Pith. sign in

REVIEW 3 major objections 3 minor 64 references

Spreading activation over an automatically built knowledge graph improves multi-hop question answering, delivering up to a 39% absolute gain over naive RAG when combined with chain-of-thought retrieval.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 15:41 UTC pith:6B4OC523

load-bearing objection Plausible training-free SA retriever over an auto-built KG, but the 25–39% gains are in-sample: hyperparameters are tuned on the same 100-question test sets, so the empirical claim isn't established yet. the 3 major comments →

arxiv 2512.15922 v3 pith:6B4OC523 submitted 2025-12-17 cs.AI

Leveraging Spreading Activation for Improved Document Retrieval in Knowledge-Graph-Based RAG Systems

Jovan Pavlovi\'c , Mikl\'os Kr\'esz , L\'aszl\'o Hajdu This is my paper
classification cs.AI
keywords spreading activationknowledge graphretrieval-augmented generationmulti-hop question answeringdocument retrievalchain-of-thoughtgraph-based RAGsmall open-weight models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard retrieval-augmented generation treats all retrieved text as equally relevant and tends to miss the 'bridge' documents that connect evidence across multiple sources. This paper argues that a classic cognitive-science algorithm—spreading activation—can fix that by searching a knowledge graph built automatically from the corpus. The paper claims that on two standard multi-hop QA benchmarks, a single retrieval step using spreading activation matches or beats several training-free RAG baselines, and that plugging it into chain-of-thought iterative retrieval yields up to a 39% absolute improvement in answer correctness over naive RAG, all with small open-weight language models. The reason to care is that the method needs no fine-tuning and no LLM-guided graph traversal, so it points toward cheaper, more robust retrieval for grounded reasoning.

Core claim

The core discovery, stated on the paper's own terms, is that spreading activation over an automatically constructed knowledge graph is a more effective document retriever for multi-hop question answering than dense vector similarity alone. The pipeline builds a heterogeneous graph whose nodes are entities, entity descriptions, and document chunks, connected by 'describes' links and 'related_to' links extracted by a prompt-tuned LLM and weighted by cosine similarity from an off-the-shelf embedding model. At query time, the system fetches the top-k matching entity descriptions, expands their neighborhoods by n hops, runs a breadth-first spreading activation from the matched seed entities with

What carries the argument

The spreading activation algorithm over the entity subgraph: seed entities receive activation 1, activation propagates breadth-first along 'related_to' edges scaled by cosine-similarity weights and capped at 1, and entities whose final activation exceeds a threshold are selected; document chunks describing activated entities (plus high-weight relationship texts) become the context fed to the generator. This converts a lexical query match into a set of associatively related entities, which is what lets the system surface bridge documents that never mention the query terms.

Load-bearing premise

The method assumes the automatically extracted knowledge graph is faithful—that entities are correctly identified and merged and that cosine-similarity edge weights reflect how strongly entities are really related; if the graph misrepresents the corpus, spreading activation will route retrieval to the wrong documents and the reported gains will not survive.

