Paper deep dive
D$^2$F-ReAG: Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation
Jiaoyang Li, Junhao Ruan, Shengwei Tang, Kaiyan Chang, Zhengtao Yu, Tong Xiao, Jingbo Zhu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/9/2026, 3:49:07 AM
Summary
The paper introduces D2F-ReAG, a framework for multi-hop reasoning-augmented generation that dynamically decomposes complex questions into sub-questions based on the reliability of root-level reasoning. It employs an adaptive filtering mechanism to verify sub-question answers and integrates them to refine the final answer, outperforming existing graph-based and prompt-based RAG methods on benchmarks like HotpotQA, 2WikiMultiHopQA, and MuSiQue.
Entities (12)
Relation Signals (11)
D2F-ReAG → evaluatedon → HotpotQA
confidence 95% · Experiments on three multi-hop benchmarks... HotpotQA
D2F-ReAG → evaluatedon → 2WikiMultiHopQA
confidence 95% · Experiments on three multi-hop benchmarks... 2WikiMultiHopQA
D2F-ReAG → evaluatedon → MuSiQue
confidence 95% · Experiments on three multi-hop benchmarks... MuSiQue
D2F-ReAG → usesmechanism → Dynamic Decomposition
confidence 95% · D2F-ReAG... adaptively controls reasoning depth... question is logically decomposed into sub-questions
D2F-ReAG → usesmechanism → Reliability Filtering
confidence 95% · D2F-ReAG... adaptively controls reasoning depth by judging the reliability of the root-level reasoning
LLM-Acc → measures → Semantic Correctness
confidence 90% · LLM-Acc... assesses whether the prediction is semantically equivalent to the reference
Str-Acc → measures → Lexical Correctness
confidence 90% · Str-Acc measures lexical correctness by checking whether the prediction exactly matches the ground-truth answer
D2F-ReAG → outperforms → LogicRAG
confidence 90% · D2F-ReAG consistently outperforms all baselines... Against the strongest prompt-based baseline LogicRAG, it yields up to 5.4-point Str-Acc and 6.4-point LLM-Acc improvements
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) often generate inaccurate answers due to their reliance on static internal knowledge. Retrieval-augmented generation (RAG) addresses this limitation by integrating external knowledge and excelling at single-hop queries. However, it struggles with multi-hop questions that require cross-document reasoning. Existing methods, such as graph structured RAG or question decomposition, often lack dynamic decomposition and effective filtering, which leads to lower efficiency and accuracy. To overcome these limitations, we propose Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation (D2F-ReAG), a novel paradigm that adaptively controls reasoning depth by judging the reliability of the root-level reasoning. If the root reasoning is reliable, the model directly generates the answer. Otherwise, the question is logically decomposed into sub-questions, and the verified reasoning derived from these sub-questions is used to refine the root reasoning. Experiments on three multi-hop benchmarks demonstrate the effectiveness of our method in handling complex multi-hop questions.
Tags
Links
- Source: https://arxiv.org/abs/2608.04444v1
- Canonical: https://arxiv.org/abs/2608.04444v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
30,818 characters extracted from source content.
