Pith. sign in

REVIEW 5 major objections 6 minor 5 cited by

Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that systematically tuning a modular knowledge-graph RAG pipeline—over chunk size, retriever, top-k, and prompts—delivers consistent but task-dependent gains over default settings on HotPotQA, TwoWikiMultiHop, and MuSiQue.

desk verdict Useful empirical measurements of TPE-based tuning in Cognee, but the 12-question holdout sets and missing statistical detail keep it a solid conditional, not a definitive result. read the letter →

arxiv 2505.24478 v1 pith:ZUBDPEMM submitted 2025-05-30 cs.AI cs.CL

Vasilije Markovic , Lazar Obradovic , Laszlo Hajdu , Jovan Pavlovic This is my paper
classification cs.AIcs.CL
keywords hyperparameteroptimizationknowledgegraphsretrieval-augmentedgenerationmulti-hopquestionansweringTree-structuredParzenEstimatorgraph-basedretrievalLLM-basedevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that systematic hyperparameter tuning of a knowledge-graph retrieval-augmented generation pipeline is not just possible but practically valuable. It studies the full pipeline—chunking, graph construction, retrieval, prompting, and evaluation—as one tunable system and optimizes six parameters with a Tree-structured Parzen Estimator on three multi-hop QA benchmarks. The authors report consistent gains over default settings across all nine dataset–metric combinations, while emphasizing that the gains vary by task and metric and that no single configuration wins everywhere. The payoff is a concrete argument that configuration-level choices alone can improve downstream question answering without architectural changes, and that evaluation-metric design is itself a bottleneck.

What carries the argument

The central object is the Cognee pipeline treated as a parameterized objective function, optimized by a Tree-structured Parzen Estimator over six configuration knobs: chunk size, retriever strategy, top-k, QA prompt template, graph-construction prompt, and whether document summaries are generated. Each trial runs the full extract–construct–retrieve–answer loop and returns one scalar score, either exact match, F1, or an LLM-graded correctness score. The main mechanism carrying the argument is the retriever choice: graph completion returns nodes and triplets formatted as structured text, while chunk retrieval passes raw text segments to the model, and this difference, combined with answer-format alignment, largely explains where gains appear.

What would settle it

Run the same TPE optimization over the full, unfiltered training and validation splits of HotPotQA, TwoWikiMultiHop, and MuSiQue. If optimized configurations no longer beat the defaults by comparable margins—or lose on some metrics—then the paper's central claim only holds for its cleaned subset.

Watch

Extended reading notes

Core claim

The central claim is that targeted, task-aware tuning of a modular graph-based RAG pipeline yields meaningful and repeatable answer-quality improvements. On the paper's filtered subsets, the optimized configuration beat the default on every benchmark–metric pair: relative training-set gains reached 1496% for HotPotQA exact match, 351% for MuSiQue F1, and between 62.8% and 71.2% for LLM-based correctness, with several default scores near zero because the default prompt produced conversational answers where benchmarks demand terse ones. Most gains persisted on held-out questions, though they shrank. The authors attribute the pattern to a few interacting levers—retriever type, chunk size, and prompt format—and conclude that systematic tuning is achievable and useful, but that metric choice and dataset characteristics strongly shape what counts as improvement.

Load-bearing premise

The load-bearing premise is that the authors' manual filter of each benchmark—keeping 24 training and 12 test questions judged to be well-formed and supported—leaves a representative slice, so gains measured on that slice would also appear on the full benchmarks.

Editorial extensions

If this is right

  • Pipeline configuration should be treated as part of the optimization problem in graph-RAG systems, not as a fixed default.
  • Task-specific tuning generalizes reasonably to unseen questions from the same benchmark, though with smaller margins.
  • Exact match, F1, and LLM-based correctness can rank configurations differently, so metric choice changes the conclusions of any tuning study.
  • Prompt style, especially answer concision, is a first-order determinant of lexical metrics like exact match.
  • A shared benchmark or leaderboard for graph-augmented RAG would help cross-system comparison and robust tuning conclusions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because all gains are measured on a small manually filtered subset of each benchmark, a natural test is to rerun the same optimizer on the full unfiltered sets; if the margins collapse, the finding is about clean examples rather than realistic noise.
  • The large prompt-format effect suggests part of the measured gain is alignment with the metric's lexical expectations rather than deeper reasoning; an answer-normalization layer before scoring would separate the two.
  • The task-specific best configurations imply that per-domain tuning portraits could be precomputed and reused, letting practitioners warm-start optimization for a new dataset instead of assuming transfer.
  • The summary-generation parameter was explored at only two levels, so continuous variation of summary granularity may interact with chunk size in ways this study does not yet resolve.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper studies systematic hyperparameter optimization of Cognee, a modular knowledge-graph construction and retrieval pipeline, on three multi-hop QA benchmarks (HotPotQA, TwoWikiMultiHop, MuSiQue). Using a Tree-structured Parzen Estimator, the authors tune chunk size, retriever type, top-k, QA and graph prompts, and a task-processing parameter, scoring configurations with exact match, F1, and DeepEval's LLM-based correctness. They report consistent but variable gains over Cognee's default configuration on training and held-out sets, and argue that configuration-level changes alone can improve downstream performance. The evaluation uses 24 training and 12 test manually filtered questions per dataset, with 50 optimization trials per benchmark-metric combination.

Significance. If the result holds, the paper provides useful empirical evidence that task-specific tuning of graph-based RAG pipelines matters and that default configurations can be far from optimal. The modular treatment of the Cognee pipeline and the use of multiple benchmarks and metrics are strengths, as is the explicit separation of training and hold-out sets. However, the significance is tempered by the very small manual-filtered evaluation sets, the absence of displayed confidence intervals or significance tests, the lack of a hold-out baseline comparison, and the potential confound between prompt-format alignment and graph-construction/retrieval improvements. The paper's central claim is plausible but currently not established at the level of statistical confidence the wording implies.

major comments (5)
  1. [Section 5, Table 3] The hold-out evaluation uses only 12 manually filtered questions per dataset, so each answer changes EM/F1 by about 8.3 percentage points. Section 5 states that bootstrap confidence intervals were computed, but no intervals appear in Table 3 or elsewhere, and no significance tests are reported. Moreover, Table 3 contains only optimized train and test scores; the claimed 'gains over the baseline' on the hold-out set are not shown because no baseline test-set scores are reported (the baseline values in Table 2 are training-set baselines). Without these numbers, the central claim that tuning produces consistent hold-out gains is not supported.
  2. [Section 5] The evaluation set is created by randomly sampling instances and then manually filtering out examples that are 'ungrammatical, ambiguous, mislabeled, or unsupported.' The criteria are subjective, and the paper gives no counts of how many or which examples were removed, no examples of filtered items, and no inter-annotator agreement. This is load-bearing because all training and test scores, and hence all relative gains, are computed on these filtered subsets. If the filtering removed hard or noisy instances, the observed gains may not transfer to the unfiltered benchmarks; the authors should quantify the filtering and demonstrate representativeness.
  3. [Section 6.1] The paper attributes the large exact-match gains to a mismatch between the default prompt style ('conversational') and the benchmarks' preferred answer style ('short and dry'). If this is correct, the headline improvements are largely a prompt-format alignment effect rather than an effect of KG construction or retrieval hyperparameters. Because the QA prompt template is one of the jointly tuned parameters, the experimental design does not isolate its contribution. An ablation that fixes the QA prompt and varies chunk size, retriever type, top-k, and task getter would be needed to support the claim that configuration-level changes to the graph-based interface drive the gains.
  4. [Section 4.2 and Section 5] The paper does not provide the exact parameter values, the LLM models and versions used for graph construction, answer generation, and DeepEval grading, the random seeds, or the best configurations found in each experiment. It also does not release code or configuration files. Section 4.2 gives only ranges for chunk size and top-k and qualitative descriptions of prompt templates and retriever types; Section 5 reports 50 trials and 30-minute runtimes but no trial-level detail. These omissions make the experiments impossible to reproduce or independently verify, which is a serious limitation for an empirical tuning study.
  5. [Section 6.3] The discussion acknowledges that the DeepEval LLM grader 'introduced noise' and that several near-verbatim answers received less than full credit. Because the optimizer maximizes this LLM-based correctness score, part of the training-set improvement may be an artifact of the grader rather than a genuine improvement in answer quality. The manuscript should report grader reliability (e.g., agreement with deterministic metrics) or show that correctness gains persist when the final configurations are re-scored with deterministic metrics.
minor comments (6)
  1. [Table 2] The relative gain for HotPotQA EM is reported as 1496.0%, but combining the displayed baseline (0.042) and optimized (0.667) values gives approximately 1488% (or about 1588% if the gain is computed as a ratio rather than a percentage increase), so the calculation should be reconciled.
  2. [Throughout] The benchmark name is spelled inconsistently as both 'Musique' and 'MuSiQue'; the latter matches the cited reference and should be used consistently.
  3. [Section 5] The sentence 'In addition to point estimates, we report confidence intervals computed using non-parametric bootstrap resampling' is not followed by any reported intervals in Section 6 or the tables; either add the intervals or remove the claim.
  4. [Figure 1] The captions of the running-maximum plots do not explain whether the horizontal baseline lines correspond to training-set baseline scores or hold-out scores, and no error bars or variability information is shown; the caption should be clarified.
  5. [Table 5 and Section 4.2] Cognee supports seven retriever types according to Table 5, but the optimization study considers only two retrieval strategies (cognee completion and cognee graph completion), so the parameter-space coverage is narrower than the framework description implies; this should be stated explicitly.
  6. [Section 7] The final paragraph on 'cognification' is a philosophical digression that is not connected to the experimental results and reads as a framework-motivation statement rather than a technical conclusion; consider removing it or moving it to an appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the study empirically tunes a pipeline against a training subset and evaluates on a separate held-out subset, with no derivation chain that reduces to its own inputs.

full rationale

The paper is an empirical hyperparameter optimization study, not a derivation. The central claim—that TPE-based tuning of Cognee configurations improves multi-hop QA scores—is tested against external benchmarks (HotPotQA, TwoWikiMultiHop, MuSiQue) using a train/hold-out split. The objective (EM/F1/DeepEval correctness on 24 training questions) and the reported results (best configuration on 12 held-out questions) are distinct quantities; the hold-out evaluation is not a fit by construction. The manual filtering of benchmark instances is a selection-bias and validity concern, but the filtering was performed once before tuning and the paper does not define the gains in terms of the filter. The paper acknowledges metric artifacts (EM/F1 style penalties, LLM-grader noise), which is a measurement concern, not a definitional circularity. No load-bearing self-citation, no imported uniqueness theorem, and no renaming of a known result as derivation are present. The 'cognification' discussion is an interpretation, not part of the evidence chain. Therefore no circular step can be identified.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on six tuned hyperparameters whose optimal values are not disclosed, and on assumptions about data filtering, LLM grading, and trial sufficiency. Dreamify is a wrapper, not a new entity, and Cognee is prior open-source infrastructure.

free parameters (6)
  • chunk_size
    One of six hyperparameters optimized by TPE; range 200-2000 tokens, best value not reported.
  • retriever_type
    Categorical choice between text-chunk completion and graph completion; best value not reported.
  • top_k
    Number of retrieved items, range 1-20; best value not reported.
  • qa_prompt_template
    Three variants tested; best variant not reported.
  • graph_prompt
    Three variants tested; best variant not reported.
  • task_getter_type
    Whether summaries are generated during graph construction; best value not reported.
assumptions (4)
  • ad hoc to paper Manual filtering of benchmark instances does not bias results
    Section 5 states 24 training and 12 test instances were manually reviewed and excluded if ungrammatical, ambiguous, mislabeled, or unsupported; the paper assumes this one-time filtering does not distort relative comparisons.
  • domain assumption DeepEval LLM-based correctness is a valid objective and metric
    Section 5 uses DeepEval's correctness as both optimization objective and evaluation metric; the paper assumes the LLM grader's scores reflect answer quality.
  • domain assumption TPE with 50 trials adequately explores the parameter space
    Section 4.1 justifies TPE over grid/random search; the paper assumes 50 trials suffice for reliable ranking of configurations.
  • domain assumption Cognee pipeline is deterministic enough for reproducible trials
    Section 4.1 claims trials are independent and reproducible despite LLM-based graph construction variability; no seeds are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning." pith.science (2026). https://pith.science/paper/ZUBDPEMM

@misc{pith2026250524478,
  author       = {Pith},
  title        = {Pith review of: Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUBDPEMM}},
  note         = {Machine review of arXiv:2505.24478}
}
read the original abstract

Integrating Large Language Models (LLMs) with Knowledge Graphs (KGs) results in complex systems with numerous hyperparameters that directly affect performance. While such systems are increasingly common in retrieval-augmented generation, the role of systematic hyperparameter optimization remains underexplored. In this paper, we study this problem in the context of Cognee, a modular framework for end-to-end KG construction and retrieval. Using three multi-hop QA benchmarks (HotPotQA, TwoWikiMultiHop, and MuSiQue) we optimize parameters related to chunking, graph construction, retrieval, and prompting. Each configuration is scored using established metrics (exact match, F1, and DeepEval's LLM-based correctness metric). Our results demonstrate that meaningful gains can be achieved through targeted tuning. While the gains are consistent, they are not uniform, with performance varying across datasets and metrics. This variability highlights both the value of tuning and the limitations of standard evaluation measures. While demonstrating the immediate potential of hyperparameter tuning, we argue that future progress will depend not only on architectural advances but also on clearer frameworks for optimization and evaluation in complex, modular systems.

Figures

Figures reproduced from arXiv: 2505.24478 by the authors.

Figure 1
Figure 1. Running maximum performance curves for Musique, TwoWikiMulti [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Setoka: A Benchmark for Hierarchical User Understanding in Personalized Agents over Heterogeneous Data

    cs.AI 2026-07 conditional novelty 8.0 of 10

    Setoka evaluates memory-augmented agents on four levels of user understanding—semantic memory, episodic memory, behavior patterns, and personality traits—over synthesized heterogeneous user data, and finds performance...

  2. Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions

    cs.CL 2025-07 unverdicted novelty 7.0 of 10

    MemoryAgentBench is a new multi-turn benchmark assessing four memory competencies in LLM agents—accurate retrieval, test-time learning, long-range understanding, and selective forgetting—showing that existing methods ...

  3. Know It, Act on It: Investigating Memory Utilization in LLM Personalization

    cs.CL 2026-07 conditional novelty 6.0 of 10

    LLM agents often pass a direct recall question about a user's preference yet fail to act on the same preference in a realistic request — a Know–Act gap that persists even in the best systems and is widest, on average,...

  4. Cost and Accuracy of Long-Term Memory in Distributed Multi-Agent Systems Based on Large Language Models

    cs.IR 2026-01 reject novelty 5.0 of 10

    A two-framework testbed comparison claims mem0 is Pareto-optimal over Graphiti for distributed LLM agents because its lower cost is paired with accuracy that is not significantly different.

  5. Improving Language Agents through BREW: Bootstrapping expeRientially-learned Environmental knoWledge

    cs.AI 2025-11 conditional novelty 5.0 of 10

    An LLM agent can improve by storing and Monte-Carlo-refining concept-level natural-language recipes mined from its own past trajectories.

Reference graph

Works this paper leans on

42 extracted references · 19 canonical work pages · cited by 5 Pith papers

  1. [1]

    The im- pact of large language models on scientific discovery: a preliminary study using gpt-4

    Microsoft Research AI4Science and Microsoft Azure Quantum. “The im- pact of large language models on scientific discovery: a preliminary study using gpt-4”. In: arXiv preprint arXiv:2311.07361 (2023)

    arXiv 2023

  2. [2]

    Self-rag: Learning to retrieve, generate, and critique through self-reflection

    Akari Asai et al. “Self-rag: Learning to retrieve, generate, and critique through self-reflection”. In: The Twelfth International Conference on Learn- ing Representations. 2023

    work page 2023

  3. [3]

    Faster, Cheaper, Better: Multi-Objective Hyper- parameter Optimization for LLM and RAG Systems

    Matthew Barker et al. “Faster, Cheaper, Better: Multi-Objective Hyper- parameter Optimization for LLM and RAG Systems”. In: arXiv preprint arXiv:2502.18635 (2025)

    arXiv 2025

  4. [4]

    Neural le- gal judgment prediction in English

    Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. “Neural le- gal judgment prediction in English”. In: arXiv preprint arXiv:1906.02059 (2019). 14

    arXiv 2019

  5. [5]

    Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning

    Yiqun Chen et al. “Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning”. In:arXiv preprint arXiv:2501.15228 (2025)

    arXiv 2025

  6. [6]

    Complex logical reasoning over knowledge graphs using large language models

    Nurendra Choudhary and Chandan K Reddy. “Complex logical reasoning over knowledge graphs using large language models”. In: arXiv preprint arXiv:2305.01157 (2023)

    arXiv 2023

  7. [7]

    From local to global: A graph rag approach to query- focused summarization

    Darren Edge et al. “From local to global: A graph rag approach to query- focused summarization”. In: arXiv preprint arXiv:2404.16130 (2024)

    arXiv 2024

  8. [8]

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

    Wenqi Fan et al. “A survey on rag meeting llms: Towards retrieval-augmented large language models”. In: Proceedings of the 30th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining . 2024, pp. 6491–6501

    work page 2024

Show all 42 references
  1. [9]

    AutoRAG-HP: Automatic Online Hyper-Parameter Tuning for Retrieval-Augmented Generation

    Jia Fu et al. “AutoRAG-HP: Automatic Online Hyper-Parameter Tuning for Retrieval-Augmented Generation”. In:arXiv preprint arXiv:2406.19251 (2024)

    2024 arXiv

  2. [10]

    Retrieval-augmented generation for large language models: A survey

    Yunfan Gao et al. “Retrieval-augmented generation for large language models: A survey”. In: arXiv preprint arXiv:2312.10997 2 (2023)

    2023 arXiv

  3. [11]

    A compre- hensive survey of retrieval-augmented generation (rag): Evolution, cur- rent landscape and future directions

    Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. “A compre- hensive survey of retrieval-augmented generation (rag): Evolution, cur- rent landscape and future directions”. In:arXiv preprint arXiv:2410.12837 (2024)

    2024 arXiv

  4. [12]

    Hipporag: Neurobiologically inspired long-term memory for large language models

    Bernal Jim´ enez Guti´ errez et al. “Hipporag: Neurobiologically inspired long-term memory for large language models”. In: The Thirty-eighth An- nual Conference on Neural Information Processing Systems . 2024

    2024

  5. [13]

    Retrieval augmented language model pre-training

    Kelvin Guu et al. “Retrieval augmented language model pre-training”. In: International conference on machine learning. PMLR. 2020, pp. 3929– 3938

    2020

  6. [14]

    Retrieval-augmented generation with graphs (graphrag)

    Haoyu Han et al. “Retrieval-augmented generation with graphs (graphrag)”. In: arXiv preprint arXiv:2501.00309 (2024)

    2024 arXiv

  7. [15]

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

    Xiaoxin He et al. “G-retriever: Retrieval-augmented generation for textual graph understanding and question answering”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 132876–132907

    2024

  8. [16]

    Constructing a multi-hop qa dataset for comprehen- sive evaluation of reasoning steps

    Xanh Ho et al. “Constructing a multi-hop qa dataset for comprehen- sive evaluation of reasoning steps”. In: arXiv preprint arXiv:2011.01060 (2020)

    2020 arXiv

  9. [17]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

    Lei Huang et al. “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions”. In: ACM Transac- tions on Information Systems 43.2 (2025), pp. 1–55

    2025

  10. [18]

    Introducing a new hyper-parameter for RAG: Context Window Utilization

    Kush Juvekar and Anupam Purwar. “Introducing a new hyper-parameter for RAG: Context Window Utilization”. In:arXiv preprint arXiv:2407.19794 (2024). 15

    2024 arXiv

  11. [19]

    Knowledge Graphs, Large Language Models, and Hallucinations: An NLP Perspective

    Ernests Lavrinovics et al. “Knowledge Graphs, Large Language Models, and Hallucinations: An NLP Perspective”. In: Journal of Web Semantics 85 (2025), p. 100844

    2025

  12. [20]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis et al. “Retrieval-augmented generation for knowledge-intensive nlp tasks”. In: Advances in neural information processing systems 33 (2020), pp. 9459–9474

    2020

  13. [21]

    The dawn after the dark: An empirical study on factuality hallucination in large language models

    Junyi Li et al. “The dawn after the dark: An empirical study on factuality hallucination in large language models”. In:arXiv preprint arXiv:2401.03205 (2024)

    2024 arXiv

  14. [22]

    Reasoning on graphs: Faithful and interpretable large language model reasoning

    Linhao Luo et al. “Reasoning on graphs: Faithful and interpretable large language model reasoning”. In: arXiv preprint arXiv:2310.01061 (2023)

    2023 arXiv

  15. [23]

    When Not to Trust Language Models: Investigating Ef- fectiveness of Parametric and Non-Parametric Memories

    Alex Mallen et al. “When Not to Trust Language Models: Investigating Ef- fectiveness of Parametric and Non-Parametric Memories”. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguis- tics (Volume 1: Long Papers) . Ed. by Anna Rogers, Jordan B...

    2023

  16. [24]

    Unifying large language models and knowledge graphs: A roadmap

    Shirui Pan et al. “Unifying large language models and knowledge graphs: A roadmap”. In: IEEE Transactions on Knowledge and Data Engineering 36.7 (2024), pp. 3580–3599

    2024

  17. [25]

    Graph retrieval-augmented generation: A survey

    Boci Peng et al. “Graph retrieval-augmented generation: A survey”. In: arXiv preprint arXiv:2408.08921 (2024)

    2024 arXiv

  18. [26]

    Language models as knowledge bases?

