Pith. sign in

REVIEW 4 major objections 5 minor 55 references

The paper claims JERR, a synopsis-plus-graph-plus-MCTS pipeline, beats every baseline on three long-context QA 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 ↗

JERR builds a directed graph of summarized text chunks and uses Monte Carlo Tree Search to retrieve relevant nodes, reportedly improving long-context QA over baselines.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A reasonable training-free pipeline with a real but unproven headline claim; the evaluation needs held-out tuning and significance tests before the numbers are convincing. the 4 major comments →

arxiv 2508.20351 v1 pith:77IQM5JO submitted 2025-08-28 cs.CL

Joint Enhancement of Relational Reasoning for Long-Context LLMs

Zhirui Chen , Wei Shen , Jiashui Huang , Ling Shao This is my paper
classification cs.CL
keywords long-context LLMsrelational reasoningMonte Carlo Tree Searchgraph-based question answeringsynopsis extractionentity deduplicationdirected acyclic graphretrieval-augmented generation
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

JERR's claim is that long-context question answering improves when the LLM does not read the raw document at all, but works from a compressed and linked version of it: each chunk is reduced to a synopsis, the entities in those synopses become nodes of a directed graph whose edges name relationships, and Monte Carlo Tree Search picks the handful of nodes the model should actually consider. The paper reports this pipeline as the best performer on all three of its benchmarks — 86.39% accuracy on QuALITY, F1 of 0.505 on MuSiQue, F1 of 0.269 on NarrativeQA — ahead of retrieval-augmented reading, direct long-context prompting, and earlier agent-based readers. The payoff, if the claim holds, is a training-free way to extend an LLM's effective reach, with an answer trail made of the graph nodes the model actually used — a form of transparency plain long-context reading does not offer. The paper's own limitations, stated in Section 6, are that evidence covers only three datasets and that generalization, graph-building cost, and performance on simpler tasks are not yet established.

Core claim

JERR's central claim is that long-context QA improves when the document is rebuilt as a searchable graph: chunks become synopses, their entities become deduplicated nodes linked by relational attributes in a directed acyclic graph (no cycles), and Monte Carlo Tree Search selects the top-k nodes the LLM should read before answering. Reported results put JERR highest on every metric in its comparisons — 86.39% on QuALITY, F1 0.505 on MuSiQue, F1 0.269 on NarrativeQA — ahead of RAG variants, direct long-context reading, ReadAgent, LongRAG, GraphRAG, and GraphReader on GPT-4. The paper credits MCTS for the gain (PageRank in its place costs about five points on QuALITY) and claims the chosen grap

What carries the argument

The load-bearing object is a directed acyclic graph (edges point one way, no path loops back): nodes are deduplicated entities annotated with their chunk synopsis; edges are prompt-extracted relational attributes such as 'causes' or 'belongs to'. Paired with it is Monte Carlo Tree Search (MCTS), which from a root node repeatedly selects children by an exploration-exploitation score, expands to neighbors, simulates paths whose reward is keyword overlap with the question, and back-propagates the reward; the top-k ranked nodes (k=5) join the synopsis in the final answer prompt. The DAG is what the paper credits for efficient, cycle-free traversal; the ablation shows the search carries the gain,

Load-bearing premise

The framework's efficiency and its 'avoids cycles' property rest on the constructed graph being acyclic, but Eq. (5) joins any two entities in a chunk that share a relational attribute, with no topological ordering or cycle check, so acyclicity is assumed rather than guaranteed.

What would settle it

Run a standard cycle-detection pass on the graphs JERR builds for, say, the 200 MuSiQue documents; a single directed cycle would falsify the DAG premise that the traversal's efficiency rests on. Separately, measure the stability of JERR's margins over GraphRAG and ReadAgent across repeated runs (or seeds): the reported edges — roughly one point on QuALITY and a few hundredths of F1 elsewhere — are small enough that if they dissolve under reruns, the 'consistently outperforms' claim fails.

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

If this is right

  • Training-free reach: JERR extends what an off-the-shelf LLM can answer from a long document without retuning weights or enlarging the context window — the pipeline is prompts plus a graph copy of the text, and the paper demonstrates it on qwen-plus-128k and GPT-4.
  • One build, many questions: the graph is query-independent, so it can be constructed once and reused; the cost table reports per-question token use dropping from 98.54k to 44.33k when construction is skipped.
  • Search rule matters: with the same graph, swapping MCTS for PageRank lowers QuALITY accuracy from 86.39% to 81.69%.
  • Traceable answers: the nodes MCTS selects form a visible path of entity relationships, giving an answer trail that direct reading and RAG do not expose.
  • Transfer across base models: JERR is reported ahead of baselines with both qwen-plus-128k and GPT-4, indicating the wrapper's benefit is not tied to one LLM.

Where Pith is reading between the lines

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

  • A cycle-detection audit would test the DAG premise: Eq. (5) creates directed edges between any two entities in a chunk that share a relational attribute, without a topological ordering or cycle check, so mutually referencing entities can produce cycles; adding a topological-ordering or cycle-pruning step would protect the traversal-efficiency claim.
  • The keyword-overlap reward in MCTS simulation (Eq. 11) is the fragile link: questions whose answers share no words with the question would give the search little signal, so embedding-based relevance is a natural, testable upgrade.
  • JERR's economics favor repeated use of the same corpus: the reported 44.33k tokens per question without graph reconstruction implies the framework pays off most when many questions are asked against one document collection — a regime the paper does not foreground.
  • The largest margins are on multi-hop MuSiQue, consistent with edge structure doing the reasoning work; a stress test on documents beyond the tested 16k–64k token range would show where graph-based search saturates.
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

4 major / 5 minor

Summary. The paper proposes JERR, a training-free agent for long-context question answering that combines three stages: (i) chunking and synopsis extraction via LLM prompts, (ii) construction of a directed acyclic graph (DAG) whose nodes are deduplicated entities and whose edges are relational attributes, and (iii) Monte Carlo Tree Search (MCTS) over this graph to retrieve top-k nodes relevant to a query. The retrieved nodes and synopses are fed to an LLM to generate answers. Experiments on QuALITY, MuSiQue, and NarrativeQA with Qwen-Plus-128k (and a GPT-4-based comparison in §4.5) report that JERR outperforms RAG, long-context LLM, and agent baselines on accuracy, ROUGE, F1, and LLM-Rater metrics.

Significance. If the empirical claims hold, JERR offers a modular and interpretable graph-based retrieval approach that could be useful for long-context reasoning, and the inclusion of a cost analysis and a GPT-4 robustness check are valuable. The paper explicitly positions itself as a training-free framework, and the MCTS-based retrieval with graph construction is a reasonable design. However, the reported advantages over the strongest baselines are small (e.g., 1.37% accuracy on QuALITY, 0.017 F1 on MuSiQue, 0.015 F1 on NarrativeQA), and the experimental methodology has load-bearing weaknesses that currently prevent the central claim of 'consistent outperformance' from being established. The contribution is potentially significant but requires a substantially strengthened empirical validation.

