Pith. sign in

REVIEW 3 major objections 4 minor 86 references

An 8B model, trained from SPARQL-scaffolded teacher traces with reinforcement learning, outperforms frozen frontier LLMs on three knowledge-graph question answering benchmarks.

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 ↗

An 8B model trained by supervised imitation of a SPARQL-guided teacher plus GRPO reinforcement learning outperforms all compared frozen frontier-LLM systems on WebQSP, CWQ, and GrailQA.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A well-engineered 8B KG navigator whose main results hold up under oracle linking, but the paper must fix internal number conflicts and clearly scope the headline claim. the 3 major comments →

arxiv 2607.18481 v1 pith:P5RBN7LI submitted 2026-07-20 cs.CL cs.IR

Search-on-Graph-R1: Training Large Language Models to Search Knowledge Graphs with Reinforcement Learning

Jia Ao Sun , Hao Yu , Fengran Mo , Zhan Su , Yuchen Hui , Bang Liu , Jian-Yun Nie This is my paper
classification cs.CL cs.IR
keywords knowledge graph question answeringreinforcement learningGRPOSPARQL scaffoldingtool-use trajectoriesLLM post-trainingFreebasemulti-hop reasoning
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 reading

Search-on-Graph-R1 (SOG-R1) aims to show that knowledge-graph navigation, normally done by prompting a frontier LLM to call a retrieval tool, can be compressed into an 8B model's weights. The key move is to scaffold a teacher with each question's gold SPARQL query, so the teacher walks a known answer-bearing path through a live knowledge graph and produces tool-use trajectories that are grounded by construction. The student is then fine-tuned on those trajectories and refined with GRPO under an exact-match reward plus a mild turn-count penalty. On WebQSP, CWQ, and GrailQA, the resulting 8B agent surpasses every frozen frontier-LLM system in the comparison and posts the strongest CWQ accuracy among all compared systems, with no auxiliary module at inference and no LLM judge during training. The comparisons assume oracle topic-entity linking, so the claim is about navigation after the topic entities are already identified.

Core claim

The paper claims that a compact model can internalize live KG traversal rather than delegating it to a frontier model or a static subgraph. Gold SPARQL queries, present in every KGQA training set, are used not as student supervision but as a scaffold that tells a teacher LLM which path to walk; the teacher's executed Search calls against a live Freebase server become training trajectories. Supervised fine-tuning on these trajectories cold-starts navigation, and GRPO with a deterministic outcome reward (format gate * exact match * turn-count factor) refines it, teaching recovery from wrong turns and compression of path length. In the paper's experiments, the 8B SOG-R1 exceeds all frozen front

What carries the argument

The load-bearing mechanism is a two-stage post-training pipeline. First, SPARQL-scaffolded trace generation: a teacher is shown the gold SPARQL query for a question and asked to translate it into calls to a single batched 1-hop neighbour retrieval tool, Search, which executes live against Freebase and returns tables of adjacent triples. This turns the dataset's hidden answer path into grounded, multi-turn teacher trajectories. Second, student post-training: LoRA SFT on those trajectories, then GRPO with a reward that multiplicatively combines a binary format gate (must call the tool and finish cleanly), a binary exact-match reward against the gold answer, and a floored turn-count penalty tha

Load-bearing premise

The assumption that each question's topic entities are already correctly linked to graph nodes: all three benchmarks are evaluated with oracle entity IDs supplied, and the paper's own limitations say performance under self-linking is presumed lower.

What would settle it

Run SOG-R1's evaluation with an automatic entity linker replacing the oracle topic-entity IDs on WebQSP, CWQ, or GrailQA. If exact-match accuracy drops substantially — toward the no-KG floor or below the frozen frontier baselines — then the paper's headline comparison measures oracle-assisted navigation rather than end-to-end KGQA. A second check: apply the pipeline to a knowledge graph whose training questions lack gold SPARQL queries (e.g., a Wikidata-based benchmark) and see whether the cold-start stage still produces grounded trajectories.

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

If this is right

  • If the central claim holds, a single compact 8B policy can replace a frontier-LLM-plus-orchestration stack for KGQA, cutting inference cost and latency while matching or exceeding accuracy.
  • The SFT and RL stages contribute complementary gains: SFT provides the navigation prior; RL adds recovery and compression, with RL alone underperforming SFT alone.
  • RL learns to reach answers in fewer Search calls than its SFT initialization on all three benchmarks, so efficiency pressure need not come at the cost of correctness.
  • The pipeline needs only a few thousand annotated questions: RL compensates for small cold-start pools and the curve flattens once the pool reaches roughly 8,700 questions.
  • The recipe transfers across model families and scales down to a 4B model, with the same ordering Base < SFT-only < full pipeline.