What would settle it

Construct the graph for a sample of multi-hop questions, record retrieval output, then corrupt the graph by removing a random 10% of related_to edges that are not in the gold evidence paths and rerun the same questions. If correctness stays roughly flat, the spreading activation and its edges are not the source of the gains; if it drops materially, the graph edges are load-bearing.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • As a single-step retriever, SA-RAG outperforms or matches several training-free RAG baselines on the two benchmarks examined, including iterative and decomposition-based approaches.
  • When used as a plug-and-play retriever inside chain-of-thought iterative retrieval, it yields the largest gains: up to 39% absolute correctness improvement over naive RAG and up to 10% over a graph-based baseline.
  • Because it requires no fine-tuning of retriever or generator and works with small open-weight language models, the approach is practical for settings with limited compute.
  • The method avoids LLM-guided graph traversal, reducing reliance on large models during retrieval and making the retrieval step more deterministic.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to expose the activation paths (which entities were activated and through which relations) as structured evidence, so the retrieval step doubles as an explanation of the reasoning chain rather than an opaque context dump.
  • The paper's own limitation suggests a natural experiment: fine-tune the embedding model on the relation triplets in the knowledge graph, then check whether the results where retrieval context seemed sufficient but reasoning lagged improve—if not, the ceiling is the generator, not the retriever.
  • Because the graph is built from overlapping chunks, entity reconciliation (same entity with multiple names) is a finite resource; an inference is that SA-RAG would benefit from an entity-resolution step before indexing, and that the current alias matching may set the upper bound on bridge recall.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes SA-RAG, a retrieval-augmented generation framework that uses spreading activation over an automatically constructed heterogeneous knowledge graph to retrieve documents for multi-hop question answering. The indexing stage uses a prompt-tuned LLM to extract entities, descriptions, and relations; at query time, top-k entity descriptions seed a subgraph, edge weights are set by embedding cosine similarity, and a BFS-based spreading activation algorithm identifies activated entities whose associated document chunks and relation descriptions are passed to an LLM generator. Experiments on MuSiQue and 2WikiMultiHopQA (100 questions each) compare SA-RAG to Naive RAG, CoT-RAG, query decomposition, and HippoRAG 2, reporting that SA-RAG combined with CoT retrieval improves correctness by 25–39% absolute over Naive RAG. The authors also release code and claim the method works with small open-weight LLMs.

Significance. If the empirical claims are validated, the paper makes a useful contribution: it demonstrates a training-free, plug-and-play retrieval module that uses a classical graph algorithm rather than LLM-guided traversal, and it ships reproducible code. The method is clearly specified and the use of spreading activation on an automatically built KG is a plausible and underexplored direction. However, the headline improvements are currently supported only by an in-sample evaluation with a small number of questions and without statistical safeguards, so the magnitude of the claimed gains is not yet established.

major comments (3)
  1. [§5.4–5.5, Table 1] The central empirical claim (25–39% absolute improvement) is based on hyperparameters chosen on the same 100-question test sets. The paper states that k, n, c, τa, τd, τr were set to values that 'yielded the best results' on these benchmarks. With seven tuned parameters and n=100, the reported gains are in-sample estimates and may include a large selection-bias component. The absence of a validation split, confidence intervals, or significance tests makes the headline numbers uninterpretable. Please report results on a held-out split (or a nested validation procedure) and provide standard errors or bootstrap intervals.
  2. [§4, Table 1] The manual Correctness metric is load-bearing for the main comparison (e.g., '75' vs '56' on MuSiQue with phi4), but the protocol is not described. There is no statement of who performed the annotation, whether it was blinded to system identity, or how inter-annotator agreement was measured. Without this, the correctness numbers are not reproducible. Please provide the annotation instructions, a blinded protocol, and agreement statistics; alternatively, report only EM/F1 or a validated automatic metric.
  3. [§3.1, §6] The evaluation uses only 100 randomly selected questions per benchmark. While the authors acknowledge this in §6 as a limitation, the combination of a small sample and the in-sample tuning described above means the paper's contribution is not yet empirically supported. A more extensive evaluation or, at minimum, a carefully designed validation protocol is needed before the 25–39% claim can be accepted.
minor comments (3)
  1. [Algorithm 1] The activation update uses min(1, value + weight·value), but the threshold τa is set to 0.5. This is fine, but the cap at 1.0 makes the threshold effectively a value in (0,1]; please clarify whether the cap is intended to prevent unbounded activation and why τa is not defined relative to this cap.
  2. [§5.4] The parameter values k=3 vs k=10 and n=4 vs n=3 for the two benchmarks are reported without explanation. It would be helpful to show how sensitive the final results are to these choices, especially given the in-sample tuning concern.
  3. [Appendix B] The prompts are detailed, which is good, but several listings have formatting artifacts (e.g., stray '``' and line breaks). Please clean up the appendix display so the prompts can be copied directly.