Expand or collapse full text
D 2 F-ReAG: Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation Jiaoyang Li 1 , Junhao Ruan 1 , Shengwei Tang 1 , Saihan Chen 1 , Kaiyan Chang 1 , Zhengtao Yu 2 , Tong Xiao 1,3 , and Jingbo Zhu 1,3 1 Northeastern University, Shenyang, China 2 Kunming University of Science and Technology, Kunming, China 3 NiuTrans Research, Shenyang, China Abstract. Large language models (LLMs) often generate inaccurate answers due to their reliance on static internal knowledge. Retrieval- augmented generation (RAG) addresses this limitation by integrating external knowledge and excelling at single-hop queries. However, it strug- gles with multi-hop questions that require cross-document reasoning. Ex- isting methods, such as graph structured RAG or question decomposi- tion, often lack dynamic decomposition and effective filtering, which leads to lower efficiency and accuracy. To overcome these limitations, we pro- pose Dynamic Decomposition and Filtering for Multi-Hop Reasoning- Augmented Generation (D 2 F-ReAG), a novel paradigm that adaptively controls reasoning depth by judging the reliability of the root-level rea- soning. If the root reasoning is reliable, the model directly generates the answer. Otherwise, the question is logically decomposed into sub- questions, and the verified reasoning derived from these sub-questions is used to refine the root reasoning. Experiments on three multi-hop bench- marks demonstrate the effectiveness of our method in handling complex multi-hop questions. Keywords: Multi-Hop Reasoning· Dynamic Decomposition· Reasoning- Augmented Generation 1 Introduction Large Language Models (LLMs) have demonstrated strong capabilities in lan- guage understanding and generation across a wide range of tasks [1]. Despite these advances, they still suffer from inherent limitations such as outdated para- metric knowledge and hallucinations, which often lead to factually incorrect or unsupported outputs and undermine their reliability in knowledge-intensive sce- narios [11]. Retrieval-augmented Generation (RAG) mitigates these issues by incorporating external knowledge sources into the generation process, thereby improving factuality and reducing the model’s reliance on its static internal knowledge [13]. arXiv:2608.04444v1 [cs.CL] 5 Aug 2026 2J. Li et al. For simple factual questions, retrievers can often locate the necessary ev- idence within a single retrieval step, making the process both efficient and straightforward [17]. In contrast, multi-hop reasoning questions require connect- ing multiple pieces of evidence that are typically scattered across different docu- ments[3]. To address this challenge, Graph-based RAG (Graph RAG) organizes facts through pre-built graph structures, enabling effective evidence traversal and multi-hop retrieval across interconnected knowledge units [6]. However, such graphs often suffer from inherent limitations: they may be incomplete due to the difficulty of capturing all relevant relations, are costly to construct and maintain at scale, and become less effective when the underlying knowledge is frequently updated or evolving [14]. Recent advances such as LogicRAG [2] reduce the reliance on pre-constructed graphs by decomposing complex queries into sub-questions and iteratively com- pressing retrieved evidence into a document-level memory, enabling multi-hop reasoning without explicit graph structures. However, such compression cannot fully filter out redundant or erroneous content, allowing noise to accumulate across iterations and gradually mislead subsequent reasoning [23]. Moreover, their query decomposition is often rigid and lacks adaptive control over gran- ularity, so that simple questions tend to be over-decomposed while genuinely complex ones are not decomposed deeply enough, ultimately degrading reason- ing performance [12]. To overcome these limitations, we propose D 2 F-ReAG, a multi-hop reason- ing framework that performs on-demand question decomposition driven by rea- soning reliability and augments generation with verified sub-question reasoning paths. Concretely, D 2 F-ReAG continuously assesses the model’s confidence in its current reasoning, and triggers further decomposition only when the