major comments (4)
  1. [§4.3 and §4.1 (hyperparameter selection)] The ablation study in §4.3 uses the same benchmarks that produce the headline results: top-k is set to 5 based on Figure 2 on QuALITY, chunk size 2k is chosen from Figure 3 on NarrativeQA, and chunk size 600 is chosen from Figure 5 on QuALITY. These values are then applied in Tables 2–3 (see §4.1 Implementation Details). No held-out validation or nested model selection is described. Given that the reported margins over the best baselines are small (e.g., QuALITY 86.39 vs 85.02, MuSiQue F1 0.505 vs 0.488, NarrativeQA F1 0.269 vs 0.254), the 'consistent outperformance' claim could be an artifact of test-set tuning. The authors should re-run evaluations with hyperparameters chosen on a separate validation split or report sensitivity across datasets.
  2. [§4 Tables 2–3 (statistical reliability)] The paper reports no error bars, confidence intervals, or significance tests for any metric. Since the differences between JERR and the next-best baseline are often in the range of 0.01–0.02 F1 or 1–2% accuracy, it is impossible to determine whether these differences are real or within run-to-run noise. The authors should provide multiple runs with different random seeds and report variance, or apply paired bootstrap / significance testing (e.g., per-sample paired tests) to support the central claim of consistent improvement.
  3. [§3.3, Eq. (5) and Appendix D (DAG acyclicity)] The paper defines a DAG and relies on acyclicity for efficient traversal and to 'inherently avoid cycles' (Appendix D). However, Eq. (5) adds a directed edge for every pair of entities in a chunk that has any relational attribute, without any topological ordering or cycle check. In natural language, mutually referring entities will produce cycles, so the constructed graph is not guaranteed to be a DAG. This is an internal inconsistency: either the graph construction must enforce acyclicity (e.g., by ordering nodes or pruning back edges), or the DAG-based claims and the MCTS 'avoids cycles' property must be revised. This does not invalidate the empirical numbers, but it undermines a stated component of the method.
  4. [Appendix B, Algorithm 1 vs. §3.4 (MCTS description)] The MCTS description is internally inconsistent. §3.4 states that simulation computes scores by evaluating keyword matches along the simulated path. Appendix B says 'a random simulation is conducted' and Eq. (11) defines reward as a sum of keyword-overlap counts over the simulated trajectory. However, Algorithm 1 (lines 15–28) does not perform a random simulation; it repeatedly moves to the 'best child' by win/visit ratio, and the score is incremented for a node before moving to the next node, with no random rollouts. This discrepancy concerns the core retrieval mechanism and should be resolved: either the algorithm should match the description (random simulation) or the text should describe the actual deterministic greedy rollout used in the implementation.
minor comments (5)
  1. [References] Several references are incomplete or placeholders, e.g., 'Okapi BM25. 2020. Title of the webpage.' and 'Author Name qwen-plus 128k. 2024.' These need full bibliographic details.
  2. [§4.5, Table 5] The table appears to use percentages (e.g., 60.5) while Table 3 uses ratios (0.455). The text should state the unit/scale and clarify that these are aligned with the GraphReader paper's presentation. Also, GraphReader is excluded from the main experiments due to 'absence of public source code' but appears in Table 5; this discrepancy should be explained.
  3. [§3.2 and general notation] Equation (2) uses an unusual union notation ('n U i=1') and the paper should define the operation clearly. Similarly, Eq. (3) uses 'c = ddpexact(sj)' inside a set definition in a way that is confusing; please clarify.
  4. [Implementation details] The paper does not specify several MCTS parameters that are part of the method, such as the exploration coefficient κ, the number of simulations N, and the simulation depth d. These should be reported for reproducibility.
  5. [Throughout] Minor typos: 'MusiQUE' in §4.5, 'MTCS' in §5, and inconsistent hyphenation of 'top-k'. A careful proofreading pass is recommended.

Circularity Check

0 steps flagged

No significant circularity: JERR's empirical comparisons rest on an experimental pipeline, not on a self-referential derivation or fitted parameters renamed as predictions.

full rationale

The paper's central claim — that JERR outperforms baselines on ROUGE, F1, and LLM-Rater metrics — is an empirical result produced by running a pipeline (chunking, synopsis extraction, graph construction, MCTS retrieval, and prompted generation) on benchmark datasets. No equation in the paper defines a predicted metric in terms of a fitted parameter, and no benchmark number is shown to equal its own input by construction. The MCTS reward is keyword overlap (Eq. 11), not a function fit to the target metric. Hyperparameters such as top-k=5 and chunk sizes are selected via ablations on QuALITY/NarrativeQA (Section 4.3, Figures 2, 3, 5) and then used in the headline comparisons; this is a test-set-tuning / overfitting concern about experimental validity, not a circular reduction. The asserted DAG acyclicity (Eq. 5) is internally inconsistent because edges are added between any entity pair with a non-empty relational attribute without a topological check, but an incorrect structural assumption is a correctness risk, not circularity. There are no load-bearing self-citations, imported uniqueness theorems, or renamed known results. The paper's derivations are therefore not circular.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The framework's central claim rests on several unshown assumptions: the LLM prompts are reliable for extraction, keyword overlap approximates relevance, and the graph is actually a DAG. The main benchmark-specific choices (top-k, chunk size, SimHash threshold, MCTS parameters) are free parameters, several left unspecified.

free parameters (6)
  • top-k (relevant nodes) = 5 (default)
    Informed by ablation in Figure 2 on QuALITY; reported results use this value on the same benchmark.
  • chunk size = 600 tokens (QuALITY), 2000 tokens (MuSiQue/NarrativeQA)
    Tuned per dataset via ablations (Figures 3 and 5) on the same datasets used for evaluation.
  • SimHash similarity threshold θ = unspecified (text says e.g., 3 bits)
    Threshold for near-duplicate entity merging; value not reported.
  • MCTS exploration coefficient κ = unspecified
    Used in UCB in Eq. (8); value not given, affects search behavior.
  • MCTS simulation count N = unspecified
    Number of MCTS iterations; not reported, but controls retrieval quality.
  • MCTS simulation depth d = 10 (in Algorithm 1 only)
    Max depth in reward simulation; fixed in pseudocode but not reported as a tuned parameter.
axioms (4)
  • domain assumption Prompt-based LLM summarization and entity extraction produce reliable synopses and components.
    Sections 3.2 and 3.3 rely on LLM prompts (A.1-A.3) to distill chunks and extract entities, but no quality control is reported.
  • ad hoc to paper Keyword overlap between query and node text is a sufficient relevance signal for retrieval.
    Eq. (11) and Algorithm 1 define MCTS reward as the count of keyword matches; this assumes lexical overlap captures the semantic relevance needed for QA, which is not demonstrated.
  • ad hoc to paper The graph construction yields a DAG.
    Section 3.3 asserts DAG construction, but Eq. (5) adds edges without acyclicity constraints; this is an unflagged assumption the traversal depends on.
  • standard math Standard data structures and algorithms (Bloom Filter, Trie, SimHash, MCTS) behave as documented.
    Used in Section 3.3 and Appendix B; no new mathematical claims are made about these.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint Enhancement of Relational Reasoning for Long-Context LLMs." pith.science (2026). https://pith.science/paper/77IQM5JO

@misc{pith2026250820351,
  author       = {Pith},
  title        = {Pith review of: Joint Enhancement of Relational Reasoning for Long-Context LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/77IQM5JO}},
  note         = {Machine review of arXiv:2508.20351}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Despite significant progress, large language models (LLMs) still struggle with long contexts due to memory limitations and their inability to tackle complex and long-context tasks. Additionally, LLMs often suffer from a lack of transparency and are prone to producing hallucinations. To address these challenges, we propose \textbf{JERR}, a novel framework designed to enhance long-context comprehension via graph-based reasoning in LLMs. JERR integrates three key components: synopsis extraction, graph construction, and relational reasoning. First, synopsis is extracted by chunking text strategically, allowing the model to summarize and understand information more efficiently. Second, we build a directed acyclic graph (DAG) to resolve redundancy, ensuring logical consistency and clarity. Finally, we incorporate Monte Carlo Tree Search (MCTS) to help the model navigate complex reasoning paths, ensuring more accurate and interpretable outputs. This framework provides a novel solution that enables LLMs to handle extended contexts and complex reasoning tasks with improved reliability and transparency. Experimental results show that JERR consistently outperforms all baselines on the ROUGE and F1 metrics, achieving the highest scores on the LLM-Rater evaluation.

Figures

Figures reproduced from arXiv: 2508.20351 by Jiashui Huang, Ling Shao, Wei Shen, Zhirui Chen.

Figure 1
Figure 1. Figure 1: The overall framework of JERR consists of three main steps: 1) Given a question and a long context, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of JERR with different top-k [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance of JERR with different chunk [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Study of case from the second passage of QuALITY dev set with the corresponding question (1 out of 9) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance of JERR with different chunk [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

55 extracted references · 31 canonical work pages

  1. [1]

    Uthus, Mandy Guo, James Lee - Thorp, Yi Tay, Yun - Hsuan Sung, and Sumit Sanghai

    Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Onta \ n \' o n, Siddhartha Brahma, Yury Zemlyanskiy, David C. Uthus, Mandy Guo, James Lee - Thorp, Yi Tay, Yun - Hsuan Sung, and Sumit Sanghai. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.309 Colt5: Faster long-range transformers with conditional computation . In Proceedings of the 2023 Conference ...