Circularity Check

2 steps flagged

Headline 25–39% gains are in-sample: retrieval hyperparameters are tuned on the same 100-question test sets that produce the reported Table 1.

specific steps
  1. fitted input called prediction [Section 5.4 (Subgraph fetching); Section 4 (Results, Table 1)]
    "We set the values of k and n to 3 and 4, respectively, for the MuSiQue experiments, and to 10 and 3 for the 2WikiMultiHopQA experiments, as these parameter choices yielded the best results."

    The subgraph-fetching parameters k and n are explicitly chosen to yield the best results on the same 100-question benchmarks whose results are later reported in Table 1. The subsequent claim that SA-RAG+CoT gives 25–39% absolute improvements over Naive RAG is therefore a restatement of the optimized configuration, not an out-of-sample prediction. Without a validation split, the headline improvements are in-sample selections rather than independent evidence.

  2. fitted input called prediction [Section 5.5 (Spreading activation and document retrieval); Section 4 (Results, Table 1)]
    "In our experiments, we found that setting c=0.4 provided the best results in terms of the recall/precision tradeoff. ... The values of τa, τd, and τr that produced the best performance in our experiments are 0.5, 0.45, and 0.5, respectively."

    The activation scaling factor c and the three thresholds τa, τd, τr are selected as the values that produce the best performance on the exact benchmarks that are then used as evidence for the method. The reported improvements (e.g., '25% to 39% relative to Naive RAG') are optimized in-sample quantities, not unbiased estimates from a held-out evaluation. The fitting step is acknowledged in the text but the results section presents the selected configuration as the method's achieved performance.

full rationale

The algorithmic pipeline itself is not circular: spreading activation is a standard, externally defined algorithm; the knowledge graph is constructed from the corpus with LLM prompts; and no load-bearing self-citations or imported uniqueness theorems appear. The circularity is confined to the empirical support for the headline claim. The retrieval hyperparameters (k, n, c, τa, τd, τr) are tuned on the same 100-question test sets from which Table 1 is derived, and the paper explicitly states that these values 'yielded the best results' / 'produced the best performance.' Therefore the claimed 25–39% absolute improvements over Naive RAG are in-sample maxima, not independent predictions. This is a genuine fitted-input-called-prediction pattern. The severity is partial rather than total: the method could still be genuinely effective, and a validation-set-based or pre-registered evaluation would resolve the issue, but the current evidence for the central quantitative claim is statistically forced by the tuning procedure.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new theoretical entities are postulated; the graph node types (entities, entity descriptions, document chunks, related_to edges) are architectural components. The central claim depends on unverified LLM extraction quality, embedding-based edge weights, and the representativeness of a 100-question evaluation.

free parameters (7)
  • k (number of seed entity descriptions) = 3 (MuSiQue), 10 (2WikiMultiHopQA)
    Section 5.4: 'We set the values of k and n to 3 and 4, respectively, for the MuSiQue experiments, and to 10 and 3 for the 2WikiMultiHopQA experiments, as these parameter choices yielded the best results.' Tuned per benchmark.
  • n (hop neighborhood expansion) = 4 (MuSiQue), 3 (2WikiMultiHopQA)
    Same sentence in Section 5.4; tuned per benchmark for best results.
  • c (linear edge-weight rescaling factor) = 0.4
    Section 5.5: 'setting c=0.4 provided the best results in terms of the recall/precision tradeoff.'
  • tau_a (activation threshold) = 0.5
    Section 5.5: 'The values of tau_a, tau_d, and tau_r that produced the best performance in our experiments are 0.5, 0.45, and 0.5.'
  • tau_d (document pruning threshold) = 0.45
    Section 5.5, same sentence as tau_a.
  • tau_r (relation relevance threshold) = 0.5
    Section 5.5, same sentence as tau_a.
  • chunk size / overlap = 500 words / 200 words
    Section 5.3: word-based chunking with 500-word chunks and 200-word overlap; a hand-chosen design decision affecting graph granularity.
axioms (5)
  • domain assumption LLM-extracted entities, descriptions, and relations faithfully represent the corpus content
    Section 5.3 indexing relies on prompt-tuned LLM extraction; if extraction is noisy or incomplete, graph retrieval cannot recover missing evidence.
  • domain assumption Cosine similarity between query/entity-description embeddings identifies the correct seed entities, and cosine similarity of relation links captures relation strength
    Sections 5.4-5.5 seed the graph and weight edges using an off-the-shelf embedding model; authors admit in Section 6 that this adaptation is simple and unoptimized.
  • domain assumption MuSiQue and 2WikiMultiHopQA, each sampled to 100 random questions, are representative enough to support general claims
    Section 3.1 restricts to 100 questions per benchmark due to computational cost; no seed or subset details are given.
  • domain assumption Word-based chunking at 500 words with 200-word overlap preserves the atomic facts needed for multi-hop reasoning
    Section 5.3; chunk granularity affects which entities co-occur and which documents link to which entities.
  • standard math The probability identity in Section 5.1 is a valid modeling decomposition of the generation process
    Section 5.1 writes p(a*|q,KB)=sum_D p(a*|q,D)p(D|q,KB); this is stated as the formal objective but is not used in the rest of the method.