rea- soning over the current question is judged unreliable. Once a sub-question is solved with sufficient confidence and deemed relevant to the query, its reason- ing trace is propagated upward to update and guide the reasoning of the root question, so that accurate intermediate results are effectively integrated rather than discarded or compressed away. To prevent over-reasoning and unnecessary computational overhead, D 2 F-ReAG also adopts an early-stopping strategy that terminates the process as soon as the root question can be answered reliably. In summary, our contributions are: – We propose D 2 F-ReAG, a novel framework that dynamically decomposes complex questions into a sequence of logical sub-questions and progressively augments the reasoning process of the root question with the correct reason- ing paths derived from these sub-questions, ensuring that accurate interme- diate results are effectively propagated back to guide the final answer rather than being lost in lossy memory compression. – We dynamically and adaptively balance the decomposition depth accord- ing to question complexity through a confidence-driven control mechanism, enabling the framework to perform deeper decomposition only when nec- essary while avoiding redundant decomposition for simpler questions, thus Title Suppressed Due to Excessive Length3 achieving truly on-demand decomposition that adapts to varying reasoning difficulty. – We leverage reliable reasoning chains derived from sub-questions to itera- tively update and correct the root reasoning process, which effectively identi- fies and fixes intermediate errors, thereby reducing error accumulation across reasoning steps and improving the overall correctness of the final answer. 2 Related Work Graph-based methods organize external knowledge into structured graph repre- sentations to better support multi-hop retrieval and cross-document reasoning. GraphRAG [6] performs hierarchical community search by jointly leveraging local and global queries, allowing retrieval to operate at varying levels of gran- ularity. LightRAG [7] further improves large-scale retrieval efficiency through a two-stage, graph-augmented indexing pipeline that balances coverage and cost. RAPTOR [16] constructs hierarchical summaries over the corpus to enable multi-granular retrieval at different levels of abstraction, thereby accommodating queries of diverse specificity. HippoRAG [8] and its extension HippoRAG 2 [9] ad- ditionally enhance long-context coherence by ranking nodes with PageRank [15] and integrating paragraph-level memory, which helps preserve contextual conti- nuity across retrieval steps. Despite these advances, graph-based methods gen- erally rely on pre-constructed structures that are costly to build and maintain, and may become outdated when the underlying knowledge evolves. Prompt-based methods, in contrast, interleave reasoning and retrieval through carefully designed prompts without modifying model parameters, offering a more flexible and lightweight alternative. ReAct [21] alternates between reasoning and retrieval in a step-by-step manner, enabling the model to dynamically decide when to retrieve based on its current reasoning state and observed evidence. ChainRAG [22] decomposes complex questions into sub-questions guided by prompts, and subsequently performs iterative sentence-graph retrieval to cap- ture both local fine-grained details and global document-level structures. Logi- cRAG [2] similarly decomposes questions into logical components and aggregates evidence across multiple reasoning steps, leveraging logical relationships among sub-questions to enhance retrieval coherence and reduce redundancy. SiGIR [4] refines the reasoning process iteratively to progressively approach the final an- swer; however, it decomposes questions purely by atomicity and lacks explicit correctness verification during retrieval, which may allow erroneous intermediate results to propagate and accumulate across reasoning steps, ultimately degrading the quality of the final answer. 