0
72.7
73.5
65.3
77.0
71.8
✓
79.5
80.3
81.4
77.8
76.5
81.0
84.6
✓
✓
82.9
83.5
83.0
82.6
78.6
83.7
89.7
✓
✓
83.7
85.7
88.4
82.6
80.6
83.3
87.2
✓
✓
✓
85.8
87.3
90.7
84.6
84.7
84.1
92.3
Table 4: Ablation study on Qwen3-8B in the single-choice setting. Base Retriever denotes tag-constrained dense retrieval; CAER and CSSR denote the two proposed CogRAG modules.
Figure 4: Distribution of error onset levels across target cognitive levels before and after CogRAG. Each group corresponds to the target cognitive level of the question, and stacked bars indicate the earliest cognitive stage at which the reasoning error begins.
Method
LOW (Fact-Centric)
HIGH (Option-Centric)
Rel. Score
Accuracy (%)
Rel. Score
Accuracy (%)
Base Retriever
75.2
78.9
71.5
80.2
+ CAER
77.8
82.7
74.8
83.0
Table 5: Effect of Cognition-Adaptive Evidence Refinement (CAER) on evidence quality and answer accuracy.
Method
Level
Unanswered (%)
Consistency (%)
Accuracy (%)
Base Retriever
LOW
0.0
–
78.9
HIGH
0.0
–
80.2
Overall
0.0
–
79.5
+ Standard CoT
LOW
6.5
98.6
81.6
HIGH
8.6
96.9
77.5
Overall
7.6
97.4
80.1
Table 6: Validity, reasoning-answer consistency, and accuracy comparison of different reasoning strategies under the same base retrieval setting. Direct denotes direct option selection based on the retrieved evidence. Unanswered denotes the proportion of outputs without a valid final option. Consistency measures the proportion of valid responses in which the final selected option is supported by the generated reasoning trajectory.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1: Baseline accuracy across question formats and cognitive levels. The model is evaluated on 811 single-choice questions and 379 scenario-based questions.
Figure A.2: Effect of cognitive-level injection on single-choice questions. The baseline model is compared with a variant that explicitly includes the cognitive level in the prompt.
Figure A.3: Comparison of prompt-level cognitive guidance strategies. Level Only injects only the cognitive-level label, whereas Level + Description additionally includes its textual description.
Setting
Direct 2-way
5-to-2 Mapping
Overall
Low
High
Overall
Low
High
Zero-shot
58.1
98.1
14.7
70.2
63.0
73.0
Few-shot
63.4
96.7
27.3
80.6
87.2
74.3
Appendix
Table B.1: Comparison of routing hit rates between direct binary prediction and five-class-to-binary mapping.
Method
Overall
Macro
Rem.
Und.
App.
Ana.
Eva.
Baseline
73.4
72.0
72.7
73.5
65.3
77.0
71.8
BM25
72.1
71.0
69.8
72.7
66.3
74.6
71.8
Dense
78.2
77.8
80.6
77.1
74.5
79.8
76.9
Hybrid
71.9
70.0
69.8
73.4
62.2
75.4
69.2
Appendix
Table C.1: Comparison of retrieval methods under the Qwen3-8B single-choice setting.