Organizations: School of Computer Science and Engineering, Southeast University, China · Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China · School of Law, Southeast University, China
Retrieval-Augmented Generation (RAG) systems have demonstrated remarkable effectiveness in mitigating hallucinations by incorporating external knowledge. However, the retrieval process inevitably introduces noise, posing significant challenges to RAG robustness. Fundamentally, noise robustness is a composite competency encompassing noise identification, retrieval filtering, and robust reasoning. Unfortunately, existing methods often fail to capture this holistic nature, typically restricted to single-dimensional enhancements or hampered by noise-induced optimization interference. To address these limitations, we decompose RAG robustness into three sub-tasks: knowledge boundary perception, retrieval verification, and robust reasoning. Accordingly, we propose ARMOR, an Adaptive Curriculum Meta-Learning for Noise Robust RAG Reasoning. Inspired by the teaching process, ARMOR introduces an adaptive task scheduler as a teacher to curate training plans according to the absolute learning progress of RAG generators (Student). Distinct from GRPO, ARMOR utilizes inter-group advantage estimation during learning to effectively mitigate mode collapse caused by limited output spaces. Experiments on three complex reasoning datasets (2WikiMultiHopQA, MusiQue, and HotPotQA) indicate that ARMOR achieves synergistic improvements in retrieval noise robustness, significantly outperforming baselines in output quality.
Figures & tables
Figure 1: The overall framework of ARMOR. ( Left ) The Adaptive Task Scheduler functions as a Teacher , dynamically prioritizing one of the three robustness sub-tasks (Knowledge Boundary Perception, Retrieval Verification, and Robust Reasoning) for each training batch based on the Absolute Learning Progress (ALP). ( Right ) The RAG Generator ( Student ) generates multiple rollouts to learn the selected task. ( Bottom Right ) Inter-Group Advantage Estimation computes reward advantages across the global batch scope rather than within groups, effectively mitigating mode collapse and stabilizing the decoupled training process.
Subset
2Wiki
MuSiQue
HotpotQA
Train Warm-Up
4,467
496
–
Train RL
7,716
857
–
Test
1,352
375
500
Validation
1,352
375
–
Table 1: Experiment dataset statistics.
Method
2WikiMultiHopQA
MusiQue
HotpotQA
Avg.
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Qwen-3-1.7B
No-RAG
37.72
37.21
32.99
3.73
2.96
0.00
24.20
22.44
15.67
21.88
20.87
16.22
Vanilla-RAG
47.63
44.67
38.46
24.00
27.43
17.33
45.07
45.72
38.33
38.90
39.27
31.37
CoT
39.64
17.13
11.39
8.27
6.27
1.33
22.47
18.03
13.60
23.46
13.81
8.77
CoN
63.61
13.13
4.73
29.33
12.36
5.33
54.53
22.65
17.07
49.16
16.05
9.04
Table 2: Performance comparison of ARMOR with baselines. We report Correctness (Cor.), F1 score, and Exact Match (EM) in percentage (%) using Qwen-3-1.7B and Llama-3.2-3B as backbones. The best results are highlighted in bold . Full results are provided in Appendix D .
Figure 2: Robustness analysis under controlled noise ratios (from 0% to 100%) with fixed retrieval size of K=5 .
Method
2WikiMultiHopQA
MusiQue
HotpotQA
Avg.
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Qwen-3-1.7B
ARMOR
79.14
79.22
77.22
54.67
57.22
46.67
57.73
54.86
58.00
63.85
63.77
60.63
w/o CRL -random
67.66
67.97
66.48
39.33
40.59
28.67
37.27
35.45
27.33
48.09
48.00
40.83
w/o CRL -seq
52.22
52.69
50.89
23.65
26.50
15.20
24.67
21.76
11.27
33.51
33.65
25.79
w/o CRL -outcome sup.
60.22
60.45
58.59
32.33
32.37
23.00
30.80
27.72
18.73
41.12
40.18
33.44
Table 3: Ablation analysis on the effectiveness of the Adaptive Curriculum Learning (CRL) strategy and Inter-Group Advantage Estimation (inter-group adv). The best results are highlighted in bold .
Figure 3: Task selection probabilities pn(τ) during the training.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: An example of the structured reasoning chain..
Figure 5: Prompt for generating the first-hop sub question in 2WikiMultiHopQA.
Figure 6: Prompt for generating multi-hop sub-questions (second-hop onward) in 2WikiMultiHopQA.
Figure 7: Prompt for reconstructing the sub question in MusiQue.
Figure 8: Prompt for reconstructing the sub answer in 2WikiMultiHopQA.
Figure 9: Prompt for reconstructing the thinking process in 2WikiMultiHopQA and MusiQue.
Figure 10: Prompt for golden document augmentation.
Datasets
Golden
Inconsequential
Irrelevant
Multi-Hop Question
2Wiki
21.87
30.03
48.10
MuSiQue
12.92
35.65
51.43
HotpotQA
17.14
33.89
48.97
Decomposed Sub-Question
2Wiki
25.06
29.17
45.77
Appendix
Table 4: Distribution of retrieved document types for multi-hop questions and their decomposed sub-questions across three datasets.
Figure 11: Prompt for direct answer.
Figure 12: Inference prompt of ARMOR.
Method
2WikiMultiHopQA
MusiQue
HotpotQA
Avg.
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Cor.
F1
EM
Qwen-3-0.6B
No-RAG
37.43
35.55
32.99
3.73
2.58
0.00
20.27
16.26
10.47
20.48
18.13
14.49
Vanilla-RAG
34.47
28.16
21.15
14.67
11.63
2.67
34.87
28.24
21.27
28.00
22.68
15.03
CoT
34.17
31.29
28.55
2.67
5.99
0.00
19.67
18.16
14.33
18.84
18.48
14.29
CoN
56.15
19.33
9.91
19.73
15.98
8.27
46.13
32.50
26.53
40.67
22.60
14.90
Appendix
Table 5: Performance comparing our method with baselines on three multi-hop QA datasets using Qwen-3-0.6B, Qwen-3-4B, Llama-3.2-1B as backbones. We report Correctness (Cor.), F1 score, and Exact Match (EM) in percentage (%). The best results are highlighted in bold .
Figure 13: Sensitivity analysis of rollout number ( N ) on training stability using Qwen-3-1.7B.
Figure 14: Impact of retrieval size ( K ) on generation correctness using Qwen-3-1.7B.
Question:
When was Ole Rynning’s father born?
Ground Truth:
May 30, 1778
ARMOR:
<think> To find out when Ole Rynning’s father was born, I first need to identify who Ole Rynning’s father is. </think> <sub-question> Who is Ole Rynning’s father? </sub-question> <search> True </search> <doc> … He was the son of the priest Jens Rynning (1778–1857) … </doc> <doc-type> golden </doc-type> <doc> … </doc> <doc-type> irrelevant </doc-type> … <sub-answer> Ole Rynning’s father is Jens Rynning. </sub-answer> <think> Now that I know Ole Rynning’s father is Jens Rynning, I need to find his date of birth. </think> <sub-question> When was Jens Rynning born? </sub-question> <search> True </search> <doc> Jens Rynning (May 30, 1778 – June 11, 1857) was a … </doc> <doc-type> golden </doc-type> <doc> Ole Rynning (April 4, 1809 – September 1838) was a … </doc> <doc-type> inconsequential </doc-type> … <sub-answer> Jens Rynning was born on May 30, 1778. </sub-answer> <main-answer> May 30, 1778. </main-answer>
Search-R1:
<think> Okay, I need to find out when Ole Rynning’s father was born. First, I should check if I have any existing information about Ole Rynning. I remember that Ole Rynning is a Danish footballer… </think> <search> Ole Rynning birth date </search> <information> … He was the son of the priest Jens Rynning (1778–1857) … </information> <search> Ole Rynning father birth date </search> <information> … He was born in Stavanger as a son of Ole Bergesen … </information> <answer> Jens Rynning </answer>
IRCoT:
After interleaving chain-of-thought reasoning with iterative retrieval, the QA reader generates the final answer: <doc> Jens Rynning (May 30, 1778 – June 11, 1857) was a … </doc> <doc> Ole Rynning (April 4, 1809 – September 1838) was a … </doc> <doc> … He was the son of the priest Jens Rynning (1778–1857) … </doc> <doc> … </doc> … <main-answer> April 4, 1809. </main-answer>
Appendix
Table 6: Comparative case study of ARMOR , Search-R1 , and IRCoT on multi-hop QA with noisy retrieval.