    doi:10.18653/v1/2023.emnlp-main.309 2023

  2. [2]

    Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150

    Pith/arXiv arXiv 2020

  3. [3]

    Okapi BM25. 2020. Title of the webpage. https://github.com/dorianbrown/rank_bm25

    work page 2020

  4. [4]

    Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. 2012. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1--43

    work page 2012

  5. [5]

    Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, and Xiao Huang. 2024 a . http://papers.nips.cc/paper\_files/paper/2024/hash/1b57aaddf85ab01a2445a79c9edc1f4b-Abstract-Conference.html Entity alignment with noisy annotations from large language models . In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Inform...

    work page 2024

  6. [6]

    Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595

    Pith/arXiv arXiv 2023

  7. [7]

    Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. 2024 b . https://openreview.net/forum?id=6PmJoRfdaK Longlora: Efficient fine-tuning of long-context large language models . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

    work page 2024

  8. [8]

    Michiel de Jong, Yury Zemlyanskiy, Nicholas FitzGerald, Fei Sha, and William W. Cohen. 2022. https://openreview.net/forum?id=OY1A8ejQgEX Mention memory: incorporating textual knowledge into transformers through entity mention attention . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    work page 2022

  9. [9]

    Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. https://openreview.net/forum?id=r1l73iRqKm Wizard of wikipedia: Knowledge-powered conversational agents . In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net

    work page 2019

  10. [10]

    Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. 2024. https://openreview.net/forum?id=ONOtpXLqqw Longrope: Extending LLM context window beyond 2 million tokens . In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net

    work page 2024

  11. [11]

    Guanting Dong, Rumei Li, Sirui Wang, Yupeng Zhang, Yunsen Xian, and Weiran Xu. 2023. Bridging the kb-text gap: Leveraging structured knowledge-aware pre-training for kbqa. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 3854--3859

    work page 2023

  12. [12]

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

    Pith/arXiv arXiv 2024

  13. [13]

    Thibault F \' e vry, Livio Baldini Soares, Nicholas FitzGerald, Eunsol Choi, and Tom Kwiatkowski. 2020. https://doi.org/10.18653/V1/2020.EMNLP-MAIN.400 Entities as experts: Sparse memory access with entity supervision . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020 , pag...

    work page doi:10.18653/v1/2020.emnlp-main.400 2020

  14. [14]

    Kanishk Gandhi, Jan-Philipp Fr \"a nken, Tobias Gerstenberg, and Noah Goodman. 2024. Understanding social reasoning in language models with language models. Advances in Neural Information Processing Systems, 36

    work page 2024

  15. [15]

    Zitian Gao, Boye Niu, Xuzheng He, Haotian Xu, Hongzhang Liu, Aiwei Liu, Xuming Hu, and Lijie Wen. 2024. Interpretable contrastive monte carlo tree search reasoning. arXiv preprint arXiv:2410.01707

    Pith/arXiv arXiv 2024

  16. [16]

    Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. 2024 a . Lm-infinite: Zero-shot extreme length generalization for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3991--4008

    work page 2024

  17. [17]

    Buntine, and Ehsan Shareghi

    Jiuzhou Han, Nigel Collier, Wray L. Buntine, and Ehsan Shareghi. 2024 b . https://doi.org/10.18653/V1/2024.FINDINGS-ACL.400 Pive: Prompting with iterative verification improving graph-based generative capability of llms . In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 , pag...

    work page doi:10.18653/v1/2024.findings-acl.400 2024

  18. [18]

    Linmei Hu, Zeyi Liu, Ziwang Zhao, Lei Hou, Liqiang Nie, and Juanzi Li. 2023. A survey of knowledge enhanced pre-trained language models. IEEE Transactions on Knowledge and Data Engineering

    work page 2023

  19. [19]

    Wenyu Huang, Guancheng Zhou, Hongru Wang, Pavlos Vougiouklis, Mirella Lapata, and Jeff Z. Pan. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-EMNLP.927 Less is more: Making smaller language models competent subgraph retrievers for multi-hop KGQA . In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-1...

    work page doi:10.18653/v1/2024.findings-emnlp.927 2024

  20. [20]

    Nourhan Ibrahim, Samar Aboulela, Ahmed Ibrahim, and Rasha Kashef. 2024. A survey on augmenting knowledge graphs (kgs) with large language models (llms): models, evaluation metrics, benchmarks, and challenges. Discover Artificial Intelligence, 4(1):76

    work page 2024

  21. [21]

    Gautier Izacard and Edouard Grave. 2021. https://doi.org/10.18653/V1/2021.EACL-MAIN.74 Leveraging passage retrieval with generative models for open domain question answering . In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021 , pages 874--880....

    doi:10.18653/v1/2021.eacl-main.74 2021

  22. [22]

    Boran Jiang, Yuqi Wang, Yi Luo, Dawei He, Peng Cheng, and Liangcai Gao. 2024 a . https://doi.org/10.1109/ICKG63256.2024.00026 Reasoning on efficient knowledge paths: Knowledge graph guides large language model for domain question answering . In IEEE International Conference on Knowledge Graph, ICKG 2023, Shanghai, China, December 1-2, 2023 , pages 142--149. IEEE

    arXiv 2024

  23. [23]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin - Yew Lin, Yuqing Yang, and Lili Qiu. 2024 b . https://doi.org/10.18653/V1/2024.ACL-LONG.91 Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P...

    doi:10.18653/v1/2024.acl-long.91 2024

  24. [24]

    Ziyan Jiang, Xueguang Ma, and Wenhu Chen. 2024 c . Longrag: Enhancing retrieval-augmented generation with long-context llms. arXiv preprint arXiv:2406.15319

    Pith/arXiv arXiv 2024

  25. [25]

    Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia - Yuan Chang, Huiyuan Chen, and Xia Hu. 2024. https://openreview.net/forum?id=nkOMLBIiI7 LLM maybe longlm: Selfextend LLM context window without tuning . In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net

    work page 2024

  26. [26]

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2020. https://openreview.net/forum?id=HklBjCEKvH Generalization through memorization: Nearest neighbor language models . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net

    work page 2020

  27. [27]

    Jack Lanchantin, Shubham Toshniwal, Jason Weston, Sainbayar Sukhbaatar, et al. 2024. Learning to reason and memorize with self-notes. Advances in Neural Information Processing Systems, 36

    work page 2024

  28. [28]

    Canny, and Ian Fischer

    Kuang - Huei Lee, Xinyun Chen, Hiroki Furuta, John F. Canny, and Ian Fischer. 2024. https://openreview.net/forum?id=OTmcsyEO5G A human-inspired reading agent with gist memory of very long contexts . In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net

    work page 2024

  29. [29]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459--9474

    2020

  30. [30]

    Qing Li and Guanzhong Wu. 2025. Explainable reasoning over temporal knowledge graphs by pre-trained language model. Information Processing & Management, 62(1):103903

    work page 2025

  31. [31]

    Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, Wenbo Su, and Bo Zheng. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-EMNLP.746 Graphreader: Building graph-based agent to enhance long-context abilities of large language models . In Findings of the Association for Computational Lingu...

    doi:10.18653/v1/2024.findings-emnlp.746 2024

  32. [32]

    Haochen Liu, Song Wang, Chen Chen, and Jundong Li. 2024 a . Few-shot knowledge graph relational reasoning via subgraph adaptation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3346--3356

    work page 2024

  33. [33]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024 b . Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157--173

    2024

  34. [34]

    Linhao Luo, Yuan - Fang Li, Gholamreza Haffari, and Shirui Pan. 2024. https://openreview.net/forum?id=ZGNWW7xZ6Q Reasoning on graphs: Faithful and interpretable large language model reasoning . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

    2024

  35. [35]

    Yansong Ning and Hao Liu. 2024. http://papers.nips.cc/paper\_files/paper/2024/hash/decd42d78c42cea59c95c7c3d40d5e0f-Abstract-Conference.html Urbankgent: A unified large language model agent framework for urban knowledge graph construction . In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems ...

    work page 2024

  36. [36]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

    2023

  37. [37]