3 Method We propose D 2 F-ReAG, a dynamic framework for multi-hop reasoning. As shown in Figure 1, D 2 F-ReAG begins with (I) Retrieval & Generation, where first re- trieves top-k documents to generate question’s reasoning. In (I) Judge Relia- bility, a judge model evaluates the reliability of the reasoning. If reliable, the 4J. Li et al. Which film was released first, Aas Ka Panchhi or Phoolwari? Original QuestionRetrieverdocuments Judge model solved final answer Decompose I.Retrieval & Generation I.Judge Reliability Which film was released first, Aas Ka Panchhi or Phoolwari? When was Aas Ka Panchhi released? When was Phoolwari released? Which year is earlier? #1 and #2? solved & relevant No information..... may be 1933. Aas Ka Panchhi is a 1961..... Panchhi, which is 1961. Decompose Which year is earlier? 1961 and 1933? Rewrite 1933. Aas Ka Panchhi is a 1961..... Panchhi, which is 1961. To determine which film was released first between Aas Ka Panchhi......Therefore, Phoolwari (1933) 1946 released before Aas Ka Panchhi (1955). 1961 Reasoning When was Phoolwari released? What is another name for Phoolwari? What is Phoolwari? According to..... is The Bower. Rewrite What is The Bower released? Acrodd..... is 1946. Accord..... is 1946. solved & relevant correct & irrelevant I.Decomposition & Rewriting correct & irrelevant IV.ReAG When was #4 released? Stop Fig. 1. The D 2 F-ReAG framework consists of four stages: (I) Retrieval & Generation, which retrieves relevant documents to generate a reasoning. (I) Judge Reliability, where an LLM evaluates reasoning reliability against a threshold. (I) Decomposition & Rewriting, which decomposes and rewrites the question when reliability is insuf- ficient. (IV) ReAG (Reasoning-Augments generation), which integrates reliable and relevant sub-question reasoning to update the root reasoning. The process stops once the root question is answered reliably. question is solved; otherwise, (I) Decomposition & Rewriting decomposes the question into sub-questions and rewrites them to retrieve more relevant evidence. Finally, (IV) ReAG (Reasoning-Augmented Generation) augments root reason- ing with reliable, relevant reasoning traces of sub-questions. D 2 F-ReAG stops once the root reasoning score exceeds a threshold. 3.1 Retrieval & Generation For the root question q root and each sub-question q i , we retrieve the top-k most relevant documents D(q) from the corpus C using a dense retriever, and then generate the corresponding reasoning process r(q) conditioned on the retrieved evidence. The retrieved documents serve as external knowledge that grounds the model’s reasoning, mitigating hallucinations and providing factual support for downstream judgment. Formally, for any question q ∈q root ,q 1 ,...,q n , r(q) = Generator q,D(q) ,(1) where Generator denotes the model that generates detailed reasoning based on relevant documents. By explicitly conditioning on D(q), the generator is encour- aged to perform evidence-aware reasoning rather than relying solely on its para- Title Suppressed Due to Excessive Length5 metric memory, which is particularly important for knowledge-intensive multi- hop questions. 3.2 Judge Reliability After generating the reasoning for the root or sub-questions, we judge its re- liability using an LLM-based scoring mechanism. Rather than treating every generated reasoning chain as trustworthy, we introduce an explicit verification step that assesses whether the reasoning is logically coherent, factually consis- tent with the retrieved documents, and sufficient to answer the question. We denote the reliability score of the reasoning process as s r (q), which is calculated by LLM: s r (q) = Score(q,r(q)),(2) where r(q) represents the reasoning process, and Score(q,r(q)) is the reasoning score provided by the LLM according to our predefined rubric. The rubric jointly considers multiple dimensions, including evidence grounding, logical consistency, and answer completeness, so that the resulting score reflects a holistic assessment of reasoning quality. The reasoning score s r (q) ∈ [0, 10] measures reasoning reliability (higher is better). We compare it to a threshold θ (set to 7 in our experiments): if s r (q) > θ, we consider the question solved. Otherwise, D 2 F-ReAG thinks the reasoning unreliable and further decomposes the question for continued reason- ing (Section 3.3). This adaptive mechanism allows the system to allocate more computational effort only to questions that genuinely require deeper exploration, while terminating early on questions that are already well-addressed. Judge = ( Solved,if s r (q) > θ Decompose, if s r (q)≤ θ. (3) If the current question is the root question and is deemed solved, the LLM gen- erates the final answer based on the corresponding reasoning process, ensuring that the answer is directly traceable to verified reasoning steps: a root = Answer(r q root ).