Figure 15: Prompt for the evaluation of correctness.
Figure 16: Comparison of correctness based on (1) Human and (2) GPT-5 Judgement. Legend entries such as “ 2Wiki (0.6)” denote results of using 2WikiMultiHopQA with 60% noise ratio.
Retrieval-Augmented Generation (RAG) was introduced to enhance the capabilities of Large Language Models (LLMs) beyond their encoded prior knowledge. This is achieved by providing LLMs with an external source of knowledge, which helps reduce factual hallucinations and enables access to new information not available during pretraining. However, inconsistent retrieved information can negatively affect LLM responses. The Retrieval-Augmented Generation Benchmark (RGB) was introduced to evaluate the robustness of RAG systems under such conditions. In this work, we use the RGB corpus to evaluate LLMs in four scenarios: noise robustness, information integration, negative rejection, and counterfactual robustness. We perform a comparative analysis between the RGB RAG baseline and GraphRAG, a knowledge graph based retrieval system. We test three GraphRAG customizations to improve robustness. Results show improvements over the RGB baseline and provide insights for designing more reliable RAG systems for real world scenarios.
Hazem Amamou, Stéphane Gagnon, Alan Davoust +1
Institut national de la recherche scientifique (INRS-EMT), Montréal, Québec, Canada · INRS-UQO Mixed Research Unit on Cybersecurity, Gatineau, Québec, Canada · Université du Québec en Outaouais, Gatineau, Québec, Canada
Standard Retrieval-Augmented Generation (RAG) systems predominantly rely on semantic relevance as a proxy for utility. However, this assumption collapses in realistic decision-making scenarios where user queries are laden with cognitive biases, such as false premises or confirmation bias. In such cases, maximizing relevance paradoxically promotes the retrieval of sycophantic evidence that reinforces hallucinations, a critical failure we term the ``Relevance-Robustness Gap''. To bridge this gap, we propose CoRM-RAG (Counterfactual Risk Minimization for RAG), a framework that aligns retrieval with decision safety rather than mere similarity. Grounded in causal intervention, we introduce a Cognitive Perturbation Protocol to simulate user biases during training, which is then distilled into a lightweight Evidence Critic. This scoring module learns to identify documents that possess sufficient evidential strength to steer the model toward correctness despite adversarial query perturbations. Extensive experiments on decision-making benchmarks demonstrate that CoRM-RAG significantly outperforms strong dense retrievers and LLM-based rerankers in adversarial settings, while enabling effective risk-aware abstention through reliable robustness scoring. Our code is available at https://github.com/PeiYangLiu/CoRM-RAG.git.
Peiyang Liu, Qiang Yan, Ziqiang Cui +3
National Engineering Research Center for Software Engineering, Peking University Beijing, China · PX Securities Shenzhen, China · City University of Hong Kong Hong Kong SAR, China +2
Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, effectively mitigating their inherent knowledge limitations. However, RAG remains vulnerable to poisoning attacks that manipulate retrieved texts to mislead model outputs. Existing defense mechanisms often lack theoretical robustness guarantees and perform unreliably when the LLM has limited knowledge of the retrieved content. In this work, we propose PRA-RAG, a provably robust retrieval aggregation algorithm designed to defend against poisoning attacks on retrieved texts. PRA-RAG samples multiple combinations of retrieved texts and utilizes geometric structures in the embedding space to identify a robust subset, from which a stable aggregated representation is derived. We provide theoretical bounds on the maximum impact of poisoned retrieved content and establish a quantitative measure of RAG's robustness. Experiments across multiple benchmarks and RAG architectures demonstrate that PRA-RAG reduces the attack success rate to as low as 1% while maintaining an accuracy of 71%, significantly outperforming representative state-of-the-art methods.
Xue Tan, Yi Zheng, Chang Huo +7
School of Computer Science, Fudan University, Shanghai, China · Institute of Big Data, Fudan University, Shanghai, China · Department of Computer Science, Worcester Polytechnic Institute, MA, USA