Where Pith is reading between the lines

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

  • An end-to-end version that performs its own entity linking would be the real test; the paper's oracle topic-entity IDs mean the reported numbers measure navigation skill, not the full question-answering chain.
  • The same SPARQL-scaffolding trick should apply wherever a training set ships with executable queries that encode a traversal — SQL over relational databases is the obvious neighbour — but the paper does not test that.
  • Because the reward is deterministic and needs no LLM judge, the cost of scaling RL to new graphs is mainly the annotation of gold queries; on graphs without them, the cold-start stage would need another source of grounded trajectories.
  • The batched Search tool's effect on dropped entities could be measured in isolation: comparing single-entity calls against batched calls with the same training data would show how much of the gain comes from this interface change.
Share X Bluesky LinkedIn Reddit HN

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 SOG-R1, a compact 8B KGQA agent trained in three stages: (1) a frontier teacher (Qwen3-235B) is scaffolded with each training question's gold SPARQL query and asked to traverse the answer path using a batched 1-hop Search tool over a live Freebase server, producing grounded multi-turn trajectories; (2) the student (Llama-3.1-8B-Instruct) is LoRA fine-tuned on these trajectories (SOG-SFT); (3) the student is refined with GRPO under a deterministic reward combining an exact-match answer term, a format gate, and a turn-count efficiency factor (SOG-R1). At inference the student receives the question plus oracle topic-entity IDs and interleaves Search calls until it emits a final answer. On WebQSP, CWQ, and GrailQA, the paper reports 91.7/88.7/90.0 exact match, which surpasses all frozen frontier-LLM baselines in the comparison and gives the strongest CWQ result among all compared systems. Additional experiments cover stage ablations, path compression, data-scale effects, and transfer to a second backbone (Qwen3-4B).

Significance. If the results hold, the paper makes a substantial practical and methodological contribution: it demonstrates that a compact 8B model can internalize a live knowledge-graph search policy that is competitive with or better than frozen GPT-4-class orchestrators on standard KGQA benchmarks, with no LLM judge in the reward loop. The SPARQL-scaffolded teacher stage is a simple and credible solution to the cold-start problem for tool-using KG agents, and the choice of a deterministic outcome reward is principled. The paper also provides useful ablations, a data-scale study, and cross-family transfer evidence. However, the manuscript currently contains contradictory ablation numbers and a pool-size inconsistency that must be resolved before the stage-contribution and data-scale claims can be taken at face value. The oracle topic-entity linking assumption is acknowledged in Limitations but qualifies the headline 'no auxiliary module' and 'surpasses every frozen frontier-LLM system' claims; the paper should either test self-linking or state this condition more prominently.

major comments (3)
  1. [§5.3, Table 3, Table 5, §5.4] The stage-ablation numbers are internally contradictory. Table 3 reports Base 50.9/35.3/55.5, SOG-SFT 89.6/86.1/88.8, and SOG-R1-ZERO 80.9/63.0/60.7, while Table 5 and the §5.3 text report Base 49.2/32.2/54.1 and SOG-SFT 90.9/86.3/88.8, and the text gives SOG-R1-ZERO as 81.9/63.0/60.0. Section 5.4's RL uplift of +0.8/+2.4/+1.2 is consistent only with the Table 5 SOG-SFT values; using Table 3 would give +2.1/+2.6/+1.2. The stage ablation is a central contribution, so the authors must correct the tables/text and state which set of numbers is final.
  2. [§5.5, Table 1, §5.1] Table 1 and §5.1 state that the GrailQA training pool is 2,700 questions, yet §5.5 says 'GrailQA is omitted; its EM was unchanged from 2,700 to 8,700 questions.' This implies that an 8,700-question GrailQA pool was used in an experiment not listed in Table 1, and it also raises uncertainty about which pool size the main GrailQA configuration actually used. Please reconcile the pool counts and state explicitly whether any GrailQA run used more than 2,700 questions.
  3. [§3, Algorithm 1, Limitations] The paper's central claim is conditioned on oracle topic-entity linking. Section 3 and Algorithm 1 define the input as (q, Eq) with Eq gold entity IDs, and the Limitations concede that 'performance under self-linking is presumed to be lower.' The comparison is internally fair because the baselines are also evaluated under standard linking conventions, but the abstract's phrase 'no auxiliary module at inference' and the headline 'surpasses every frozen frontier-LLM system' are narrower in practice. The oracle interface is load-bearing for the deployment claim: if self-linking is nontrivial on these benchmarks, the margin over frozen systems could shrink or reverse. Please add a self-linking experiment with a standard EL tool, or explicitly qualify the abstract and conclusion to state that all results assume gold topic-entity linking.
minor comments (4)
  1. [Table 3 caption] The caption says results are 'mean ± standard deviation over three seeds,' but the table shows only point estimates with no standard deviations. Either include the deviations or revise the caption.
  2. [§5.3, Table 3] The text reports SOG-R1-ZERO as 81.9/63.0/60.0 while Table 3 gives 80.9/63.0/60.7; this is part of the same inconsistency as major comment 1 and should be fixed in the same pass.
  3. [Abstract] Minor typo: 'fewerSearchcalls' should be 'fewer Search calls.'
  4. [Figure 2] The data-scale plot has no error bars or variance information. Since other results claim three-seed runs, it would be helpful to report variance for the data-scale points as well.

