GroupRAG: Cognitively Inspired Group-Aware Retrieval and Reasoning via Knowledge-Driven Problem Structuring

Xinyi Duan; Yuanrong Tang; Jiangtao Gong

cs.IRMar 26, 2026

GroupRAG: Cognitively Inspired Group-Aware Retrieval and Reasoning via Knowledge-Driven Problem Structuring

Authors: Xinyi Duan, Yuanrong Tang, Jiangtao Gong

Organizations: Tsinghua University

Abstract

The performance of language models is commonly limited by insufficient knowledge and constrained reasoning. Prior approaches such as Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) address these issues by incorporating external knowledge or enforcing linear reasoning chains, but often degrade in real-world settings. Inspired by cognitive science, which characterizes human problem solving as search over structured problem spaces rather than single inference chains, we argue that inadequate awareness of problem structure is a key overlooked limitation. We propose GroupRAG, a cognitively inspired, group-aware retrieval and reasoning framework based on knowledge-driven keypoint grouping. GroupRAG identifies latent structural groups within a problem and performs retrieval and reasoning from multiple conceptual starting points, enabling fine-grained interaction between the two processes. Experiments on MedQA (medical) and Bar Exam QA (legal) show that GroupRAG outperforms representative RAG- and CoT-based baselines. These results suggest that explicitly modeling problem structure, as inspired by human cognition, is a promising direction for robust retrieval-augmented reasoning.

Figures & tables

Figure 1: Reasoning Paradigms Comparison. Traditional CoT follows Linear/Divergent paths on unstructured sequences. GroupRAG transforms monolithic inputs into a structured problem space, employing a Convergent net via keypoint grouping for real-world alignment.

Figure 2: Illustration of the GroupRAG reasoning process on a clinical case, featuring keypoint extraction, knowledge-driven grouping, local and global reasoning, and answer alignment.

Figure 3: An Abstract Overview of GroupRAG.

	Extract F1		Group F1		Local Acc.(%)		Global WIF		Answer Acc.(%)
	Med	Bar	Med	Bar	Med	Bar	Med	Bar	Med	Bar
GroupRAG	0.96	0.92	0.80	0.74	73.1	74.7	1.13	1.02	71.7	66.4
w/o Ext. T.	0.95	0.90	0.81	0.73	70.2	71.8	1.06	0.95	68.5	62.1
w/o Gro. T.	0.96	0.92	0.71	0.65	64.5	67.9	0.93	0.87	63.0	59.7
w/o Loc. T.	0.96	0.92	0.80	0.74	61.8	63.1	0.85	0.76	64.2	62.5
w/o Glo. T.	0.96	0.92	0.80	0.74	73.1	74.7	0.69	0.67	68.2	63.7

Table 1: Leave-One-Out Ablation Results. Each variant is compared with the full pipeline (first row) to demonstrate the marginal contribution of each module. “w/o” indicates removing each component independently from the full pipeline (T.: Training; R.: RAG; Med: MedQA; Bar: Bar Exam QA).

	Extract F1		Group F1		Local Acc.(%)		Global WIF		Answer Acc.(%)
	Med	Bar	Med	Bar	Med	Bar	Med	Bar	Med	Bar
GroupRAG	0.96	0.92	0.80	0.74	73.1	74.7	1.13	1.02	71.7	66.4
$-$ Ext. T.	0.95	0.90	0.81	0.73	70.2	71.8	1.06	0.95	68.5	63.5
$-$ Gro. T.	0.95	0.90	0.72	0.67	62.1	63.9	0.94	0.84	63.0	58.7
$-$ Gro. R.	0.95	0.90	0.70	0.63	61.8	62.1	0.88	0.81	61.0	57.3
$-$ Loc. T.	0.95	0.90	0.70	0.63	58.5	58.1	0.75	0.71	56.7	53.2

Table 2: Progressive Ablation Results. Each variant is compared with the previous row to demonstrate the cumulative effect of the integrated modules. “

-

” indicates cumulative removal (T.: Training; R.: RAG; Med: MedQA; Bar: Bar Exam QA).

Dataset	Model	Method
Dataset	Model	Direct (%)	+CoT Prompting(%)	+Naive RAG(%)	+Group RAG(%)
MedQA	GPT-4o	$\mathbf{89.1}_{\pm 0.8}$	$90.3_{\pm 0.7}$	$88.7_{\pm 0.7}$	$\mathbf{91.5}_{\pm 0.5}$
	L3.1-8B(base)	$\mathbf{48.1}_{\pm 1.0}$	$54.4_{\pm 1.0}$	$53.9_{\pm 1.1}$	$\mathbf{61.0}_{\pm 0.8}$
	L3.1-8B(trained)	$\mathbf{52.9}_{\pm 0.6}$	$61.3_{\pm 0.9}$	$59.1_{\pm 0.8}$	$\mathbf{71.7}_{\pm 0.7}$
BarExam	GPT-4o	$\mathbf{77.5}_{\pm 0.7}$	$79.3_{\pm 0.8}$	$78.2_{\pm 0.9}$	$\mathbf{80.7}_{\pm 0.8}$
	L3.1-8B(base)	$\mathbf{42.9}_{\pm 1.0}$	$48.6_{\pm 0.8}$	$47.8_{\pm 0.9}$	$\mathbf{55.0}_{\pm 0.9}$

Table 3: Joint comparison across datasets, models, and methods. All values report mean accuracy over 5 independent runs with standard deviation. L3.1-8B refers to LLaMA 3.1-8B.

Retrieval-Augmented Reasoning

See all 27 papers

CogRAG: Tackling Heterogeneous Cognitive Demands in RAG via Stratified Retrieval and Reasoning
Apr 1, 2026Xudong Wang, Zilong Wang, Kui Su +1Structured Reasoning in Language Models Adaptive RAG
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
When to Retrieve During Reasoning: Adaptive Retrieval for Large Reasoning Models
Apr 29, 2026Dongxin Guo, Jikun Wu, Siu Ming YiuRetrieval-Augmented Generation Retrieval-Augmented Reasoning
The University of Hong Kong Hong Kong, China · Brain Investing Limited Hong Kong, China · Stellaris AI Limited Hong Kong, China
Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
Jun 11, 2026Zilin Xiao, Qi Ma, Chun-cheng Jason Chen +4Analogical Reasoning RL for Language Model Reasoning
Meta Superintelligence Labs · Rice University · Work partially done during internship at Meta

Apr 1, 2026J/K move · Enter open · S save

CogRAG: Tackling Heterogeneous Cognitive Demands in RAG via Stratified Retrieval and Reasoning

Xudong Wang, Zilong Wang, Kui Su, Zhaoyan Ming

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

Structured Reasoning in Language Models Adaptive RAG RAG Robustness Reasoning Consistency in Language Models Large Reasoning Models

Open paper AskPDF arXiv

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.

Abstract

Retrieval-Augmented Reasoning

Explore similar work

CogRAG: Tackling Heterogeneous Cognitive Demands in RAG via Stratified Retrieval and Reasoning

When to Retrieve During Reasoning: Adaptive Retrieval for Large Reasoning Models

Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning