Rethinking Retrieval-Augmented Generation as a Cooperative Decision-Making Problem

Lichang Song; Ting Long; Yi Chang

cs.CLFeb 21, 2026

Rethinking Retrieval-Augmented Generation as a Cooperative Decision-Making Problem

Authors: Lichang Song, Ting Long, Yi Chang

Organizations: Jilin University

Abstract

Retrieval-Augmented Generation (RAG) has demonstrated strong effectiveness in knowledge-intensive tasks by grounding language generation in external evidence. Despite its success, many existing RAG systems are built based on a ranking-centric, asymmetric dependency paradigm, where the generation quality of the generator is highly dependent on reranking results of the reranker. To overcome this limitation, we propose Cooperative Retrieval-Augmented Generation (CoRAG), a framework that treats the reranker and the generator as peer decision-makers rather than being connected through an asymmetric dependency pipeline. By jointly optimizing their behaviors toward a shared task objective, the reranker and generator are encouraged to cooperate, ensuring that document reranking and generation work in concert to improve the final response. Experimental results demonstrate good generalization and improved generation stability of CoRAG, even when the model is trained on only around 10K PopQA samples. Our model released in https://github.com/CoderrrSong/CoRAG.

Figures & tables

Figure 1: Comparison with previous works. Previous works assume an asymmetric dependency between the reranker and the generator, whereas CoRAG treats them as equal participants optimized under a shared task-oriented reward.

Figure 2: CoRAG overview. The reranker and generator cooperate to generate responses. The task-oriented reward derived from response guides GRPO-aligned training of the reranker and GRPO optimization of the generator.

Dataset	Train	Test
PopQA	12,868	1,399
TriviaQA	78,785	11,313
NaturalQuestions	79,168	3,610
ASQA	4,353	948
2WikiMultiHopQA	167,454	12,576

Table 1: Dataset statistics.

Method	PopQA	TriviaQA	NQ	2WikiMultiHopQA	ASQA
Method	(acc)	(acc)	(acc)	(acc)	(em)	(pre)	(rec)
Baselines w/o Retrieval
Vanilla Zero-shot Prompting
ChatGPT	29.3	74.3	–	–	35.3	–	–
Llama-3-Instruct ${}_{8\text{B}}$	22.8	69.4	46.6	45.6	30.6	–	–
Llama-3-Instruct ${}_{70\text{B}}$	28.9	80.6	57.9	57.5	39.1	–	–

Table 2: Overall results of our method and baselines on five benchmarks. Baseline results are reported in InstructRAG Wei et al. (2024) . - indicates the results are not reported or not applicable. The best and second-best performances are highlighted in bold and with an underline, respectively. Notably, our model is trained only on PopQA Mallen et al. (2023) , while all other datasets are used exclusively for evaluation.

	PopQA	NQ	TriviaQA	2Wiki
RTrain	66.19	62.71	75.57	49.36
GTrain	66.54	63.37	76.25	51.53
RGReplace	65.83	63.21	74.97	49.72
GReplace	66.26	62.46	75.00	50.00
RReplace	70.12	69.77	80.76	56.05
CoRAG (Ours)	71.26	72.49	81.00	58.26

Table 3: Ablation study.

Figure 3: Cross-validation results with different reranker and generator combinations (Top-3 setting).

Figure 4: Impact of the document number.

Metric	Llama-3-Instruct ${}_{8\text{B}}$	InstructRAG	CoRAG
human-eval
pass@1	64.45	55.85	63.96
pass@10	83.53	76.82	85.97
human-eval+
pass@1	56.58	49.26	57.43
pass@10	77.43	70.12	79.26

Table 4: The cross-task evaluation. The InstructRAG in the table is trained on the Pop dataset. WTQ represents WikiTable Questions. The best and second-best performances are highlighted in bold and with an underline.

Method	InstructRAG	RetRobust	CoRAG
Pattern-based	66.19	56.68	71.26
LlaMa	75.63	25.52	89.21
GPT	60.83	54.90	65.90
DeepSeek	59.69	34.10	64.26
Qwen	60.69	57.26	66.12
Average	64.60	45.69	71.35

Table 5: Evaluation with LLM-as-a-judge (PopQA).

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Prompt : You are an expert evaluator for large model responses. Your core task is to determine whether the large model’s response points to any of the correct answers. Please conduct the evaluation based on the following information: 1. Question: {question} 2. List of correct answers: {answers} 3. Large model’s response: {response} Evaluation Rules: 1. If the core information and key content of the large model’s response point to any answer in the correct answer list (semantic consistency is acceptable, no need for word-for-word matching), please output "yes". 2. If the large model’s response deviates from all correct answers, contains obvious errors, or fails to effectively respond to the question, please output "no". 3. You only need to output a single word: either "yes" or "no", without any additional redundant content.

Output : yes

Appendix

Table 6: Prompt of LLM-as-a-Judge

Prompt : You are tasked with answering the given question by analyzing a set of documents. Please follow this STRICT TWO-STEP PROCESS: — ### STEP 1: Document Analysis For each document: - First, extract potentially relevant information from the original document. This includes facts, names, dates, or statements that may relate to the question, even if the connection is not immediately obvious. - Then, explain the reason for the information extraction in your previous step based on the question. (e.g., how the document addresses the question’s focus). ### STEP 2: Final Answer - Summarize your answer to the question based on the analysis above. - If none of the documents are helpful or relevant, answer based on your own general knowledge. In that case, clearly state that you are doing so. — Use the following format for your response: ### Step 1: Document Analysis Document 1: - Extraction: … - Explanation: … Document 2: - Extraction: … - Explanation: … ### Step 2: Final Answer Well-supported answer, based on the relevant documents. If no relevant documents, answer based on general knowledge and say so explicitly. — EXAMPLE: Question: Who is the author of The Mahdi? ### Step 1: Document Analysis Document 1: - Extraction: ’Mahdi’ is a thriller novel by Philip Nicholson written in 1981 under the identity of a. J. Quinnell … - Explanation: This document directly states that the author of ’Mahdi’ is Philip Nicholson. Upon re-examining the question "Who is the author of The Mahdi?", the document mentions the corresponding book title and provides the author information requested. Document 2: - Extraction: ’The Mahdi’ was published by Philip Nicholson in 1981 under the pen name A.J. Quinnell, establishing his presence in the thriller genre with a novel known for its gripping plot and enduring popularity… - Explanation: This document directly states that ’The Mahdi’ was published by Philip Nicholson. Upon re-examining the question "Who is the author of The Mahdi?", the document mentions the corresponding book title and provides the author information requested. ### Step 2: Final Answer Based on Documents 1 and 2, The answer to question ’Who is the author of The Mahdi?’ is A.J. Quinnell, a pseudonym of Philip Nicholson. — Now it is your turn to analyze the following documents and answer the given question by following the two-step process. {context} {question}

Appendix

Table 7: Prompt of Inference

	PopQA		NQ
	Ranked	Shuffle	Ranked	Shuffle
Llama-3-Instruct ${}_{8\text{B}}$	65.83	65.68	64.65	63.68
InstructRAG-Gen	66.69	66.33	64.45	61.49
CoRAG-Gen	71.26	71.05	72.46	72.10

Appendix

Table 8: Performance comparison under shuffled vs. ranked document order on PopQA and NQ datasets

Retrieval-Augmented Generation

Retrieval-augmented generation connects language models to external knowledge so answers can use current, domain-specific evidence.

See all 766 papers

Guaranteeing Knowledge Integration with Joint Decoding for Retrieval-Augmented Generation
Apr 9, 2026Zhengyi Zhao, Shubo Zhang, Zezhong Wang +7Hallucination in RAG Retrieval-Augmented Generation
The Chinese University of Hong Kong · Ministry of Education Key Laboratory of High Confidence Software Technologies, CUHK · University of International Relations +2
Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation
May 2, 2026Peiyang Liu, Qiang Yan, Ziqiang Cui +3Retrieval-Augmented Generation LLM Hallucination Mitigation
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
Latent Abstraction for Retrieval-Augmented Generation
Apr 20, 2026Ha Lan N. T, Minh-Anh Nguyen, Dung D. LeRetrieval-Augmented Generation Adaptive RAG
Center for AI Research, VinUniversity, Vietnam

Apr 9, 2026J/K move · Enter open · S save

Guaranteeing Knowledge Integration with Joint Decoding for Retrieval-Augmented Generation

Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Yuxi Zhang+6

The Chinese University of Hong Kong · Ministry of Education Key Laboratory of High Confidence Software Technologies, CUHK · University of International Relations+2

Hallucination in RAG Retrieval-Augmented Generation Knowledge Conflicts in Language Models

Open paper AskPDF arXiv

Retrieval-Augmented Generation (RAG) significantly enhances Large Language Models (LLMs) by providing access to external knowledge. However, current research primarily focuses on retrieval quality, often overlooking the critical ''integration bottleneck'': even when relevant documents are retrieved, LLMs frequently fail to utilize them effectively due to conflicts with their internal parametric knowledge. In this paper, we argue that implicitly resolving this conflict in a single generation pass is suboptimal. We introduce GuarantRAG, a framework that explicitly decouples reasoning from evidence integration. First, we generate an ''Inner-Answer'' based solely on parametric knowledge to capture the model's reasoning flow. Second, to guarantee faithful evidence extraction, we generate a ''Refer-Answer'' using a novel Contrastive DPO objective. This objective treats the parametric Inner-Answer as a negative constraint and the retrieved documents as positive ground truth, forcing the model to suppress internal hallucinations in favor of external evidence during this phase. Finally, rather than naive concatenation or using the DPO trained model directly, we propose a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Answer with the factual precision of the Refer-Answer at the token level. Experiments on five QA benchmarks demonstrate that GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.

Abstract

Retrieval-Augmented Generation

Explore similar work

Guaranteeing Knowledge Integration with Joint Decoding for Retrieval-Augmented Generation

Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation

Latent Abstraction for Retrieval-Augmented Generation