Circularity Check

1 steps flagged

Mostly self-contained; minor reward-encoded 'path compression' observation is disclosed and not load-bearing.

specific steps
  1. fitted input called prediction [§4.4 (Reward) and §5.4 (Path Compression)]
    "The turn-count factor penalizes verbosity through a floored linear penalty on excess turns, f(n_t) = max(f_min, 1 − λ·max(0, n_t − T_d)) ... since rem = 0 on any incorrect trajectory, f(n_t) acts as a tie-breaker that prefers shorter paths without ever losing correctness. ... A central claim of SOG-R1 is that the RL stage teaches the agent to navigate the KG more concisely than its SFT initialization."

    The compression reported in §5.4 is the exact quantity the reward's f(n_t) term is designed to reduce: GRPO maximizes r(τ) = g(τ)·r_em(τ)·f(n_t), where f(n_t) penalizes correct rollouts whose tool-call count exceeds T_d, and T_d is per-dataset set to the 75th percentile of teacher-trajectory turn counts. Thus the observed drop from SOG-SFT to SOG-R1 in average Search calls is a consequence of the optimization objective, not an independent discovery. The paper is transparent about this design, the accuracy gain confirms correctness is not abandoned, and the finding does not support the headline held-out accuracy comparison, so the circularity is minor and non-load-bearing.

full rationale

The central empirical claims — SOG-R1 outperforming frozen frontier-LLM systems and post-trained baselines on WebQSP, CWQ, and GrailQA — are evaluated on held-out test splits with no test-label fitting, and the teacher-scaffolding procedure is a legitimate distillation pipeline rather than a derivation whose output equals its input. The reliance on oracle topic-entity linking, disclosed in Section 3 and the Limitations ('the student receives the linked entity IDs E_q rather than performing entity linking itself'), narrows the practical scope but is an evaluation convention shared with the baselines, not a circular step. The only candidate circularity is the path-compression result, because the turn-count efficiency factor is an explicit reward component and T_d is fit to teacher trajectories; however, the paper openly states this reward design, the measured effect is consistent with the objective rather than tautologically equal to it, and it is secondary to the main accuracy results. No load-bearing self-citation or imported uniqueness claim is present. Overall the paper is self-contained with a minor, disclosed reward-conditioned observation, warranting a low score.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central performance claim rests on oracle entity linking and gold-SPARQL availability at training time, plus the assumption that live Freebase 1-hop queries are complete enough to ground the trajectories. The RL reward adds only a few genuinely free constants (lambda, f_min, T_d); the rest of the recipe is standard supervised/RL training. No new entities are postulated.

free parameters (4)
  • Turn-count penalty rate lambda = 0.05
    Fixed at a 'conservative' value with no tuning (Appendix B); part of the RL reward f(n_t) = max(fmin, 1 - lambda*max(0, n_t - T_d)). Affects the margin between correct rollouts, not the correctness advantage.
  • Reward floor f_min = 0.3
    Caps the maximum penalty in f(n_t); fixed, not tuned.
  • Free-zone threshold T_d = 75th-percentile turn count of teacher trajectories (per dataset)
    Set per dataset from teacher-trajectory turn counts (Section 4.4). A data-dependent threshold that influences the efficiency signal.
  • LoRA rank / alpha = r=64, alpha=128
    SFT adapter hyperparameters chosen for the 8B backbone; hand-chosen, not fitted to test data.
axioms (5)
  • domain assumption The topic entities Eq in each question are correctly linked to KG nodes (oracle entity linking) and provided to the agent at train and test time.
    Stated in Section 3 and in Limitations: 'the student receives the linked entity IDs Eq rather than performing entity linking itself, and performance under self-linking is presumed to be lower.' All benchmark claims inherit this assumption.
  • domain assumption Every KGQA training question supplies a gold SPARQL query that encodes an answer-bearing traversal path.
    Section 4.2: 'every training question comes with a gold SPARQL query.' True of WebQSP/CWQ/GrailQA but not all KGQA settings; the paper notes Wikidata is untested.
  • domain assumption The teacher (Qwen3-235B) with gold SPARQL and the live Search tool produces trajectories that, after gold-answer filtering, are sufficient cold-start supervision for the 8B student.
    The cold-start stage rests on this; retention rates 99.6/96.0/87.2% are given, but trace sufficiency is established only empirically on the three benchmarks.
  • domain assumption Live Freebase Virtuoso queries return correct, complete 1-hop results for the Search tool, and the 50-row truncation with paged property summaries does not systematically remove answer-bearing edges.
    The groundedness claim relies on this backend (Section 4.1, Appendix A); the completeness of Freebase under the subset queries is not independently audited.
  • domain assumption GRPO with a sparse binary exact-match reward plus the format/efficiency gates is a sufficient optimizer for the navigation policy.
    The training procedure is empirical; no guarantee beyond the reported runs that the reward landscape supports the stated convergence.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Search-on-Graph-R1: Training Large Language Models to Search Knowledge Graphs with Reinforcement Learning." pith.science (2026). https://pith.science/paper/P5RBN7LI

@misc{pith2026260718481,
  author       = {Pith},
  title        = {Pith review of: Search-on-Graph-R1: Training Large Language Models to Search Knowledge Graphs with Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P5RBN7LI}},
  note         = {Machine review of arXiv:2607.18481}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Knowledge graph question answering (KGQA) requires navigating from topic entities to an answer several relations away. Recent methods prompt a frontier LLM to explore the graph through a retrieval tool, but their reliance on frontier-scale inference makes them costly to deploy. We present Search-on-Graph-R1 (\sogrone{}), which internalizes this navigation into a compact 8B model through supervised fine-tuning (SFT) followed by reinforcement learning (RL). Our central idea is to scaffold a frontier teacher with each question's gold SPARQL query, so the teacher traverses a known answer-bearing path with a live \texttt{Search} tool rather than having to discover the path itself. Since every call executes against a live Freebase server, the resulting trajectories are grounded in the knowledge graph by construction. On WebQSP, CWQ, and GrailQA, \sogrone{} at 8B surpasses every frozen frontier-LLM system in our comparison and posts the strongest results on CWQ of any system we compare against. It does so using no auxiliary module at inference and no LLM judge during training. Isolating each training stage shows that SFT and RL contribute complementary gains, our approach transfers across model families, and RL learns to reach answers in fewer \texttt{Search} calls than its SFT initialization.

Figures

Figures reproduced from arXiv: 2607.18481 by Bang Liu, Fengran Mo, Hao Yu, Jia Ao Sun, Jian-Yun Nie, Yuchen Hui, Zhan Su.

Figure 1
Figure 1. Figure 1: Overview of SOG-R1. (A) SPARQL-scaffolded teacher trace generation. We walk through the CWQ question “What is the state capital where the major export is Apple trees?” (1) Its gold SPARQL query is shown to a 235B teacher LLM. (2) Using the Search tool, the teacher reads the query as a traversal blueprint and issues the tool calls needed to navigate the knowledge graph, executing each against a live Freebas… view at source ↗
Figure 2
Figure 2. Figure 2: Data-scale ablation on CWQ: test-set EM as the post-training pool grows, for the cold-start stage alone (SOG-SFT) and for the full SFT-then-RL pipeline (SOG-R1). Both stages draw from the same pool at ev￾ery size, and the backbone, tool, prompts, validation-set size, and all hyperparameters are held fixed. The 8,700 point is the main SOG-R1 configuration of [PITH_FULL_IMAGE:figures/full_fig_p008_2.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

86 extracted references · 8 canonical work pages

  1. [1]

    Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1247--1250

    2008

  2. [2]

    Liyi Chen, Panrong Tong, Zhongming Jin, Ying Sun, Jieping Ye, and Hui Xiong. 2024. Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs. Advances in Neural Information Processing Systems, 37:37665--37691

    2024

  3. [3]

    Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 6491--6506

    2021

  4. [4]

    Zixuan Dong, Baoyun Peng, Yufei Wang, Jia Fu, Xiaodong Wang, Xin Zhou, Yongxue Shan, Kangchen Zhu, and Weiguo Chen. 2025. https://aclanthology.org/2025.coling-main.479/ E ffi QA : Efficient question-answering with strategic multi-model collaboration on knowledge graphs . In Proceedings of the 31st International Conference on Computational Linguistics, pag...

    2025

  5. [5]

    Tengfei Feng and Liang He. 2025. https://aclanthology.org/2025.coling-main.205/ RGR - KBQA : Generating logical forms for question answering using knowledge-graph-enhanced large language model . In Proceedings of the 31st International Conference on Computational Linguistics, pages 3057--3070, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  6. [7]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. https://arxiv.org/abs/2407.21783 The llama 3...

    Pith/arXiv arXiv 2024

  7. [8]

    Yu Gu, Sue Kase, Michelle Vanni, Brian Sadler, Percy Liang, Xifeng Yan, and Yu Su. 2021. Beyond iid: three levels of generalization for question answering on knowledge bases. In Proceedings of the web conference 2021, pages 3477--3488

    2021

  8. [9]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  9. [10]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys, 55(12):1--38

    2023

  10. [12]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan O Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. https://openreview.net/forum?id=Rwhi91ideu Search-r1: Training LLM s to reason and leverage search engines with reinforcement learning . In Second Conference on Language Modeling

    2025

  11. [13]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles

    2023

  12. [19]

    Haoran Luo, Haihong E, Guanting Chen, Qika Lin, Yikai Guo, Fangzhi Xu, Zemin Kuang, Meina Song, Xiaobao Wu, Yifan Zhu, and Luu Anh Tuan. 2026. https://arxiv.org/abs/2507.21892 Graph-r1: Towards agentic graphrag framework via end-to-end reinforcement learning . Preprint, arXiv:2507.21892

    Pith/arXiv arXiv 2026

  13. [21]

    Linhao Luo, Yuan-Fang Li, Reza Haffari, and Shirui Pan. 2024 b . Reasoning on graphs: Faithful and interpretable large language model reasoning. In International Conference on Learning Representations, volume 2024, pages 14400--14423

    2024

  14. [22]

    Linhao Luo, Zicheng Zhao, Gholamreza Haffari, Yuan-Fang Li, Chen Gong, and Shirui Pan. 2025. https://proceedings.mlr.press/v267/luo25t.html Graph-constrained reasoning: Faithful reasoning on knowledge graphs with large language models . In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning R...

    2025

  15. [23]

    Jie Ma, Zhitao Gao, Qi Chai, Wangchun Sun, Pinghui Wang, Hongbin Pei, Jing Tao, Lingyun Song, Jun Liu, Chen Zhang, and 1 others. 2025. Debate on graph: a flexible and reliable reasoning framework for large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24768--24776

    2025

  16. [25]

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

    2024

  17. [26]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, dahai li, Zhiyuan Liu, and Maosong Sun. 2024. https://openreview.net/forum?id=dHng2O0Jjr Tool LLM : Facilitating large language models to master 16000+ real-world API s...

    2024

  18. [27]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. https://openreview.net/forum?id=Yacmpz84TH Toolformer: Language models can teach themselves to use tools . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  19. [28]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. https://arxiv.org/abs/2402.03300 Deepseekmath: Pushing the limits of mathematical reasoning in open language models . Preprint, arXiv:2402.03300

    Pith/arXiv arXiv 2024

  20. [29]

    Tiesunlong Shen, Jin Wang, Xuejie Zhang, and Erik Cambria. 2025. https://aclanthology.org/2025.coling-main.211/ Reasoning with trees: Faithful question answering over knowledge graph . In Proceedings of the 31st International Conference on Computational Linguistics, pages 3138--3157, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  21. [30]

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256

    Pith/arXiv arXiv 2024

  22. [31]

    Yuxin Shi, Han Fu, Zhuo Li, Chenghao Liu, Xiaoxue Ren, and Jianling Sun. 2026. https://openreview.net/forum?id=g6XnP7Sgui Plan-answer-refine-on-graph: Structured planning and self-refinement for large language model reasoning on knowledge graphs . In The Fourteenth International Conference on Learning Representations

    2026

  23. [34]

    Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. 2024. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In International Conference on Learning Representations, volume 2024, pages 3868--3898

    2024

  24. [39]

    Shuai Wang and Yinan Yu. 2026. https://arxiv.org/abs/2603.21440 Kg-hopper: Empowering compact open llms with knowledge graph reasoning via reinforcement learning . Preprint, arXiv:2603.21440

    Pith/arXiv arXiv 2026

  25. [41]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  26. [42]

    Shiqi Yan, Yubo Chen, Ruiqi Zhou, Zhengxi Yao, Shuai Chen, Tianyi Zhang, Shijie Zhang, Wei-Qiang Zhang, Yongfeng Huang, Haixin Duan, and Yunqi Zhang. 2026. https://openreview.net/forum?id=NfuBj8jleE Explore-on-graph: Incentivizing autonomous exploration of large language models on knowledge graphs with path-refined reward modeling . In The Fourteenth Inte...

    2026

  27. [46]

    Donghan Yu, Sheng Zhang, Patrick Ng, Henghui Zhu, Alexander Hanbo Li, Jun Wang, Yiqun Hu, William Yang Wang, Zhiguo Wang, and Bing Xiang. 2023. https://openreview.net/forum?id=XHc5zRPxqV9 Dec AF : Joint decoding of answers and logical forms for question answering over knowledge bases . In The Eleventh International Conference on Learning Representations

    2023

  28. [47]

    Zhiqiang Zhang, Liqiang Wen, and Wen Zhao. 2025. https://aclanthology.org/2025.coling-main.562/ Rule- KBQA : Rule-guided reasoning for complex knowledge base question answering with large language models . In Proceedings of the 31st International Conference on Computational Linguistics, pages 8399--8417, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  29. [48]

    Zhiqiang Zhang and Wen Zhao. 2025. https://aclanthology.org/2025.coling-main.712/ A collaborative reasoning framework powered by reinforcement learning and large language models for complex questions answering over knowledge graph . In Proceedings of the 31st International Conference on Computational Linguistics, pages 10672--10684, Abu Dhabi, UAE. Associ...

    2025

  30. [50]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  31. [51]

    Publications Manual , year = "1983", publisher =

    1983

  32. [52]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  33. [53]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  34. [54]

    Dan Gusfield , title =. 1997

    1997

  35. [55]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  36. [56]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  37. [57]

    C omp KBQA : Component-wise Task Decomposition for Knowledge Base Question Answering

    Tian, Yuhang and Song, Dandan and Wu, Zhijing and Yang, Pan and Zhou, Changzhi and Yang, Jun and Wang, Hao and Ma, Huipeng and Li, Chenhao and Zhang, Luan. C omp KBQA : Component-wise Task Decomposition for Knowledge Base Question Answering. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025....

    work page doi:10.18653/v1/2025.emnlp-main.16 2025

  38. [58]

    D - RAG : Differentiable Retrieval-Augmented Generation for Knowledge Graph Question Answering

    Gao, Guangze and Li, Zixuan and Yuan, Chunfeng and Li, Jiawei and Jianzhuo, Wu and Zhang, Yuehao and Jin, Xiaolong and Li, Bing and Hu, Weiming. D - RAG : Differentiable Retrieval-Augmented Generation for Knowledge Graph Question Answering. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.e...

    work page doi:10.18653/v1/2025.emnlp-main.1793 2025

  39. [59]

    i QUEST : An Iterative Question-Guided Framework for Knowledge Base Question Answering

    Wang, Shuai and Yu, Yinan. i QUEST : An Iterative Question-Guided Framework for Knowledge Base Question Answering. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.760

    work page doi:10.18653/v1/2025.acl-long.760 2025

  40. [60]

    Digest the Knowledge: Large Language Models empowered Message Passing for Knowledge Graph Question Answering

    Wan, Junhong and Yu, Tao and Jiang, Kunyu and Fu, Yao and Jiang, Weihao and Zhu, Jiang. Digest the Knowledge: Large Language Models empowered Message Passing for Knowledge Graph Question Answering. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.750

    work page doi:10.18653/v1/2025.acl-long.750 2025

  41. [61]

    Ontology-Guided Reverse Thinking Makes Large Language Models Stronger on Knowledge Graph Question Answering

    Liu, Runxuan and Luo, Bei and Li, Jiaqi and Wang, Baoxin and Liu, Ming and Wu, Dayong and Wang, Shijin and Qin, Bing. Ontology-Guided Reverse Thinking Makes Large Language Models Stronger on Knowledge Graph Question Answering. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.1865...

    doi:10.18653/v1/2025.acl-long.741 2025

  42. [62]

    KG -Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph

    Jiang, Jinhao and Zhou, Kun and Zhao, Wayne Xin and Song, Yang and Zhu, Chen and Zhu, Hengshu and Wen, Ji-Rong. KG -Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.468

    doi:10.18653/v1/2025.acl-long.468 2025

  43. [63]

    RGR - KBQA : Generating Logical Forms for Question Answering Using Knowledge-Graph-Enhanced Large Language Model

    Feng, Tengfei and He, Liang. RGR - KBQA : Generating Logical Forms for Question Answering Using Knowledge-Graph-Enhanced Large Language Model. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  44. [64]

    C hat KBQA : A Generate-then-Retrieve Framework for Knowledge Base Question Answering with Fine-tuned Large Language Models

    Luo, Haoran and E, Haihong and Tang, Zichen and Peng, Shiyao and Guo, Yikai and Zhang, Wentai and Ma, Chenghao and Dong, Guanting and Song, Meina and Lin, Wei and Zhu, Yifan and Luu, Anh Tuan. C hat KBQA : A Generate-then-Retrieve Framework for Knowledge Base Question Answering with Fine-tuned Large Language Models. Findings of the Association for Computa...

    doi:10.18653/v1/2024.findings-acl.122 2024

  45. [65]

    The Fourteenth International Conference on Learning Representations , year=

    Explore-on-Graph: Incentivizing Autonomous Exploration of Large Language Models on Knowledge Graphs with Path-refined Reward Modeling , author=. The Fourteenth International Conference on Learning Representations , year=

  46. [66]

    International Conference on Learning Representations , volume=

    Reasoning on graphs: Faithful and interpretable large language model reasoning , author=. International Conference on Learning Representations , volume=

  47. [67]

    and Han, Zhen and Zhu, Qi and Robinson, Ian and Thompson, Bryan and Rangwala, Huzefa and Karypis, George

    Mavromatis, Costas and Adeshina, Soji and Ioannidis, Vassilis N. and Han, Zhen and Zhu, Qi and Robinson, Ian and Thompson, Bryan and Rangwala, Huzefa and Karypis, George. BYOKG - RAG : Multi-Strategy Graph Retrieval for Knowledge Graph Question Answering. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10....

    work page doi:10.18653/v1/2025.emnlp-main.1417 2025

  48. [68]

    arXiv preprint arXiv:2604.12487 , year=

    KG-Reasoner: A Reinforced Model for End-to-End Multi-Hop Knowledge Graph Reasoning , author=. arXiv preprint arXiv:2604.12487 , year=

    Pith/arXiv arXiv

  49. [69]

    International Conference on Learning Representations , volume=

    Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph , author=. International Conference on Learning Representations , volume=

  50. [70]

    The Fourteenth International Conference on Learning Representations , year=

    Plan-Answer-Refine-on-Graph: Structured Planning and Self-Refinement for Large Language Model Reasoning on Knowledge Graphs , author=. The Fourteenth International Conference on Learning Representations , year=

  51. [71]

    A Framework of Knowledge Graph-Enhanced Large Language Model Based on Question Decomposition and Atomic Retrieval

    Li, Yading and Song, Dandan and Zhou, Changzhi and Tian, Yuhang and Wang, Hao and Yang, Ziyi and Zhang, Shuhao. A Framework of Knowledge Graph-Enhanced Large Language Model Based on Question Decomposition and Atomic Retrieval. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.670

    doi:10.18653/v1/2024.findings-emnlp.670 2024

  52. [72]

    F i D e L i S : Faithful Reasoning in Large Language Models for Knowledge Graph Question Answering

    Sui, Yuan and He, Yufei and Liu, Nian and He, Xiaoxin and Wang, Kun and Hooi, Bryan. F i D e L i S : Faithful Reasoning in Large Language Models for Knowledge Graph Question Answering. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.436

    work page doi:10.18653/v1/2025.findings-acl.436 2025

  53. [73]

    and Meng, Helen M

    Li, Kun and Zhang, Tianhua and Wu, Xixin and Luo, Hongyin and Glass, James R. and Meng, Helen M. Decoding on Graphs: Faithful and Sound Reasoning on Knowledge Graphs through Generation of Well-Formed Chains. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1186

    work page doi:10.18653/v1/2025.acl-long.1186 2025

  54. [74]

    E ffi QA : Efficient Question-Answering with Strategic Multi-Model Collaboration on Knowledge Graphs

    Dong, Zixuan and Peng, Baoyun and Wang, Yufei and Fu, Jia and Wang, Xiaodong and Zhou, Xin and Shan, Yongxue and Zhu, Kangchen and Chen, Weiguo. E ffi QA : Efficient Question-Answering with Strategic Multi-Model Collaboration on Knowledge Graphs. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  55. [75]

    RJE : A Retrieval-Judgment-Exploration Framework for Efficient Knowledge Graph Question Answering with LLM s

    Lin, Can and Jiang, Zhengwang and Zheng, Ling and Zhao, Qi and Zhang, Yuhang and Song, Qi and Zhou, Wangqiu. RJE : A Retrieval-Judgment-Exploration Framework for Efficient Knowledge Graph Question Answering with LLM s. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.873

    work page doi:10.18653/v1/2025.emnlp-main.873 2025

  56. [76]

    A Collaborative Reasoning Framework Powered by Reinforcement Learning and Large Language Models for Complex Questions Answering over Knowledge Graph

    Zhang, Zhiqiang and Zhao, Wen. A Collaborative Reasoning Framework Powered by Reinforcement Learning and Large Language Models for Complex Questions Answering over Knowledge Graph. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  57. [77]

