CARA: Causally Guided Retrieval Agent for Question Answering

Weicheng Liu; Xiangcheng Wu; Depeng Xu; Haopeng Zhang; Xi Niu

cs.AIOct 15, 2026

CARA: Causally Guided Retrieval Agent for Question Answering

Authors: Weicheng Liu, Xiangcheng Wu, Depeng Xu, Haopeng Zhang, Xi Niu

Organizations: University of North Carolina at Charlotte

Abstract

Retrieval-augmented generation over text-rich knowledge graphs requires identifying evidence that contributes to answering a question, rather than merely matching its textual content. Existing agentic retrievers adaptively balance global search and multi-hop exploration but primarily rely on textual similarity to guide retrieval. We propose CARA, a training-free, CAusally guided Retrieval Agent that explicitly evaluates the causal relevance of retrieved evidence through two complementary criteria: consequentiality, measuring whether adding evidence improves answerability, and essentiality, assessing whether removing evidence weakens answerability. CARA integrates these assessments with path consistency to guide graph exploration and employs reciprocal rank fusion to aggregate evidence from multiple agents. We evaluate CARA on three STaRK datasets and augment STaRK with ground-truth answers to enable downstream question answering evaluation. Experimental results demonstrate that CARA consistently outperforms existing retrievers. CARA also improves answer quality and faithfulness while examining fewer graph nodes and requiring fewer retrieval operations. These findings demonstrate that causal relevance assessment enables more focused and effective agentic graph retrieval, improving both retrieval performance and downstream question answering.

Figures & tables

Figure 1: Overview of CARA. Each retrieval agent explores the knowledge graph with GlobalSearch and Neighbors . The causal reasoning module

\mathcal{C}

assesses the neighborhood candidates. The scores guide the agent’s next action. The agents’ result lists are aggregated into one ranking, which the answer generator uses to answer

Q

Conventional retrievers
	Amazon			MAG			Prime
Method	Hit@1	R@20	MRR	Hit@1	R@20	MRR	Hit@1	R@20	MRR
BM25	45.3	53.7	55.4	25.9	45.7	34.9	12.7	31.3	19.8
ada-002	39.7	53.8	50.9	29.1	48.4	38.6	12.6	36.0	21.3
GritLM-7B	49.8	61.8	60.5	39.1	61.4	50.5	15.0	38.4	23.4
KAR	49.6	59.7	59.0	50.1	63.9	58.6	23.3	46.1	32.2

Table 1: Retrieval performance (%) on the full STaRK test sets ( Amazon

n{=}1{,}642

, MAG

n{=}2{,}665

, Prime

n{=}2{,}801

). † Reported by Yu et al. (2025) (LLaMA3-8B-Instruct, test-0.1 subset). ∗ CARA version:

\lambda=0.5,\alpha=0.5

. Bold : best; underline : second best.

	Amazon		MAG		Prime
Method	BS	Faith.	BS	Faith.	BS	Faith.
No retrieval	7.5	–	16.7	–	27.9	–
KAR	28.3	88.3	33.0	86.6	36.0	88.6
ARK	30.1	87.4	39.2	78.6	40.5	80.5
CARA (ours) ∗	30.7	88.0	40.2	82.6	41.3	86.1

Table 2: Question answering on the full STaRK test sets. All methods share the same answer generator. BERTScore (BS,

\times

100) compares each answer with the ground-truth answer. Faithfulness (Faith., %) is the fraction of answer claims supported by the retrieved context. Faithfulness is not defined in the “No retrieval” approach. ∗ CARA version:

\lambda=0.5,\alpha=0.5

. Bold : best; underline : second best.

Figure 2: Retrieval (Recall@20, %) and answer quality (BERTScore, $\times$ 100) of CARA on Amazon and Prime under different exploration budgets: maximum steps per agent $T_{\max}$ (rows) and number of agents (columns). The outlined cell marks each panel’s peak (earliest budget when tied). All values are on the STaRK test-0.1 subset (10% of each test set).

	Amazon ( $n{=}1{,}642$ )					MAG ( $n{=}2{,}665$ )					Prime ( $n{=}2{,}801$ )
	Retrieval			QA		Retrieval			QA		Retrieval			QA
Variant	Hit@1	R@20	MRR	BS	Faith.	Hit@1	R@20	MRR	BS	Faith.	Hit@1	R@20	MRR	BS	Faith.
CARA (ours) ∗	62.4	60.5	71.3	30.7	88.0	84.9	80.8	88.1	40.2	82.6	59.5	73.4	67.7	41.3	86.1
w/o $\mathcal{C}$	60.5	60.2	69.1	30.4	87.7	80.3	77.9	83.5	39.2	78.5	58.1	72.1	65.4	40.8	80.7
count-based	62.2	60.5	70.3	30.6	88.0	84.4	80.8	87.4	40.1	82.8	60.0	73.1	67.2	40.9	86.3
w/o $\mathcal{C}$ , count-based	59.0	60.2	67.7	30.1	87.4	80.1	77.9	83.1	39.2	78.6	58.6	71.9	65.3	40.5	80.5

Table 3: Ablation of the causal reasoning module

\mathcal{C}

and the rank-based aggregation (RRF) on the full STaRK test sets. w/o

\mathcal{C}

: agents without

\mathcal{C}

. count-based: nodes ranked by the frequency of agents’ votes instead of RRF; w/o

\mathcal{C}

, count-based is equivalent to ARK (Tables 1 and 2 ). ∗ CARA version:

\lambda=0.5,\alpha=0.5

. Bold : best; underline : second best.

Dataset	Method	Agent	$\mathcal{C}$ module	Total	(k)	GlobalSearch	Neighbors	examined	(s)
		LLM calls			Tokens	Tool calls		Nodes	Latency
Amazon	ARK	26.9	–	26.9	310	18.2	15.2	152	105
	CARA (ours) ∗	22.1	8.4	30.6	244	13.2	10.9	122	145
MAG	ARK	31.5	–	31.5	338	16.1	30.4	163	136
	CARA (ours) ∗	25.8	20.1	45.9	340	11.9	22.3	123	223
Prime	ARK	41.2	–	41.2	469	20.2	47.8	249	141

Table 4: Per-query cost on the full STaRK test sets, summed over the three agents. Lower is better. Total counts agent calls and

\mathcal{C}

module calls. Tokens cover these calls. Latency is the mean wall-clock time per query on Nvidia H200 GPUs (Qwen3.8-27B served with vLLM). ∗ CARA version:

\lambda=0.5,\alpha=0.5

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Backbone
LLM	Qwen3.8-27B for every call, thinking mode off
Serving	vLLM 0.30.0; 6 replicas, 1 H200 each
Agents
Agents	$m=3$ per question, run in parallel
Step budget	$T_{\max}=150$ LLM calls per agent
Decoding	temperature 1.0, top- $p$ 0.95, top- $k$ 20

Appendix

Table 5: Settings of the implementation

Agentic Retrieval

Agentic retrieval lets AI agents iteratively search, inspect, and verify evidence to answer complex information needs.

See all 97 papers

CausalRAG2: Hierarchical Causal Knowledge Graph Design for RAG
Feb 4, 2026Nengbo Wang, Tuo Liang, Vikash Singh +6KG Construction Retrieval-Augmented Generation
Department of Computer and Data Sciences, Case Western Reserve University, Cleveland, OH, USA · Design and Innovation Department, Case Western Reserve University, Cleveland, OH, USA
PathRouter: Aligning Rewards with Retrieval Quality in Agentic Graph Retrieval-Augmented Generation
Jun 15, 2026Bo Wang, Heyan Huang, Yaolin Li +6Reward Shaping Graph Retrieval-Augmented Generation
Beijing Institute of Technology · Joy Future Academy
Why Neighborhoods Matter: Traversal Context and Provenance in Agentic GraphRAG
May 14, 2026Riccardo Terrenzi, Maximilian von Zastrow, Serkan AyvazRAG Citation Faithfulness Data Provenance
Centre for Industrial Software, University of Southern Denmark, Alsion 2, 6400 Sønderborg, Denmark

Feb 4, 2026J/K move · Enter open · S save

CausalRAG2: Hierarchical Causal Knowledge Graph Design for RAG

Nengbo Wang, Tuo Liang, Vikash Singh, Chaoda Song+5

Department of Computer and Data Sciences, Case Western Reserve University, Cleveland, OH, USA · Design and Innovation Department, Case Western Reserve University, Cleveland, OH, USA

KG Construction Retrieval-Augmented Generation LLM Reasoning with Graphs Graph Retrieval-Augmented Generation Knowledge Graphs

Open paper AskPDF arXiv

Retrieval augmented generation (RAG) has enhanced large language models by enabling access to external knowledge, with graph-based RAG emerging as a powerful paradigm for structured retrieval and reasoning. However, existing graph-based methods often over-rely on entity-centric node matching and lack explicit causal modeling, leading to unfaithful or spurious answers. Prior attempts to incorporate causality are typically limited to local or single-document contexts and also suffer from information isolation that arises from modular graph structures, which hinders scalability and cross-module causal reasoning. To address these challenges, we propose CausalRAG2, a framework that rethinks knowledge organization for graph-based RAG through causal gating across hierarchical modules. CausalRAG2 explicitly models causal relationships to suppress spurious correlations while enabling scalable reasoning over large-scale knowledge graphs. We also introduce HolisQA, a benchmark for holistic comprehension beyond entity-centric matching. Extensive experiments demonstrate that CausalRAG2 consistently outperforms competitive graph-based RAG baselines across multiple datasets and evaluation metrics. Our work establishes a principled foundation for structured, scalable, and causally grounded RAG systems. Our code and HolisQA benchmark are available at https://github.com/Pwnb/CausalRAG2.

Abstract

Agentic Retrieval

Explore similar work

CausalRAG2: Hierarchical Causal Knowledge Graph Design for RAG

PathRouter: Aligning Rewards with Retrieval Quality in Agentic Graph Retrieval-Augmented Generation

Why Neighborhoods Matter: Traversal Context and Provenance in Agentic GraphRAG