    Fabio Petroni et al. “Language models as knowledge bases?” In: arXiv preprint arXiv:1909.01066 (2019)

    2019 arXiv

  19. [27]

    How much knowledge can you pack into the parameters of a language model?

    Adam Roberts, Colin Raffel, and Noam Shazeer. “How much knowledge can you pack into the parameters of a language model?” In:arXiv preprint arXiv:2002.08910 (2020)

    2020 arXiv

  20. [28]

    Toward expert-level medical question answering with large language models

    Karan Singhal et al. “Toward expert-level medical question answering with large language models”. In: Nature Medicine (2025), pp. 1–8

    2025

  21. [29]

    The web as a knowledge-base for an- swering complex questions

    Alon Talmor and Jonathan Berant. “The web as a knowledge-base for an- swering complex questions”. In: arXiv preprint arXiv:1803.06643 (2018)

    2018 arXiv

  22. [30]

    From Lou- vain to Leiden: guaranteeing well-connected communities

    Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck. “From Lou- vain to Leiden: guaranteeing well-connected communities”. In: Scientific reports 9.1 (2019), pp. 1–12

    2019

  23. [31]

    MuSiQue: Multihop Questions via Single-hop Ques- tion Composition

    Harsh Trivedi et al. “MuSiQue: Multihop Questions via Single-hop Ques- tion Composition”. In: Transactions of the Association for Computational Linguistics 10 (2022), pp. 539–554

    2022

  24. [32]

    Attention is all you need

    Ashish Vaswani et al. “Attention is all you need”. In: Advances in neural information processing systems 30 (2017). 16

    2017

  25. [33]

    Cost-effective hyperparameter optimization for large language model generation infer- ence

    Chi Wang, Xueqing Liu, and Ahmed Hassan Awadallah. “Cost-effective hyperparameter optimization for large language model generation infer- ence”. In: International Conference on Automated Machine Learning. PMLR. 2023, pp. 21–1

    2023

  26. [34]

    Knowledge graph prompting for multi-document question answering

    Yu Wang et al. “Knowledge graph prompting for multi-document question answering”. In: Proceedings of the AAAI Conference on Artificial Intelli- gence. Vol. 38. 17. 2024, pp. 19206–19214

    2024

  27. [35]

    Bloomberggpt: A large language model for finance

    Shijie Wu et al. “Bloomberggpt: A large language model for finance”. In: arXiv preprint arXiv:2303.17564 (2023)

    2023 arXiv

  28. [36]

    Corrective retrieval augmented generation

    Shi-Qi Yan et al. “Corrective retrieval augmented generation”. In: (2024)

    2024

  29. [37]

    HotpotQA: A dataset for diverse, explainable multi- hop question answering

    Zhilin Yang et al. “HotpotQA: A dataset for diverse, explainable multi- hop question answering”. In: arXiv preprint arXiv:1809.09600 (2018)

    2018 arXiv

  30. [38]

    The value of semantic parse labeling for knowledge base question answering

    Wen-tau Yih et al. “The value of semantic parse labeling for knowledge base question answering”. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) . 2016, pp. 201–206

    2016

  31. [39]

    Neural, symbolic and neural-symbolic reasoning on knowledge graphs

    Jing Zhang et al. “Neural, symbolic and neural-symbolic reasoning on knowledge graphs”. In: AI Open 2 (2021), pp. 14–35

    2021

  32. [40]

    Subgraph retrieval enhanced model for multi-hop knowl- edge base question answering

    Jing Zhang et al. “Subgraph retrieval enhanced model for multi-hop knowl- edge base question answering”. In:arXiv preprint arXiv:2202.13296 (2022)

    2022 arXiv

  33. [41]

    Variational reasoning for question answering with knowledge graph

    Yuyu Zhang et al. “Variational reasoning for question answering with knowledge graph”. In: Proceedings of the AAAI conference on artificial intelligence. Vol. 32. 1. 2018

    2018

  34. [42]

    Retrieval-augmented generation for ai-generated content: A survey

    Penghao Zhao et al. “Retrieval-augmented generation for ai-generated content: A survey”. In: arXiv preprint arXiv:2402.19473 (2024). 17

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.