    Reasoning with Trees: Faithful Question Answering over Knowledge Graph

    Shen, Tiesunlong and Wang, Jin and Zhang, Xuejie and Cambria, Erik. Reasoning with Trees: Faithful Question Answering over Knowledge Graph. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  58. [78]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  59. [79]

    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=

  60. [80]

    arXiv preprint arXiv:2510.08825 , year=

    Search-on-Graph: Iterative Informed Navigation for Large Language Model Reasoning on Knowledge Graphs , author=. arXiv preprint arXiv:2510.08825 , year=

    Pith/arXiv arXiv

  61. [81]

    Donghan Yu and Sheng Zhang and Patrick Ng and Henghui Zhu and Alexander Hanbo Li and Jun Wang and Yiqun Hu and William Yang Wang and Zhiguo Wang and Bing Xiang , booktitle=. Dec. 2023 , url=

    2023

  62. [82]

    RNG - KBQA : Generation Augmented Iterative Ranking for Knowledge Base Question Answering

    Ye, Xi and Yavuz, Semih and Hashimoto, Kazuma and Zhou, Yingbo and Xiong, Caiming. RNG - KBQA : Generation Augmented Iterative Ranking for Knowledge Base Question Answering. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.417

    doi:10.18653/v1/2022.acl-long.417 2022

  63. [83]

    Search-R1: Training

    Bowen Jin and Hansi Zeng and Zhenrui Yue and Jinsung Yoon and Sercan O Arik and Dong Wang and Hamed Zamani and Jiawei Han , booktitle=. Search-R1: Training. 2025 , url=

    2025

  64. [84]

    Rule- KBQA : Rule-Guided Reasoning for Complex Knowledge Base Question Answering with Large Language Models

    Zhang, Zhiqiang and Wen, Liqiang and Zhao, Wen. Rule- KBQA : Rule-Guided Reasoning for Complex Knowledge Base Question Answering with Large Language Models. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  65. [85]

    The Web as a Knowledge-Base for Answering Complex Questions

    Talmor, Alon and Berant, Jonathan. The Web as a Knowledge-Base for Answering Complex Questions. Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018. doi:10.18653/v1/N18-1059

    doi:10.18653/v1/n18-1059 2018

  66. [86]

    The Value of Semantic Parse Labeling for Knowledge Base Question Answering

    Yih, Wen-tau and Richardson, Matthew and Meek, Chris and Chang, Ming-Wei and Suh, Jina. The Value of Semantic Parse Labeling for Knowledge Base Question Answering. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2016. doi:10.18653/v1/P16-2033

    doi:10.18653/v1/p16-2033 2016

  67. [87]

    Proceedings of the web conference 2021 , pages=

    Beyond iid: three levels of generalization for question answering on knowledge bases , author=. Proceedings of the web conference 2021 , pages=

    2021

  68. [88]

    Few-shot In-context Learning on Knowledge Base Question Answering

    Li, Tianle and Ma, Xueguang and Zhuang, Alex and Gu, Yu and Su, Yu and Chen, Wenhu. Few-shot In-context Learning on Knowledge Base Question Answering. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.385

    doi:10.18653/v1/2023.acl-long.385 2023

  69. [89]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Toolformer: Language Models Can Teach Themselves to Use Tools , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  70. [90]

    Yujia Qin and Shihao Liang and Yining Ye and Kunlun Zhu and Lan Yan and Yaxi Lu and Yankai Lin and Xin Cong and Xiangru Tang and Bill Qian and Sihan Zhao and Lauren Hong and Runchu Tian and Ruobing Xie and Jie Zhou and Mark Gerstein and dahai li and Zhiyuan Liu and Maosong Sun , booktitle=. Tool. 2024 , url=

    2024

  71. [91]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  72. [92]

    Edward J Hu and yelong shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    2022

  73. [93]

    L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2024. doi:10.18653/v1/2024.acl-demos.38

    doi:10.18653/v1/2024.acl-demos.38 2024

  74. [94]

    2024 , journal =

    HybridFlow: A Flexible and Efficient RLHF Framework , author =. 2024 , journal =

    2024

  75. [95]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  76. [96]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  77. [97]

    Proceedings of the 2008 ACM SIGMOD international conference on Management of data , pages=

    Freebase: a collaboratively created graph database for structuring human knowledge , author=. Proceedings of the 2008 ACM SIGMOD international conference on Management of data , pages=

    2008

  78. [98]

    ACM computing surveys , volume=

    Survey of hallucination in natural language generation , author=. ACM computing surveys , volume=. 2023 , publisher=

    2023

  79. [99]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    Editing factual knowledge in language models , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  80. [100]

    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

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.