Pith. sign in

REVIEW 3 major objections 4 minor 73 references

SCAIR, a training-free framework, injects schema-conditioned structural priors into agentic KG-RAG and outperforms generic methods on dense enterprise knowledge graphs.

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

2026-08-02 12:24 UTC pith:TJBW3EG3

load-bearing objection A useful enterprise KG-RAG benchmark and a sensible schema-aware agent, but the central claim that agentic iteration is essential outruns the evidence. the 3 major comments →

arxiv 2607.22571 v1 pith:TJBW3EG3 submitted 2026-06-02 cs.AI

SCAIR: Schema-Conditioned Agentic Iterative Reasoning for Enterprise Knowledge Graphs

Prateek Chaturvedi , Yuqicheng Zhu , Hongkuan Zhou , Dongzhuoran Zhou , Yunjie He , Steffen Staab , Fei Du , Jie Tang
show 1 more author
Evgeny Kharlamov
This is my paper
classification cs.AI
keywords knowledge graph question answeringretrieval-augmented generationschema-aware reasoningenterprise knowledge graphCMDBagentic reasoningmulti-hop reasoningtraining-free
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.

The paper argues that generic agentic KG-RAG methods, which work on public benchmarks, fail on dense, schema-driven enterprise knowledge graphs because they ignore structural and operational constraints. It introduces SCAIR, a training-free framework that injects schema-conditioned structural priors and enforces schema-aware traversal during multi-hop reasoning. On an enterprise CMDB benchmark, SCAIR achieves 35.14% accuracy versus 25.27% for the strongest baseline, without task-specific training. The authors conclude that reliable enterprise graph reasoning must explicitly incorporate the target domain's structural and operational constraints, not just rely on stronger language models.

Core claim

The central discovery is that in dense enterprise knowledge graphs (such as a CMDB), the bottleneck is not LLM reasoning capability but uncontrolled graph traversal: ReAct-style exploration suffers search explosion and semantic drift through high-degree attribute nodes, while plan-and-execute methods overfit to training distributions. SCAIR addresses this by generating schema-consistent relation paths, decomposing questions into depth-aligned subquestions, filtering and scoring candidate relations and entities conditioned on schema, and maintaining a union of previously discovered topic entities to balance exploration and exploitation. This structure-first, semantics-second approach yields c

What carries the argument

SCAIR (Schema-Conditioned Agentic Iterative Reasoning) is a training-free agentic framework that integrates three mechanisms: schema-conditioned planning, which generates valid relation paths and depth-aligned subquestions; schema-aware iterative reasoning, which filters candidate relations and entities using schema constraints before semantic scoring; and controlled topic entity propagation, which keeps a union of previously discovered and newly expanded entities. The schema-aware traversal loop is the central object that carries the argument, as ablations removing it degrade performance.

Load-bearing premise

The benchmark's question-answer pairs are generated from templates defined over the same schema SCAIR explicitly exploits, so if real enterprise queries are less schema-aligned, SCAIR's measured advantage may be partly an artifact of benchmark construction.

What would settle it

Give a schema-agnostic agent (e.g., a ReAct-style method) the same schema in its prompt without any other SCAIR machinery; if its performance matches SCAIR, the gain comes from schema availability, not the agentic control loop. Alternatively, evaluate on naturally collected enterprise queries rather than template-generated ones and check whether the accuracy gap persists.

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

If this is right

  • Enterprise KG-RAG deployments should prioritize schema-aware traversal control over scaling the LLM backbone, since replacing a LLaMA backbone with Qwen in a plan-and-execute method did not improve end-to-end performance.
  • Training-free schema injection offers a low-cost adaptation path: schema updates or new business rules can be handled without fine-tuning, unlike plan-and-execute methods that overfit to training query distributions.
  • The benchmark itself, comprising 19,080 questions from a real CMDB, provides a realistic stress test for evaluating KG-RAG under dense, constraint-heavy conditions.
  • The measured gains span multi-hop, intersection, and complex query types, suggesting the improvement comes from structural control rather than memorization or language modeling.
  • The higher inference cost (42.36 LLM calls per question) is expected to be mitigated by prompt caching in deployed APIs, making the approach viable where answer correctness is prioritized over per-query latency.

Where Pith is reading between the lines

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

  • The template-based benchmark may over-align questions with the schema SCAIR explicitly uses; a fair comparison would give baselines access to the same schema information, or test on naturally collected enterprise questions, to see if the advantage shrinks.
  • The principle of pruning by schema validity before semantic scoring could transfer to non-agentic pipelines such as text-to-SPARQL or embedding-based retrieval, potentially improving their robustness in dense graphs.
  • The paper's claim implies a testable prediction: if the schema is incomplete or noisy, SCAIR's advantage over generic agents should diminish; ablating schema access would quantify this dependency.
  • The focus on exact-match accuracy may understate the operational benefit visible in Hits@Any and F1; real enterprise users may tolerate partial answers, so the practical impact could exceed the headline accuracy gap.

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 / 4 minor

Summary. The paper proposes SCAIR, a training-free, schema-conditioned agentic KG-RAG framework that interleaves structured planning with schema-aware iterative traversal for enterprise knowledge graphs. The authors construct a benchmark from a real-world manufacturing CMDB, generating 19,080 template-based questions over nine query types, and report that SCAIR outperforms four existing KG-RAG baselines (RoG, G-Retriever, ToG, PoG), achieving 35.14 accuracy vs. 25.27 for the strongest baseline. They argue that generic agentic designs fail on dense, schema-governed enterprise graphs and that structural control over traversal is the key bottleneck, supported by ablations, failure case studies, and an inference-cost analysis.

Significance. If the empirical claims are substantiated, the paper makes a useful contribution: it releases an enterprise-oriented CMDB benchmark, demonstrates failure modes of existing KG-RAG methods under dense schema constraints, and shows that a training-free, schema-prompted agent can achieve strong gains without fine-tuning. The paper includes several strengths: a detailed algorithm (Algorithm 1), full prompt templates in Appendix C.2, ablation over structural controls (Figure 4), sensitivity analysis (Appendix E.2), an inference-cost comparison (Table 2), and explicit acknowledgment of a missing baseline in Section 7. However, the central claim — that SCAIR's agentic design, rather than its schema prior or model backbone, drives the improvement — is not yet established because of the confounds detailed below. The benchmark itself is also tightly coupled to the schema SCAIR receives, which limits the external-validity conclusions that can currently be drawn.

major comments (3)
  1. [§5.1, Table 4; Appendix E.3] The main comparison confounds method with backbone. RoG and G-Retriever are trained on LLaMA-2-7B, while SCAIR and the training-free baselines use GPT-4.1-mini. Since the central claim is that SCAIR substantially improves over existing KG-RAG methods, a same-backbone baseline is required — e.g., RoG/G-Retriever with GPT-4.1-mini, or SCAIR with LLaMA-2-7B. Appendix E.3 only varies RoG's backbone between LLaMA and Qwen and does not compare against SCAIR under a shared backbone. Without this control, the reported 10-point accuracy gain over G-Retriever could be partly attributable to model strength.
  2. [§3.2, Appendix A.2, Appendix C.2, §7] The benchmark is constructed by instantiating templates defined over the CMDB schema, and SCAIR's planning stage explicitly receives that same schema to generate relation paths and subquestions. The evaluation compares SCAIR only against baselines that do not receive the schema as a structured prior. The paper itself acknowledges in Section 7 that non-agentic schema-constrained traversal and text-to-SPARQL baselines are missing. This is load-bearing: without such controls, the reported gains do not isolate the contribution of agentic iterative reasoning from the schema-conditioned prior. The tight alignment between the template-generated questions and SCAIR's schema prompting further raises the risk of benchmark-construction artifact. Please add a schema-only non-agentic control and, ideally, a same-schema baseline for ToG or PoG.
  3. [Appendix E.2, §5.2] The beam width w=6 is selected after evaluating w∈{3,4,6} on the benchmark itself (Figure 10) and is then used for SCAIR in the main comparison, whereas baselines use settings from their original papers. This is effectively tuning SCAIR on the test set, which can inflate the reported margin. A validation split or a robustness table reporting results across beam widths for the full comparison is needed to establish that the improvement is not an artifact of hyperparameter selection.
minor comments (4)
  1. [§3.2, Appendix A.1] The paper says the benchmark is 'grounded in a real-world manufacturing CMDB,' but Appendix A.1 explains that the KG is extended with auxiliary relations (similarTo, componentStatus, manufacturer, ipAddress, macAddress) and the questions are template-generated. Please state this explicitly in the main text to avoid the impression that the benchmark contains naturally occurring user questions.
  2. [§3.2, Example 2] The statement that the example query 'cannot be expressed as a single well-defined compositional logical form' is overstated; the query is expressible as a SPARQL query with constraints. Consider softening the claim to 'cannot be easily expressed as a simple path query.'
  3. [Table 4] ToG and SCAIR differ in both search depth (3 vs. 4) and beam width (3 vs. 6). While Table 2 reports cost differences, the accompanying accuracy comparison does not control for these search-budget asymmetries. A matched-budget comparison would strengthen the claim that SCAIR's advantage comes from schema-aware control rather than from simply allocating more search.
  4. [Appendix C.2] The relation-path generation prompt shows a schema fragment (CMDBLine -hasMachines-> CMDBMachine, etc.). Please clarify whether the full schema is provided to the LLM or only this fragment, as this is important for reproducibility and for assessing the fairness of the schema prior.

Circularity Check

2 steps flagged

Benchmark templates and SCAIR share the same schema-conditioned path structure; missing schema-only control makes the central claim partly circular.

specific steps
  1. other [Appendix A.2 (Question-Answer Pair Construction) and Appendix C.2 (Prompting Design)]
    "Templates are defined over the CMDB schema and specify both the underlying graph traversal pattern and a natural language realization. ... Given the enterprise KG schema and the input question, the model first generates valid relation paths that are consistent with the ontology."

    The benchmark's gold answers are produced by executing template-defined graph traversals over the CMDB schema, and SCAIR's first planning step is prompted with the same schema and asked to output schema-consistent relation paths—the same traversal structure the templates encode. The relation-path prompt even includes the exact CMDBLine→CMDBMachine→CMDBMachineComponent→similarTo chain. Thus SCAIR is effectively handed the template structure as input, so its measured advantage over schema-blind baselines is partly built into the benchmark rather than independently demonstrated. The paper's own Section 7 concedes that non-agentic schema-constrained traversal and text-to-SPARQL baselines are needed to disentangle structural filtering from agentic iteration.

  2. fitted input called prediction [Appendix E.2 (Sensitivity to Beam Width) and Appendix C.1 (Reasoning Parameters)]
    "We therefore adopt w=6 as a balanced setting that provides strong performance while maintaining controlled exploration."

    The beam width w=6 is selected by evaluating SCAIR on the same enterprise benchmark and choosing the value that yields the best F1/accuracy, while baselines use their originally reported hyperparameters. Consequently, the reported headline margin (SCAIR 35.14 Accuracy vs. G-Retriever 25.27) includes a hyperparameter fitted to the test set. This is not a parameter-free prediction but a selection effect that partially inflates SCAIR's advantage.

full rationale

The central derivation chain is empirical: a CMDB benchmark is constructed, SCAIR is compared with existing KG-RAG methods, and the improvement is attributed to schema-aware agentic reasoning. The main circularity is that the benchmark's question-answer pairs are generated from templates defined over the CMDB schema (Appendix A.2), while SCAIR's planning stage is explicitly prompted to generate schema-consistent relation paths from that same schema (Appendix C.2). The method is therefore evaluated on queries whose answer structure matches the exact structural prior SCAIR is given, making the central claim that schema-aware traversal drives the gains partially self-fulfilling. The paper itself acknowledges the missing disentangling baseline in Section 7. Additionally, SCAIR's beam width is tuned on the benchmark itself (Appendix E.2) while baselines use their reported settings, contributing a smaller selection-effect component. Self-citations are present but not load-bearing: the evaluation-protocol citation (Zhou et al. 2025b, 2026a) is for standard metrics, and the GR-Agent citation is only a related-work mention. The ablation study shows structural controls matter, giving the claim some independent content, so this is partial circularity rather than a fully forced result.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim that SCAIR outperforms generic KG-RAG on enterprise graphs rests on a benchmark that is constructed from the same schema SCAIR is given, and on the availability of that schema. The auxiliary relations and template definitions are introduced ad hoc to the paper, and the method's hyperparameters are tuned on the test set. This makes the contribution's measured advantage partially self-fulfilling.

free parameters (4)
  • beam_width (w) = 6
    Selected based on sensitivity analysis on the same benchmark (Appendix E.2) to maximize F1; this is effectively tuning on the test set.
  • max_depth (d) = 4
    Set by the authors; no full sensitivity analysis reported, but the choice affects multi-hop completeness and is a hand-chosen hyperparameter.
  • subsampling thresholds = 20 candidates, retain top 8 per relation
    Ad hoc controls to limit search space in dense graphs; not justified by experiments, and the central results depend on them.
  • temperature settings = 0.4 exploration, 0.0 reasoning
    Chosen for controlled diversity; no experimental comparison on this choice.
axioms (4)
  • domain assumption The template-generated QA pairs faithfully represent real enterprise information needs.
    Section 3.2 and A.2; questions are automatically generated from schema-aligned templates, assuming these correspond to real business queries like component replacement and diagnostics.
  • ad hoc to paper The auxiliary relations added to the CMDB (similarTo, componentStatus, manufacturer, ipAddress, macAddress) are representative of real enterprise operational data.
    Appendix A.1 states they extend the original CMDB KG with these relations to support the considered reasoning patterns; no external validation that these reflect actual enterprise semantics.
  • domain assumption The schema is available and can be injected at inference time.
    SCAIR requires a clean, explicit schema of the KG; enterprise systems may not always expose a machine-readable schema, but paper assumes it is available (Section 4).
  • domain assumption GPT-4.1-mini's zero-shot relation/entity scoring is reliable enough for the method to work.
    All scoring steps (relation scoring, entity scoring, sufficiency checks) depend on the LLM following the prompts exactly; no calibration or human evaluation of these scores is provided (Appendix C.2).
invented entities (1)
  • similarTo relation no independent evidence
    purpose: New relation added to the CMDB-KG to support replacement queries by linking functionally interchangeable components.
    Added by the authors in Appendix A.1; no external source confirms that this reflects actual component interchangeability rules in the real CMDB. The benchmark's complex queries rely on it.

pith-pipeline@v1.3.0-alltime-deepseek · 15201 in / 11082 out tokens · 107911 ms · 2026-08-02T12:24:49.427405+00:00 · methodology

0 comments
read the original abstract

Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) enables natural language interaction with structured enterprise knowledge, yet existing agentic approaches that perform well on public benchmarks often fail to generalize to real-world enterprise Knowledge Graphs (KGs), which are dense, schema-driven, and operationally constrained. To address these limitations, we propose SCAIR (Schema-Conditioned Agentic Iterative Reasoning), a training-free framework that integrates structured planning with controlled iterative reasoning by injecting schema-conditioned structural priors and enforcing schema-aware traversal during multi-hop reasoning. Experiments on an enterprise-oriented benchmark constructed from a real-world Configuration Management DataBase (CMDB) demonstrate that SCAIR substantially improves performance over existing KG-RAG methods. Crucially, our study highlights that reliable enterprise graph reasoning cannot rely on generic agentic designs; instead, it must explicitly incorporate the target domain's structural and operational constraints into the reasoning process. We demonstrate that by aligning agent design with business logic, substantial performance gains can be achieved without the need for costly model retraining.

Figures

Figures reproduced from arXiv: 2607.22571 by Dongzhuoran Zhou, Evgeny Kharlamov, Fei Du, Hongkuan Zhou, Jie Tang, Prateek Chaturvedi, Steffen Staab, Yunjie He, Yuqicheng Zhu.

Figure 1
Figure 1. Figure 1: Representative fragment of CMDB-KG. exploration by learning explicit reasoning plans or subgraph selection strategies. However, they rely on substantial training and hyperparameter tuning to perform reliably, which is particularly challenging when adapting large models in enter￾prise environments. In practice, effective training requires sufficient coverage of the underlying query and relation distribution… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Plan-and-Execute, ReAct and SCAIR. ing the topic entity set at each depth, the algorithm maintains a union of previously discovered entities and newly expanded ones. This preserves earlier reasoning anchors (exploitation) while enabling deeper traversal into new regions of the graph (ex￾ploration). In contrast, ReAct-style methods typ￾ically update the working state using only newly expanded en… view at source ↗
Figure 4
Figure 4. Figure 4: Component Impact on Overall Performance. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy comparison across query types. Re [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of a template-based question speci [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustrative examples of structured reasoning [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hits@Any per query type for all evaluated [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance of SCAIR under different beam [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of RoG with LLaMA and Qwen backbones. In a controlled setting where gold-standard inter￾mediate triples are provided, Qwen demonstrates stronger local reasoning ability than LLaMA, cor￾rectly answering complex compositional queries that LLaMA fails to resolve. This suggests that Qwen has improved conditional reasoning capacity when the relevant evidence is explicitly given. However, when evalua… 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

73 extracted references · 5 linked inside Pith

  1. [1]

    Business & information systems engineering , volume=

    Industry 4.0 , author=. Business & information systems engineering , volume=. 2014 , publisher=

    2014

  2. [2]

    International journal of production economics , volume=

    Industry 4.0 technologies: Implementation patterns in manufacturing companies , author=. International journal of production economics , volume=. 2019 , publisher=

    2019

  3. [3]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Unifying large language models and knowledge graphs: A roadmap , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2024 , publisher=

    2024

  4. [4]

    ACM Computing Surveys (Csur) , volume=

    Knowledge graphs , author=. ACM Computing Surveys (Csur) , volume=. 2021 , publisher=

    2021

  5. [5]

    Yushan Liu and Bailan He and Marcel Hildebrandt and Maximilian Buchner and Daniela Inzko and Roger Wernert and Emanuel Weigel and Dagmar Beyer and Martin Berbalk and Volker Tresp , title =

  6. [6]

    IEEE Access , year=

    Knowledge graphs in the digital twin: A systematic literature review about the combination of semantic technologies and simulation in industrial automation , author=. IEEE Access , year=

  7. [7]

    2020 25th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA) , volume=

    Knowledge graphs for efficient integration and access of manufacturing data , author=. 2020 25th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA) , volume=. 2020 , organization=

    2020

  8. [8]

    European Semantic Web Conference , pages=

    MYAM: LLM-Supported Mapping Generation for Semantic Manufacturing Retrieval , author=. European Semantic Web Conference , pages=. 2025 , organization=

    2025

  9. [9]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Patrick Lewis and Ethan Perez and Aleksandra Piktus and Fabio Petroni and Vladimir Karpukhin and Naman Goyal and Heinrich K. Retrieval-Augmented Generation for Knowledge-Intensive. NeurIPS , year =

  10. [10]

    Gautier Izacard and Edouard Grave , title =

  11. [11]

    Improving Language Models by Retrieving from Trillions of Tokens , booktitle =

    Sebastian Borgeaud and Arthur Mensch and Jordan Hoffmann and Trevor Cai and Eliza Rutherford and Katie Millican and George van den Driessche and Jean. Improving Language Models by Retrieving from Trillions of Tokens , booktitle =

  12. [12]

    arXiv preprint arXiv:2502.06864 , year=

    Knowledge Graph-Guided Retrieval Augmented Generation , author=. arXiv preprint arXiv:2502.06864 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2501.13958 , year=

    A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models , author=. arXiv preprint arXiv:2501.13958 , year=

    arXiv

  14. [14]

    arXiv preprint arXiv:2408.08921 , year=

    Graph retrieval-augmented generation: A survey , author=. arXiv preprint arXiv:2408.08921 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2405.16506 , year=

    Grag: Graph retrieval-augmented generation , author=. arXiv preprint arXiv:2405.16506 , year=

    Pith/arXiv arXiv

  16. [16]

    Chawla and Thomas Laurent and Yann LeCun and Xavier Bresson and Bryan Hooi , title =

    Xiaoxin He and Yijun Tian and Yifei Sun and Nitesh V. Chawla and Thomas Laurent and Yann LeCun and Xavier Bresson and Bryan Hooi , title =. NeurIPS , year =

  17. [17]

    Ni and Heung

    Jiashuo Sun and Chengjin Xu and Lumingyuan Tang and Saizhuo Wang and Chen Lin and Yeyun Gong and Lionel M. Ni and Heung. Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph , booktitle =

  18. [18]

    Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning , booktitle =

    Linhao Luo and Yuan. Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning , booktitle =

  19. [19]

    Reducing hallucination in structured outputs via Retrieval-Augmented Generation , booktitle =

    Orlando Ayala and Patrice B. Reducing hallucination in structured outputs via Retrieval-Augmented Generation , booktitle =

  20. [20]

    Knowledge graph enhanced retrieval-augmented generation for failure mode and effects analysis , journal =

    Lukas Bahr and Christoph Wehner and Judith Wewerka and Jos. Knowledge graph enhanced retrieval-augmented generation for failure mode and effects analysis , journal =

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs , author=. Advances in Neural Information Processing Systems , volume=

  22. [22]

    Proceedings of the International Research and Industry Symposium on AI (IRIS-AI 25) , publisher=

    Evaluating Knowledge Graph Based Retrieval Augmented Generation Methods under Knowledge Incompleteness , author=. Proceedings of the International Research and Industry Symposium on AI (IRIS-AI 25) , publisher=. 2025 , pages=

    2025

  23. [23]

    What Breaks Knowledge Graph based RAG ? Benchmarking and Empirical Insights into Reasoning under Incomplete Knowledge

    Zhou, Dongzhuoran and Zhu, Yuqicheng and Wang, Xiaxia and Zhou, Hongkuan and He, Yuan and Chen, Jiaoyan and Staab, Steffen and Kharlamov, Evgeny. What Breaks Knowledge Graph based RAG ? Benchmarking and Empirical Insights into Reasoning under Incomplete Knowledge. Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputa...

    2026

  24. [24]

    arXiv preprint arXiv:2512.14766 , year=

    GR-Agent: Adaptive Graph Reasoning Agent under Incomplete Knowledge , author=. arXiv preprint arXiv:2512.14766 , year=

    arXiv

  25. [25]

    Proceedings of the 5th

    Bhaskarjit Sarmah and Dhagash Mehta and Benika Hall and Rohan Rao and Sunil Patel and Stefano Pasquali , title =. Proceedings of the 5th

  26. [26]

    Rose and John H

    Karthik Soman and Peter W. Rose and John H. Morris and Rabia E. Akbas and Brett Smith and Braian Peetoom and Catalina Villouta. Biomedical knowledge graph-enhanced prompt generation for large language models , journal =

  27. [27]

    Jonathan Berant and Percy Liang , title =

  28. [28]

    Jonathan Berant and Andrew Chou and Roy Frostig and Percy Liang , title =

  29. [29]

    Siva Reddy and Mirella Lapata and Mark Steedman , title =. Trans. Assoc. Comput. Linguistics , volume =

  30. [30]

    Narasimhan and Yuan Cao , title =

    Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik R. Narasimhan and Yuan Cao , title =

  31. [31]

    Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models , booktitle =

    Lei Wang and Wanyu Xu and Yihuai Lan and Zhiqiang Hu and Yunshi Lan and Roy Ka. Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models , booktitle =

  32. [32]

    The Value of Semantic Parse Labeling for Knowledge Base Question Answering , booktitle =

    Wen. The Value of Semantic Parse Labeling for Knowledge Base Question Answering , booktitle =

  33. [33]

    Alon Talmor and Jonathan Berant , title =

  34. [34]

    Bollacker and Colin Evans and Praveen K

    Kurt D. Bollacker and Colin Evans and Praveen K. Paritosh and Tim Sturge and Jamie Taylor , title =

  35. [35]

    Li Dong and Furu Wei and Ming Zhou and Ke Xu , title =

  36. [36]

    CoRR , volume =

    Antoine Bordes and Nicolas Usunier and Sumit Chopra and Jason Weston , title =. CoRR , volume =

  37. [37]

    Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering , booktitle =

    Zhentao Xu and Mark Jerome Cruz and Matthew Guevara and Tie Wang and Manasi Deshpande and Xiaofeng Wang and Zheng Li , editor =. Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering , booktitle =

  38. [38]

    Golden-Retriever: High-Fidelity Agentic Retrieval Augmented Generation for Industrial Knowledge Base , journal =

    Zhiyu An and Xianzhong Ding and Yen. Golden-Retriever: High-Fidelity Agentic Retrieval Augmented Generation for Industrial Knowledge Base , journal =

  39. [39]

    Breaking the Barrier: Utilizing Large Language Models for Industrial Recommendation Systems through an Inferential Knowledge Graph , booktitle =

    Qian Zhao and Hao Qian and Ziqi Liu and Gong. Breaking the Barrier: Utilizing Large Language Models for Industrial Recommendation Systems through an Inferential Knowledge Graph , booktitle =

  40. [40]

    CoRR , volume =

    OpenAI , title =. CoRR , volume =

  41. [41]

    LLaMA: Open and Efficient Foundation Language Models , journal =

    Hugo Touvron and Thibaut Lavril and Gautier Izacard and Xavier Martinet and Marie. LLaMA: Open and Efficient Foundation Language Models , journal =

  42. [42]

    Cole and Julian Martin Eisenschlos and Daniel Gillick and Jacob Eisenstein and William W

    Bhuwan Dhingra and Jeremy R. Cole and Julian Martin Eisenschlos and Daniel Gillick and Jacob Eisenstein and William W. Cohen , title =. Trans. Assoc. Comput. Linguistics , volume =

  43. [43]

    Xianzhi Li and Samuel Chan and Xiaodan Zhu and Yulong Pei and Zhiqiang Ma and Xiaomo Liu and Sameena Shah , title =

  44. [44]

    Ziwei Ji and Nayeon Lee and Rita Frieske and Tiezheng Yu and Dan Su and Yan Xu and Etsuko Ishii and Yejin Bang and Andrea Madotto and Pascale Fung , title =

  45. [45]

    Urvashi Khandelwal and Omer Levy and Dan Jurafsky and Luke Zettlemoyer and Mike Lewis , title =

  46. [46]

    In-Context Retrieval-Augmented Language Models , journal =

    Ori Ram and Yoav Levine and Itay Dalmedigos and Dor Muhlgay and Amnon Shashua and Kevin Leyton. In-Context Retrieval-Augmented Language Models , journal =

  47. [47]

    Rossi and Subhabrata Mukherjee and Xianfeng Tang and Qi He and Zhigang Hua and Bo Long and Tong Zhao and Neil Shah and Amin Javari and Yinglong Xia and Jiliang Tang , title =

    Haoyu Han and Yu Wang and Harry Shomer and Kai Guo and Jiayuan Ding and Yongjia Lei and Mahantesh Halappanavar and Ryan A. Rossi and Subhabrata Mukherjee and Xianfeng Tang and Qi He and Zhigang Hua and Bo Long and Tong Zhao and Neil Shah and Amin Javari and Yinglong Xia and Jiliang Tang , title =. CoRR , volume =

  48. [48]

    Alex Mallen and Akari Asai and Victor Zhong and Rajarshi Das and Daniel Khashabi and Hannaneh Hajishirzi , title =

  49. [49]

    Manning and Percy Liang and Jure Leskovec , title =

    Michihiro Yasunaga and Antoine Bosselut and Hongyu Ren and Xikun Zhang and Christopher D. Manning and Percy Liang and Jure Leskovec , title =. NeurIPS , year =

  50. [50]

    LLM-Supported Mapping Generation for Semantic Manufacturing Treasure Hunting , booktitle =

    Wilma Johanna Schmidt and Irl. LLM-Supported Mapping Generation for Semantic Manufacturing Treasure Hunting , booktitle =

  51. [51]

    Talukdar , title =

    Apoorv Saxena and Aditay Tripathi and Partha P. Talukdar , title =

  52. [52]

    2018 , url=

    Improving Language Understanding by Generative Pre-Training , author=. 2018 , url=

    2018

  53. [53]

    Improving Multi-hop Knowledge Base Question Answering by Learning Intermediate Supervision Signals

    Gaole He and Yunshi Lan and Jing Jiang and Wayne Xin Zhao and Ji. Improving Multi-hop Knowledge Base Question Answering by Learning Intermediate Supervision Signals

  54. [54]

    NeurIPS , year =

    Erik Arakelyan and Pasquale Minervini and Daniel Daza and Michael Cochez and Isabelle Augenstein , title =. NeurIPS , year =

  55. [55]

    Seeing and Knowing in the Wild: Open-domain Visual Entity Recognition with Large-scale Knowledge Graphs via Contrastive Learning , booktitle =