(4) If a sub-question is reliably solved, we use its relevant sub-questions to update the root reasoning in ReAG (Section 3.4), so that the verified intermediate con- clusions can progressively enrich the global reasoning context. 3.3 Decomposition & Rewriting When the root question or any sub-question remains unsolved, we use prompt engineering to logically decompose it into smaller, more tractable sub-questions, each focusing on a single reasoning hop or a narrower information need. This 6J. Li et al. divide-and-conquer strategy reduces the cognitive load on the generator and makes retrieval more targeted. sub(q) = Decompose(q),(5) where sub(q) is the set of sub-questions obtained by logically decomposing the question q. D 2 F-ReAG first solves each sub-question in order and judges the reliability of the reasoning. If reliable, we rewrite the related other sub-questions q ∈ S i based on the correct reasoning r q , so that subsequent sub-questions can leverage previously verified intermediate results and avoid redundant or contradictory reasoning paths. S ′ i = Rewrite(q,r q )| q ∈S i ,(6) where r q is the reasoning to the current question, and q i is the related sub- question. Through this rewriting step, ambiguous references and missing entities in the original sub-questions can be resolved using the newly acquired evidence, leading to higher-quality retrieval in the next iteration. 3.4 ReAG When a sub-question is solved reliably, we check its relevance to the root question and use its reasoning to update the root reasoning if relevant. This relevance check prevents tangential or off-topic sub-question results from contaminating the global reasoning state. Check(q i ,q) = ( relevant irrelevant (7) where Check(q i ,q) denotes the relevance of q i to the root question q. If a sub-question’s reasoning is reliable and relevant, it is used to update the root reasoning. The update operation integrates newly verified evidence and intermediate conclusions into the existing root reasoning, gradually building a more complete and coherent chain of thought toward the final answer. Formally, this can be written as: r ′ (q root ) = Update(r(q root ),r(q)),(8) where r ′ (q root ) is the updated reasoning process of the root problem q, r(q) is the reasoning process of the sub-question q i , and Update(r(q root ),r(q)) updates r(q root ) by incorporating r(q). Once the root reasoning is reliable, we stop processing the remaining sub- questions and obtain the final answer to avoid over-reasoning. This early-stopping strategy not only reduces unnecessary computational overhead but also prevents the introduction of noise from over-decomposition, which could otherwise dilute the focus of the root reasoning and degrade answer accuracy. Title Suppressed Due to Excessive Length7 Table 1. Performance comparison on HotpotQA, 2WikiMultiHopQA, and MuSiQue using Str-Acc and LLM-Acc.† denotes our reimplementation; the others are from LogicRAG. Best results are in bold, and second-best results are underlined. TypeMethod HotpotQA 2Wiki MuSiQue Str. LLM Str. LLM Str. LLM Zero-shot Llama3 (8B) 17.1 11.1 22.3 4.7 2.3 2.0 Llama3 (13B) 23.7 20.1 33.8 15.4 6.4 6.0 GPT-3.5-Turbo 31.5 35.4 24.0 22.0 7.9 10.9 GPT-4o-Mini 38.7 36.3 26.4 24.3 17.6 14.0 Graph RAPTOR48.1 57.8 47.7 45.9 25.2 29.1 GraphRAG39.6 45.2 46.3 43.3 16.5 19.4 LightRAG47.8 57.7 43.1 36.3 18.1 19.4 HippoRAG53.5 56.6 47.2 47.2 24.9 30.1 HippoRAG2 56.7 61.9 50.0 47.1 27.0 32.6 Prompt-RAG VanillaRAG43.2 53.1 43.0 42.0 20.3 23.6 ReAct † 54.2 56.5 54.9 50.8 28.8 31.9 ChainRAG † 52.1 56.6 68.9 66.231.033.8 LogicRAG † 54.2 62.565.3 62.6 29.6 36.5 OursD 2 F-ReAG55.863.4 70.3 68.9 32.2 37.9 4 Experiments 4.1 Dataset and Metrics We evaluate on three standard multi-hop reasoning benchmarks: HotpotQA [20], MuSiQue [18] and 2WikiMultiHopQA [10], covering diverse cross-document and multi-hop reasoning. Following HippoRAG 2 [9], we use the same retrieval corpus and randomly sample 1,000 questions from each validation set for evaluation, ensuring fair comparison. Following LogicRAG [2], we adopt two metrics: Str-Acc and LLM-Acc. Str-Acc measures lexical correctness by checking whether the prediction ex- actly matches the ground-truth answer after standard normalization. LLM-Acc instead employs a strong LLM as an automatic judge to assess whether the pre- diction is semantically equivalent to the reference, tolerating paraphrasing and surface-form variations. Reporting both metrics enables a balanced evaluation of strict lexical matching and flexible semantic correctness. 4.2 Baselines We compare D 2 F-ReAG with three categories of baselines that cover repre- sentative paradigms in multi-hop question answering. Zero-shot baselines di- rectly prompt LLMs without any external retrieval, including LLaMA3 (8B) and LLaMA3 (13B) [5], as well as gpt-3.5-turbo and gpt-4o-mini [1], which serve to reflect the intrinsic reasoning ability of LLMs. Graph RAG baselines incorpo- rate pre-constructed graph structures to support multi-hop retrieval, including RAPTOR [16], GraphRAG [6], LightRAG [7], HippoRAG [8], and HippoRAG 2 8J. Li et al. [9]. Prompt-based RAG baselines instead interleave reasoning and retrieval through prompt design without relying on explicit graph construction, including ReAct [21], ChainRAG [22], and LogicRAG [2]. 4.3 Implementation Details. For fair comparison, all methods adopt the same experimental configuration. We use sentence-transformers/all-MiniLM-L6-v2 [19] as the unified embedding model for dense retrieval. The number of retrieved passages (top-k) is fixed to 3 across all methods. For answer generation, we employ gpt-4o-mini [1] as the backbone LLM, ensuring that performance differences stem from the retrieval and reasoning strategies rather than the underlying generator. All experiments are conducted on a single NVIDIA RTX 3090 GPU. 4.4 Main Results Table 1 reports results on three multi-hop reasoning benchmarks (HotpotQA, 2WikiMultiHopQA, and MuSiQue). D 2 F-ReAG achieves the best or near-best performance on both Str-Acc and LLM-Acc across all three datasets, demon- strating consistent superiority over diverse baselines. Zero-shot LLMs benefit from stronger backbones, with gpt-4o-mini clearly surpassing LLaMA3 (8B/13B), yet even the best zero-shot model trails retrieval- based methods, underscoring the need for external evidence in multi-hop QA. Graph-based RAG methods (RAPTOR, GraphRAG, LightRAG, HippoRAG, HippoRAG 2) improve performance by structuring context into explicit graphs, with HippoRAG 2 being the strongest (e.g., 56.7 Str-Acc on HotpotQA), though their gains shrink on harder datasets like MuSiQue where graphs miss multi-hop relations. Prompt-based RAG methods interleave retrieval and reasoning: Chain- RAG performs well on 2Wiki via sub-question decomposition with sentence- graph retrieval, and LogicRAG attains the best baseline LLM-Acc on HotpotQA (62.5) through logical decomposition, but both suffer from fixed decomposition and noise accumulated over iterations. D 2 F-ReAG consistently outperforms all baselines on 2Wiki (70.3 / 68.9) and MuSiQue (32.2 / 37.9), and achieves the highest LLM-Acc on HotpotQA (63.4), with the largest gains on 2WikiMultiHopQA where deep multi-hop reasoning is most needed. Against the strongest prompt-based baseline LogicRAG, it yields up to 5.4-point Str-Acc and 6.4-point LLM-Acc improvements on 2Wiki, show- ing that on-demand decomposition and reliability-guided reasoning mitigate the over- or under-decomposition of fixed-depth methods. Joint gains on both met- rics further indicate that the recovered reasoning paths are not only lexically aligned with gold answers but also semantically more faithful. We report the best scores across runs, and provide a case study in Appendix 6 to illustrate how D 2 F-ReAG corrects intermediate errors via reliable sub-question reasoning. Title Suppressed Due to Excessive Length9 MuSiQue2WikiHotpotQA 0 0.2 0.4 0.6 0.8 0.59 0.63 0.8 0.41 0.37 0.2 rate w/o decomposition