    Smith, and Mike Lewis

    Ofir Press, Noah A. Smith, and Mike Lewis. 2022. https://openreview.net/forum?id=R8sQPpGCv0 Train short, test long: Attention with linear biases enables input length extrapolation . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    work page 2022

  38. [38]

    Author Name qwen-plus 128k. 2024. Title of the webpage. https://qwen.readthedocs.io/en/latest/

    work page 2024

  39. [39]

    Mohammad Sadegh Rasooli and Joel Tetreault. 2015. Yara parser: A fast and accurate dependency parser. arXiv preprint arXiv:1503.06733

    Pith/arXiv arXiv 2015

  40. [40]

    Devendra Singh Sachan, Mike Lewis, Dani Yogatama, Luke Zettlemoyer, Joelle Pineau, and Manzil Zaheer. 2023. Questions are all you need to train a dense passage retriever. Transactions of the Association for Computational Linguistics, 11:600--616

    work page 2023

  41. [41]

    Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Sch \"a rli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210--31227. PMLR

    2023

  42. [42]

    Ni, Heung - Yeung Shum, and Jian Guo

    Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel M. Ni, Heung - Yeung Shum, and Jian Guo. 2024. https://openreview.net/forum?id=nnVO1PvbTv Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, A...

    2024

  43. [43]

    Author Name text-embedding v3. 2024. Title of the webpage. https://help.aliyun.com/zh/model-studio/developer-reference/text-embedding-quick-start-1

    work page 2024

  44. [44]

    Xintao Wang, Qianwen Yang, Yongting Qiu, Jiaqing Liang, Qianyu He, Zhouhong Gu, Yanghua Xiao, and Wei Wang. 2023. Knowledgpt: Enhancing large language models with retrieval and storage access on knowledge bases. arXiv preprint arXiv:2308.11761

    Pith/arXiv arXiv 2023

  45. [45]

    Xinyi Wang, Alfonso Amayuelas, Kexun Zhang, Liangming Pan, Wenhu Chen, and William Yang Wang. 2024. https://openreview.net/forum?id=dZsEOFUDew Understanding reasoning ability of language models from the perspective of reasoning paths aggregation . In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . O...

    work page 2024

  46. [46]

    Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2024. https://openreview.net/forum?id=NG7sS51zVF Efficient streaming language models with attention sinks . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

    work page 2024

  47. [47]

    Yao Xu, Shizhu He, Jiabei Chen, Zihao Wang, Yangqiu Song, Hanghang Tong, Guang Liu, Jun Zhao, and Kang Liu. 2024. https://doi.org/10.18653/V1/2024.EMNLP-MAIN.1023 Generate-on-graph: Treat LLM as both agent and KG for incomplete knowledge graph question answering . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, E...

    doi:10.18653/v1/2024.emnlp-main.1023 2024

  48. [48]

    Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Onta \ n \' o n, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. https://proceedings.neurips.cc/paper/2020/hash/c8512d142a2d849725f31a9a7a361ab9-Abstract.html Big bird: Transformers for longer sequences . In Advances in Neural Information Pr...

    work page 2020

  49. [49]

    Yifan Zhang, Yang Yuan, and Andrew Chi-Chih Yao. 2024. On the diagram of thought. arXiv preprint arXiv:2409.10038

    Pith/arXiv arXiv 2024

  50. [50]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223

    Pith/arXiv arXiv 2023

  51. [51]

    Zirui Zhao, Wee Sun Lee, and David Hsu. 2024. Large language models as commonsense knowledge for large-scale task planning. Advances in Neural Information Processing Systems, 36

    work page 2024

  52. [52]

    Le Zhou. 2023. https://doi.org/10.1109/ACCESS.2023.3340854 Longt5-mulla: Longt5 with multi-level local attention for a longer sequence . IEEE Access , 11:138433--138444

    work page arXiv 2023

  53. [53]

    Tong Zhou, Yubo Chen, Kang Liu, and Jun Zhao. 2024. Cogmg: Collaborative augmentation between large language model and knowledge graph. arXiv preprint arXiv:2406.17231

    Pith/arXiv arXiv 2024

  54. [54]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  55. [55]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.