    Hongkuan Zhou and Lavdim Halilaj and Sebastian Monka and Stefan Schmid and Yuqicheng Zhu and Jingcheng Wu and Nadeem Nazer and Steffen Staab , editor =. Seeing and Knowing in the Wild: Open-domain Visual Entity Recognition with Large-scale Knowledge Graphs via Contrastive Learning , booktitle =. 2026 , url =. doi:10.1609/AAAI.V40I16.38370 , timestamp =

    doi:10.1609/aaai.v40i16.38370 2026

  56. [56]

    CoRR , volume =

    Hongkuan Zhou and Lavdim Halilaj and Sebastian Monka and Stefan Schmid and Yuqicheng Zhu and Bo Xiong and Steffen Staab , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2410.15981 , eprinttype =. 2410.15981 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.15981 2024

  57. [57]

    The Thirteenth International Conference on Learning Representations,

    Shi Yu and Chaoyue Tang and Bokai Xu and Junbo Cui and Junhao Ran and Yukun Yan and Zhenghao Liu and Shuo Wang and Xu Han and Zhiyuan Liu and Maosong Sun , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  58. [58]

    Prompt caching , year =

  59. [59]

    Hongyu Ren and Weihua Hu and Jure Leskovec , title =

  60. [60]

    Erik Arakelyan and Daniel Daza and Pasquale Minervini and Michael Cochez , title =

  61. [61]

    Can Pattern Learning Enhance Complex Logical Query Answering? , author=

  62. [62]

    Proceedings of the ACM on Web Conference 2025 , pages=

    Dage: Dag query answering via relational combinator with logical constraints , author=. Proceedings of the ACM on Web Conference 2025 , pages=

    2025

  63. [63]

    He, Yunjie and Hernandez, Daniel and Nayyeri, Mojtaba and Xiong, Bo and Zhu, Yuqicheng and Kharlamov, Evgeny and Staab, Steffen , title =

  64. [64]

    Predictive Multiplicity of Knowledge Graph Embeddings in Link Prediction

    Zhu, Yuqicheng and Potyka, Nico and Nayyeri, Mojtaba and Xiong, Bo and He, Yunjie and Kharlamov, Evgeny and Staab, Steffen. Predictive Multiplicity of Knowledge Graph Embeddings in Link Prediction. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024

    2024

  65. [65]

    Conformalized Answer Set Prediction for Knowledge Graph Embedding , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  66. [66]

    Predicate-Conditional Conformalized Answer Sets for Knowledge Graph Embeddings

    Zhu, Yuqicheng and Hern \'a ndez, Daniel and He, Yuan and Ding, Zifeng and Xiong, Bo and Kharlamov, Evgeny and Staab, Steffen. Predicate-Conditional Conformalized Answer Sets for Knowledge Graph Embeddings. Findings of the Association for Computational Linguistics: ACL 2025. 2025

    2025

  67. [67]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Certainty in Uncertainty: Reasoning over Uncertain Knowledge Graphs with Statistical Guarantees , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  68. [68]

    Thesis Proposal: Uncertainty in Knowledge Graph Embeddings , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Workshop) , pages=

    2025

  69. [69]

    , author=

    Towards Statistical Reasoning with Ontology Embeddings. , author=

  70. [70]

    arXiv preprint arXiv:2407.11821 , year=

    Approximating probabilistic inference in statistical el with knowledge graph embeddings , author=. arXiv preprint arXiv:2407.11821 , year=

    Pith/arXiv arXiv

  71. [71]

    Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems , pages=

    Robust Knowledge Extraction from Large Language Models using Social Choice Theory , author=. Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems , pages=

  72. [72]

    Second Conference on Language Modeling , year=

    Supposedly Equivalent Facts That Aren’t? Entity Frequency in Pre-training Induces Asymmetry in LLMs , author=. Second Conference on Language Modeling , year=

  73. [73]

    Conference on Neurosymbolic Learning and Reasoning , pages=

    ArgRAG: Explainable Retrieval Augmented Generation using Quantitative Bipolar Argumentation , author=. Conference on Neurosymbolic Learning and Reasoning , pages=. 2025 , organization=

    2025