w/ decomposition Fig. 2. Solved with vs. without decomposition on three benchmarks. Table 2. Accuracy and efficiency comparison on 2WikiMultiHopQA (with vs. without decomposition). Method Str Acc LLM Acc Avg. Time (s) Avg. Tokens ReAct54.9 50.813.9311287 LogicRAG 65.3 62.615.351998 Ours (No D) 71.9 71.09.492267 Ours (+D) 67.3 65.073.5413321 Ours (Avg.) 70.3 68.932.46057 4.5 Decomposition & Efficiency Comparison Figure 2 shows that a large proportion of questions can be solved without de- composition, demonstrating the necessity of dynamic decomposition to efficiently handle both simple and complex questions. Since ChainRAG also uses a pre-built sentence graph, we compare latency and efficiency with ReAct and LogicRAG. Table 2 shows that always decomposing and judging reliability greatly increases time and token cost. In contrast, on- demand decomposition skips unnecessary steps, speeding up easy cases while still decomposing for hard ones. While our multi-step judgments can take longer on some difficult examples, Figure 2 confirms these long cases are relatively few. Overall, D 2 F-ReAG balances between performance and efficiency and better matches human-like reasoning. 5 Conclusion Existing RAG methods still struggle with multi-hop questions, primarily due to their fixed decomposition strategies and limited ability to filter out erroneous 10J. Li et al. or irrelevant information during reasoning. To address these issues, we propose D 2 F-ReAG (Dynamic Decomposition and Filtering for Multi-Hop Reasoning- Augmented Generation), a framework that adaptively decomposes questions only when necessary and selectively filters intermediate reasoning according to its reliability. By leveraging verified reasoning from reliable sub-questions to progressively refine the root-level generation, D 2 F-ReAG effectively suppresses error propagation and produces more faithful reasoning chains, achieving strong performance across challenging multi-hop benchmarks. 6 Case Study We present a case study comparing the reasoning behaviors of D 2 F-ReAG and LogicRAG. Question: Which film has the director who died later, 45 Calibre Echo or Bons Baisers De Hong Kong? Gold Answer: Bons Baisers De Hong Kong Failure Case of LogicRAG. Model Answer: 45 Calibre Echo (Incorrect) Rounds: 1 Retrieved Contexts. LogicRAG retrieves three biography-style passages: – Bruce M. Mitchell: includes his death date (September 26, 1952), supporting reasoning about 45 Calibre Echo. – John Edward Bruce / Fred Bradley (rower): unrelated to either queried film. Dependency Analysis (sorted). Death date of Bruce M. Mitchell; Death date of Yvan Chiffre. Round 1. Query: Death date of Bruce M. Mitchell. can_answer: true. Understanding: Bruce M. Mitchell (director of 45 Calibre Echo) died on Septem- ber 26, 1952; the death date of Yvan Chiffre is not retrieved. LogicRAG never- theless concludes 45 Calibre Echo due to the missing comparison target. Correct Case of D 2 F-ReAG. Model Answer: Bons Baisers De Hong Kong (Correct) Root Rationale: The retrieved context names the directors (Bruce M. Mitchell and Yvan Chiffre) but does not provide their death dates. Judging reliability: 0.4 Decompose & Rewriting: – Who directed 45 Calibre Echo and when did the director die? – Who directed Bons Baisers De Hong Kong and when did the director die? – Which year is later? Title Suppressed Due to Excessive Length11 Filtering for Multi-Hop Reasoning-Augmented Generation. Subq 1: Director of 45 Calibre Echo is Bruce M. Mitchell. Reliability: 0.8 Subq 2: Director of Bons Baisers De Hong Kong is Yvan Chiffre. Reliability: 0.8 Updating Root Reasoning and Rewriting. Rewritten sub-questions: When did Bruce M. Mitchell die? When did Yvan Chiffre die? Filtering for Multi-Hop Reasoning-Augmented Generation. Yvan Chiffre died on 1990-01-01; Bruce M. Mitchell died on 1988-02-19. Updating Root Reasoning and Rewriting. Rewritten sub-question 3: Which year is later, 1988 or 1990? Subq 3: 1990. Reliability: 1.0 Final Answer: Bons Baisers De Hong Kong References 1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2. Chen, S., Zhou, C., Yuan, Z., Zhang, Q., Cui, Z., Chen, H., Xiao, Y., Cao, J., Huang, X.: You don’t need pre-built graphs for rag: Retrieval augmented generation with adaptive reasoning structures. arXiv preprint arXiv:2508.06105 (2025) 3. Cheng, R., Liu, J., Zheng, Y., Ni, F., Du, J., Mao, H., Zhang, F., Wang, B., Hao, J.: DualRAG: A dual-process approach to integrate reasoning and retrieval for multi-hop question answering. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2025) 4. Chu, Z., Fan, H., Chen, J., Wang, Q., Yang, M., Liang, J., Wang, Z., Li, H., Tang, G., Liu, M., Qin, B.: Self-critique guided iterative reasoning for multi-hop question answering. In: Findings of the Association for Computational Linguistics: ACL 2025 (Jul 2025) 5. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024) 6. Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., Metropolitansky, D., Ness, R.O., Larson, J.: From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024) 7. Guo, Z., Xia, L., Yu, Y., Ao, T., Huang, C.: LightRAG: Simple and fast retrieval- augmented generation. In: Findings of the Association for Computational Linguis- tics: EMNLP 2025 (2025) 8. Gutiérrez, B.J., Shu, Y., Gu, Y., Yasunaga, M., Su, Y.: Hipporag: Neurobio- logically inspired long-term memory for large language models. In: The Thirty- eighth Annual Conference on Neural Information Processing Systems (2024), https://openreview.net/forum?id=hkujvAPVsg 9. Gutiérrez, B.J., Shu, Y., Qi, W., Zhou, S., Su, Y.: From rag to mem- ory: Non-parametric continual learning for large language models (2025), https://arxiv.org/abs/2502.14802 10. Ho, X., Duong Nguyen, A.K., Sugawara, S., Aizawa, A.: Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. In: Proceedings of the 28th International Conference on Computational Linguistics (2020) 12J. Li et al. 11. Jiang, C., Qi, B., Hong, X., Fu, D., Cheng, Y., Meng, F., Yu, M., Zhou, B., Zhou, J.: On large language models’ hallucination with regard to known facts. In: Pro- ceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) (2024) 12. Kim, G.I., Kim, J.W., Jang, B.: UniRAG: A unified RAG framework for knowledge- intensive queries with decomposition, break-down reasoning, and iterative rewrit- ing. In: Findings of the Association for Computational Linguistics: EMNLP 2025 (Nov 2025) 13. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020) 14. Li, J., Ruan, J., Tang, S., Chen, S., Chang, K., Ge, Y., Xiao, T., Zhu, J.: Subqrag: Sub-question driven dynamic graph rag. arXiv preprint arXiv:2510.07718 (2025) 15. Page, L., Brin, S., Motwani, R., Winograd, T.: The pagerank citation ranking: Bringing order to the web. Tech. rep., Stanford infolab (1999) 16. Sarthi, P., Abdullah, S., Tuli, A., Khanna, S., Goldie, A., Manning, C.D.: Rap- tor: Recursive abstractive processing for tree-organized retrieval. In: International Conference on Learning Representations (ICLR) (2024) 17. Shi, Z., Zhang, S., Sun, W., Gao, S., Ren, P., Chen, Z., Ren, Z.: Generate-then- ground in retrieval-augmented generation for multi-hop question answering. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2024) 18. Trivedi, H., Balasubramanian, N., Khot, T., Sabharwal, A.: MuSiQue: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics (2022) 19. Wang, W., Bao, H., Huang, S., Dong, L., Wei, F.: MiniLMv2: Multi-head self- attention relation distillation for compressing pretrained transformers. In: Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 (Aug 2021) 20. Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W., Salakhutdinov, R., Manning, C.D.: HotpotQA: A dataset for diverse, explainable multi-hop question answering. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (2018) 21. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., Cao, Y.: ReAct: Synergizing reasoning and acting in language models. In: International Conference on Learning Representations (ICLR) (2023) 22. Zhu, R., Liu, X., Sun, Z., Wang, Y., Hu, W.: Mitigating lost-in-retrieval problems in retrieval augmented multi-hop question answering. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2025) 23. Zhuang, Z., Zhang, Z., Cheng, S., Yang, F., Liu, J., Huang, S., Lin, Q., Rajmohan, S., Zhang, D., Zhang, Q.: EfficientRAG: Efficient retriever for multi-hop question answering. In: Proceedings of the 2024 Conference on Empirical Methods in Nat- ural Language Processing (2024)