pith-pipeline@v1.3.0-alltime-deepseek · 16574 in / 15087 out tokens · 150283 ms · 2026-08-03T15:41:09.346628+00:00 · methodology

0 comments
read the original abstract

Despite initial successes and a variety of architectures, retrieval-augmented generation systems still struggle to reliably retrieve and connect the multi-step evidence required for complicated reasoning tasks. Most of the standard RAG frameworks regard all retrieved information as equally reliable, overlooking the varying credibility and interconnected nature of large textual corpora. GraphRAG approaches offer potential improvement to RAG systems by integrating knowledge graphs, which structure information into nodes and edges, capture entity relationships, and enable multi-step logical traversal. However, GraphRAG is not always an ideal solution, as it depends on high-quality graph representations of the corpus. Such representations usually rely on manually curated knowledge graphs, which are costly to construct and update, or on automated graph-construction pipelines that are often unreliable. Moreover, systems following this paradigm typically use large language models to guide graph traversal and evidence retrieval. In this paper, we propose a novel RAG framework that uses a spreading activation algorithm to retrieve information from a corpus of documents connected by an automatically constructed heterogeneous knowledge graph. This approach reduces reliance on semantic knowledge graphs, which are often incomplete due to information loss during information extraction, avoids LLM-guided graph traversal, and improves performance on multi-hop question answering. Experiments show that our method achieves better or comparable performance to several state-of-the-art RAG methods and can be integrated as a plug-and-play module with different iterative RAG pipelines. When combined with chain-of-thought iterative retrieval, it yields up to a 39% absolute improvement in answer correctness over naive RAG, while achieving these results with small open-weight language models.

Figures

Figures reproduced from arXiv: 2512.15922 by Jovan Pavlovi\'c, L\'aszl\'o Hajdu, Mikl\'os Kr\'esz.

Figure 1
Figure 1. Figure 1: High-level overview of methodology 5.3 Indexing We begin the indexing phase by performing word-based chunking, by splitting input documents into chunks of 500 words with an overlap of 200 words between consecutive chunks. Each chunk is passed to the embedding model and then to a prompt-tuned LLM, which is instructed to extract entities, relations, and entity descriptions from the text. We construct a knowl… view at source ↗
Figure 2
Figure 2. Figure 2: Knowledge graph creation during the indexing phase. The figure shows an example graph constructed from [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Subgraph fetching step: Orange nodes represent the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Spreading activation on the subgraph fetched for the query [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of applying a linear normalization factor to the edge weights of the fetched subgraph on spreading [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 21 linked inside Pith

  1. [1]

    Retrieval augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. InInternational conference on machine learning, pages 3929–3938. PMLR, 2020

    2020

  2. [2]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    2020

  3. [3]

    Retrieval-augmented generation for ai-generated content: A survey.arXiv preprint arXiv:2402.19473, 2024

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, Jie Jiang, and Bin Cui. Retrieval-augmented generation for ai-generated content: A survey.arXiv preprint arXiv:2402.19473, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    A survey on rag meeting llms: Towards retrieval-augmented large language models

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6491–6501, 2024

    2024

  5. [5]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023. 12 APREPRINT- FEBRUARY9, 2026

    Pith/arXiv arXiv 2023

  6. [6]

    A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions.arXiv preprint arXiv:2410.12837, 2024

    Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions.arXiv preprint arXiv:2410.12837, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Graph retrieval-augmented generation: A survey.arXiv preprint arXiv:2408.08921, 2024

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey.arXiv preprint arXiv:2408.08921, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Retrieval-augmented generation with graphs (graphrag)

    Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Mahantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al. Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Neural, symbolic and neural-symbolic reasoning on knowledge graphs.AI Open, 2:14–35, 2021

    Jing Zhang, Bo Chen, Lingxi Zhang, Xirui Ke, and Haipeng Ding. Neural, symbolic and neural-symbolic reasoning on knowledge graphs.AI Open, 2:14–35, 2021

    2021

  10. [10]

    Knowledge graphs, large language models, and hallucinations: An nlp perspective.Journal of Web Semantics, 85:100844, 2025

    Ernests Lavrinovics, Russa Biswas, Johannes Bjerva, and Katja Hose. Knowledge graphs, large language models, and hallucinations: An nlp perspective.Journal of Web Semantics, 85:100844, 2025

    2025

  11. [11]

    Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599, 2024

    Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599, 2024

    2024

  12. [12]

    End-to-end beam retrieval for multi-hop question answering.arXiv preprint arXiv:2308.08973, 2023

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Yong Liu, and Shen Huang. End-to-end beam retrieval for multi-hop question answering.arXiv preprint arXiv:2308.08973, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity.arXiv preprint arXiv:2403.14403, 2024

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity.arXiv preprint arXiv:2403.14403, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Knowledge graph prompting for multi-document question answering

    Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr. Knowledge graph prompting for multi-document question answering. InProceedings of the AAAI conference on artificial intelligence, volume 38 (17), pages 19206–19214, 2024

    2024

  15. [15]

    Hipporag: Neurobiologically inspired long-term memory for large language models.Advances in Neural Information Processing Systems, 37:59532–59569, 2024

    Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. Hipporag: Neurobiologically inspired long-term memory for large language models.Advances in Neural Information Processing Systems, 37:59532–59569, 2024

    2024

  16. [16]

    Agentic retrieval-augmented generation: A survey on agentic rag.arXiv preprint arXiv:2501.09136, 2025

    Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Agentic retrieval-augmented generation: A survey on agentic rag.arXiv preprint arXiv:2501.09136, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Application of spreading activation techniques in information retrieval.Artificial Intelligence Review, 11(6):453–482, 1997

    Fabio Crestani. Application of spreading activation techniques in information retrieval.Artificial Intelligence Review, 11(6):453–482, 1997

    1997

  18. [18]

    Natural language processlng using spreading activation and lateral inhibition

    Jordan Pollack and David Waltz. Natural language processlng using spreading activation and lateral inhibition. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 4, 1982

    1982

  19. [19]

    Word sense disambiguation with spreading activation networks generated from thesauri

    George Tsatsaronis, Michalis Vazirgiannis, and Ion Androutsopoulos. Word sense disambiguation with spreading activation networks generated from thesauri. InIJCAI, volume 27, pages 223–252. Hyderabad, 2007

    2007

  20. [20]

    Hitoshi Kono, Ren Katayama, Yusaku Takakuwa, Wen Wen, and Tsuyoshi Suzuki. Activation and spreading sequence for spreading activation policy selection method in transfer reinforcement learning.International Journal of Advanced Computer Science and Applications, 10(12), 2019

    2019

  21. [21]

    Kg-infused rag: Augmenting corpus- based rag with external knowledge graphs.arXiv preprint arXiv:2506.09542, 2025

    Dingjun Wu, Yukun Yan, Zhenghao Liu, Zhiyuan Liu, and Maosong Sun. Kg-infused rag: Augmenting corpus- based rag with external knowledge graphs.arXiv preprint arXiv:2506.09542, 2025

    arXiv 2025

  22. [22]

    Multi-hop question answering.Foundations and Trends® in Information Retrieval, 17(5):457–586, 2024

    Vaibhav Mavi, Anubhav Jangra, Adam Jatowt, et al. Multi-hop question answering.Foundations and Trends® in Information Retrieval, 17(5):457–586, 2024

    2024

  23. [23]

    Query2doc: Query expansion with large language models.arXiv preprint arXiv:2303.07678, 2023

    Liang Wang, Nan Yang, and Furu Wei. Query2doc: Query expansion with large language models.arXiv preprint arXiv:2303.07678, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Precise zero-shot dense retrieval without relevance labels

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Precise zero-shot dense retrieval without relevance labels. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762–1777, 2023

    2023

  25. [25]

    Query rewriting in retrieval-augmented large language models

    Xinbei Ma, Yeyun Gong, Pengcheng He, Nan Duan, et al. Query rewriting in retrieval-augmented large language models. InThe 2023 Conference on Empirical Methods in Natural Language Processing, 2023

    2023

  26. [26]

    Decomposing complex questions makes multi-hop qa easier and more interpretable.arXiv preprint arXiv:2110.13472, 2021

    Ruiliu Fu, Han Wang, Xuejun Zhang, Jun Zhou, and Yonghong Yan. Decomposing complex questions makes multi-hop qa easier and more interpretable.arXiv preprint arXiv:2110.13472, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.arXiv preprint arXiv:2212.10509, 2022

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.arXiv preprint arXiv:2212.10509, 2022. 13 APREPRINT- FEBRUARY9, 2026

    Pith/arXiv arXiv 2022

  28. [28]

    Tree of clarifications: Answering ambiguous questions with retrieval-augmented large language models

    Gangwoo Kim, Sungdong Kim, Byeongguk Jeon, Joonsuk Park, and Jaewoo Kang. Tree of clarifications: Answering ambiguous questions with retrieval-augmented large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 996–1009, 2023

    2023

  29. [29]

    Self-knowledge guided retrieval augmentation for large language models.arXiv preprint arXiv:2310.05002, 2023

    Yile Wang, Peng Li, Maosong Sun, and Yang Liu. Self-knowledge guided retrieval augmentation for large language models.arXiv preprint arXiv:2310.05002, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Self-rag: Learning to retrieve, generate, and critique through self-reflection.arXiv preprint arXiv:2310.11511, 2023

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection.arXiv preprint arXiv:2310.11511, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Corrective retrieval augmented generation.arXiv e-prints, pages arXiv–2401, 2024

    Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling. Corrective retrieval augmented generation.arXiv e-prints, pages arXiv–2401, 2024

    2024

  32. [32]

    From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Lightrag: Simple and fast retrieval-augmented generation.arXiv preprint arXiv:2410.05779, 2024

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation.arXiv preprint arXiv:2410.05779, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. Advances in Neural Information Processing Systems, 37:132876–132907, 2024

    2024

  35. [35]

    From rag to memory: Non-parametric continual learning for large language models.arXiv preprint arXiv:2502.14802, 2025

    Bernal Jiménez Gutiérrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. From rag to memory: Non-parametric continual learning for large language models.arXiv preprint arXiv:2502.14802, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Word concepts: A theory and simulation of some basic semantic capabilities.Behavioral science, 12(5):410–430, 1967

    M Ross Quillian. Word concepts: A theory and simulation of some basic semantic capabilities.Behavioral science, 12(5):410–430, 1967

    1967

  37. [37]

    Retrieval time from semantic memory.Journal of verbal learning and verbal behavior, 8(2):240–247, 1969

    Allan M Collins and M Ross Quillian. Retrieval time from semantic memory.Journal of verbal learning and verbal behavior, 8(2):240–247, 1969

    1969

  38. [38]

    A spreading-activation theory of semantic processing.Psychological review, 82(6):407, 1975

    Allan M Collins and Elizabeth F Loftus. A spreading-activation theory of semantic processing.Psychological review, 82(6):407, 1975

    1975

  39. [39]

    A spreading activation theory of memory.Journal of verbal learning and verbal behavior, 22(3):261–295, 1983

    John R Anderson. A spreading activation theory of memory.Journal of verbal learning and verbal behavior, 22(3):261–295, 1983

    1983

  40. [40]

    On the use of spreading activation methods in automatic information

    Gerard Salton and Chris Buckley. On the use of spreading activation methods in automatic information. In Proceedings of the 11th annual international ACM SIGIR conference on Research and development in information retrieval, pages 147–160, 1988

    1988

  41. [41]

    Pure spreading activation is pointless

    Michael R Berthold, Ulrik Brandes, Tobias Kötter, Martin Mader, Uwe Nagel, and Kilian Thiel. Pure spreading activation is pointless. InProceedings of the 18th ACM conference on Information and knowledge management, pages 1915–1918, 2009

    1915

  42. [42]

    Discovering Latent Information By Spreading Activation Algorithm for Document Retrieval

    Vuong M Ngo. Discovering latent information by spreading activation algorithm for document retrieval.Ngo, Vuong M.“Discovering Latent Information By Spreading Activation Algorithm for Document Retrieval. ” Academy and Industry Research Collaboration Center, January 31, 2014. https://doi. org/10.5121/ijaia. 2014.5102., 2014

    arXiv 2014

  43. [43]

    Adapting spreading activation techniques towards a new approach to content-based recommender systems

    Yolanda Blanco-Fernández, Martín López-Nores, and José J Pazos-Arias. Adapting spreading activation techniques towards a new approach to content-based recommender systems. InIntelligent interactive multimedia systems and services, pages 1–11. Springer, 2010

    2010

  44. [44]

    Context aware personalized content recommendation using ontology based spreading activation.International Journal of Information Technology, 10(2):133–138, 2018

    Sachin Papneja, Kapil Sharma, and Nitesh Khilwani. Context aware personalized content recommendation using ontology based spreading activation.International Journal of Information Technology, 10(2):133–138, 2018

    2018

  45. [45]

    Musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    2022

  46. [46]

    Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps.arXiv preprint arXiv:2011.01060, 2020

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps.arXiv preprint arXiv:2011.01060, 2020

    Pith/arXiv arXiv 2011

  47. [47]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christo- pher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    Pith/arXiv arXiv 2018

  48. [48]

    Understanding dataset design choices for multi-hop reasoning.arXiv preprint arXiv:1904.12106, 2019

    Jifan Chen and Greg Durrett. Understanding dataset design choices for multi-hop reasoning.arXiv preprint arXiv:1904.12106, 2019. 14 APREPRINT- FEBRUARY9, 2026

    Pith/arXiv arXiv 1904

  49. [49]

    Is multihop qa in dire condition? measuring and reducing disconnected reasoning.arXiv preprint arXiv:2005.00789, 2020

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Is multihop qa in dire condition? measuring and reducing disconnected reasoning.arXiv preprint arXiv:2005.00789, 2020

    Pith/arXiv arXiv 2005

  50. [50]

    provided_context

    Sewon Min, Eric Wallace, Sameer Singh, Matt Gardner, Hannaneh Hajishirzi, and Luke Zettlemoyer. Composi- tional questions do not necessitate multi-hop reasoning.arXiv preprint arXiv:1906.02900, 2019. A Implementation Details of Baseline RAG Methodologies In this appendix, we provide a detailed description of the implementation details for the baseline RAG...

    Pith/arXiv arXiv 1906

  51. [51]

    John Phan,

    **Use only information explicitly present in the original query.** * Do **not** introduce new entities, definitions, or clarifying questions that the user did not ask. * If the query says “John Phan,” do **not** ask “Who is John Phan?” because that is not needed to ,→resolve the chain

  52. [52]

    * Prefer single-hop, factual questions (entity→attribute, entity→location, entity→relation)

    **Decompose by inference steps, not by wording.** * Each subquestion should retrieve **one missing fact** or **resolve one reference** needed by a later ,→subquestion. * Prefer single-hop, factual questions (entity→attribute, entity→location, entity→relation)

  53. [53]

    * Stop decomposing when the main query becomes answerable

    **Keep only necessary subquestions.** * If a fact is already given in the original query, do **not** restate it as a question. * Stop decomposing when the main query becomes answerable

  54. [54]

    * Number the subquestions in the order they should be executed

    **Preserve dependency order.** * Later subquestions may refer to entities/answers from earlier ones. * Number the subquestions in the order they should be executed

  55. [55]

    original_question

    **Output format (JSON):** ‘‘‘json { "original_question": "<the user question>", "subquestions": [ {"id": 1, "question": "..."}, {"id": 2, "question": "..."}, {"id": 3, "question": "..."} ] } ‘‘‘

  56. [56]

    The Argentine PGA Championship record holder has won how many tournaments worldwide?

    **If the query is already single-hop, return it as one subquestion** in the same format. **Worked example** Input question: > “The Argentine PGA Championship record holder has won how many tournaments worldwide?” Decomposition: ‘‘‘json { "original_question": "The Argentine PGA Championship record holder has won how many tournaments ,→worldwide?", "subques...

    2026

  57. [57]

    born in,

    **Identifing Provided Information:** - **Input Text:** Reading the paragraph about Jack Parsons; Noting key phrases like "born in," "died ,→in," "cofounder," and the association of JPL with Caltech. - **Entity List:** Recognize the entities: Jack Parsons, Los Angeles, Pasadena, Jet Propulsion ,→Laboratory, California Institute of Technology, Aerojet Engin...

  58. [58]

    Jack Parsons was ,→born in Los Angeles

    **Extracting Explicit Relationships:** - Locating explicit statements in the text that connect two entities. For example, "Jack Parsons was ,→born in Los Angeles" and "died in Pasadena." - Identifing that Jack Parsons "co-founded" both the Jet Propulsion Laboratory and the Aerojet ,→Engineering Corporation. - Noting the phrase linking JPL to Caltech: "Jet...

  59. [59]

    triples": [ [

    **Performing Pronoun Resolution:** - Ensuring that any pronouns referring to named entities are replaced by their proper names; In this ,→case, no pronoun resolution was required since the entities are explicitly named. ### Response 19 APREPRINT- FEBRUARY9, 2026 { "triples": [ ["Jack Parsons", "born in", "Los Angeles"], ["Jack Parsons", "died in", "Pasade...

    2026

  60. [60]

    **Short knowledge paragraphs**, each describing specific entities and facts

  61. [61]

    ## Task description Your goal is to **understand and reason through the question** using only the information provided

    A **list of key relationships** between these entities. ## Task description Your goal is to **understand and reason through the question** using only the information provided. To achieve this you should:

  62. [62]

    **Identify** only the facts and relationships that are directly relevant to answering the question

  63. [63]

    - Write in full sentences

    **Paraphrase and weave** those facts into one concise, coherent paragraph called ‘provided_context ‘. - Write in full sentences. - Avoid listing facts one per line

  64. [64]

    answer_possible

    Decide whether you can answer the question with the information in your summary: - If **yes**, set ‘"answer_possible": true‘ and put your answer in ‘final_answer ‘. - If **no**, set ‘"answer_possible": false‘ and craft a specific follow-up question in ‘ ,→additional_question‘ that would help fill the gap in provided information ### Output Format (JSON) ‘‘...