Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Prompting Large Language Models with Partial Knowledge for Answering Questions with Unseen Entities

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Large language models can be awakened by partial facts they already know, turning previously unanswerable questions into correct answers.

desk verdict Useful new KGQA-incomplete-KG benchmark and a plausible effect, but the 'awakening' interpretation is not proven because the knowledge-in-model check is self-referential. read the letter →

arxiv 2508.01290 v1 pith:PJ7BBE2A submitted 2025-08-02 cs.CL

Zhichao Yan , Jiapu Wang , Jiaoyan Chen , Yanyan Wang , Hongye Tan , Jiye Liang , Xiaoli Li , Ru Li
show 1 more author
Jeff Z.Pan
This is my paper
classification cs.CL
keywords LLMknowledgeawakeningretrieval-augmentedgenerationgraphquestionansweringunseenentitypartialattentionasMarkovtransitionsprobingincompletegraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Retrieval-augmented generation usually assumes that retrieved context must contain the answer or a close clue. This paper argues the opposite for one case: a partial fact that does not contain the answer can still help, if that fact is already stored in the LLM's own parameters. On questions each tested model could not answer alone, adding such a stored partial fact turned some of them into correct answers, across two knowledge-graph QA datasets and several open and API-based models. To explain why, the paper models the transformer's attention layers as a non-homogeneous Markov chain and shows that the injected fact can open a latent path from question to answer. It then turns the phenomenon into a retrieval recipe for a new task, Unseen Entity KGQA, where the question's main entity is absent from the knowledge graph, and reports that this recipe beats embedding-similarity retrieval baselines.

What carries the argument

The carrying idea is knowledge awakening: re-introducing a KG triple that is stored in the LLM but does not contain the answer can reactivate latent internal knowledge through structured latent activation. The paper's theoretical handle on this is a layer-wise Markov propagation model of attention, in which each layer's attention matrix is treated as a context-conditioned Markov transition operator, the stack of layers forms a non-homogeneous Markov chain, and the final representation is a composition of layer-wise transitions; the existence of a nonzero attention path from the question and the partial fact to the answer, together with concavity of a log-softmax path functional, implies that some external partial fact can raise the probability of the correct answer. On the empirical side, the machinery is a set of eight awakening-knowledge types built from the gold reasoning path and its variants, a True/False probing filter that requires the fact to be affirmed by the LLM before use, and a relation-based retrieval procedure for incomplete knowledge graphs that constructs the usable knowledge types by combining the question's relation with retrieved triplets and LLM-inferred type information.

What would settle it

Take the paper's unanswerable questions and replace the probed-True partial fact with a counterfactual triple that the same LLM also judges True but that is false in the knowledge graph; if the counterfactual awakens the correct answer as often as the real stored fact, then plausibility rather than storage is doing the work, and the awakening claim fails.

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that LLMs do not reliably use the knowledge they have: a question can be unanswerable by the same model that possesses the needed fact, and the failure can be repaired by re-supplying a related fact that shares structure with the missing reasoning path but never contains the answer. The paper formalizes this as awakening: for a question Q and answer A, there is a partial fact K, stored in the model and free of A, such that the model fails on Q alone but answers correctly on Q together with K. The evidence is a three-stage benchmark: select questions the model cannot answer, probe variant triples extracted from the gold reasoning path to find those the model affirms as True, and measure how often each affirmed triple awakens the correct answer. The same mechanism is then packaged into a rule-based retrieval method for the new Unseen Entity KGQA task, where entity linking cannot be used because the head entity is missing, and the method outperforms baselines that embed questions or entities for similarity search.

Load-bearing premise

The whole awakening claim rests on the probing step: the paper takes 'the LLM answers True three times to a fact question' as proof that the fact is stored in the model's parameters, but it could instead mean that the prompt makes the fact look plausible, in which case the improvement is ordinary in-context learning rather than awakening of stored knowledge.

Editorial extensions

If this is right

  • A RAG system facing incomplete retrieval can stop discarding partial context: triples that lack the answer may still trigger stored knowledge and should be kept as awakening candidates.
  • A question's relation, not its entity, is the retrieval signal that survives knowledge-graph incompleteness; retrieval by relation plus type outperforms question- and entity-embedding retrieval when the entity is missing.
  • Awakening is stronger when the injected knowledge sits closer to the answer in the reasoning path (second-hop beats first-hop) and when the model is given more reasoning tokens before answering.
  • The effect appears across model families, sizes, and API-based models, with diminishing returns at larger scale, so it is a general property of current LLMs rather than an artifact of one architecture.
  • Feeding multiple awakening facts at once can hurt accuracy, implying that conflicts among partial facts matter and retrieval should be selective.

Reading between the lines

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

  • Editorial inference: if the effect is genuine, the practical bottleneck in some LLM failures is access rather than storage; any context that relationally resembles the latent fact may act as a retrieval cue, which reframes harmless-looking retrieval noise as a potentially useful signal.
  • Editorial inference: the layer-wise Markov model predicts that awakening strength should depend on where in the stack the injected token first attends to the answer token; this could be tested by activation patching or by ablating the injected fact at specific layers.
  • Editorial inference: the True/False probing step could be checked externally by testing counterfactual triples that are equally plausible to the LLM but false in the knowledge graph; if those awaken the model just as well, the 'stored in parameters' constraint is not what drives the improvement.
  • Editorial inference: the relation-retrieval recipe suggests a cheap improvement for RAG over incomplete knowledge bases: index triples by relation embeddings rather than entity embeddings, since relations survive entity masking.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes that LLMs can be 'awakened' by prompting them with partial knowledge from a knowledge graph that does not itself contain the answer but is already stored in the model's parameters. It formalizes this as Task 1 with two conditions—the knowledge must be in the model (D1) and the question must be unanswerable without explicit knowledge (D2)—and provides a theoretical analysis based on a layer-wise Markov propagation model of attention. The authors then introduce Task 2, Unseen Entity KGQA, where the question's head entity is absent from the KG, and evaluate a rule-based retrieval method on Mintaka, 2WikiMultiHop, and ComplexWebQuestions across multiple instruct, base, and API-based LLMs. The reported results show consistent improvements over a No-RAG baseline and over DiFaR-style baselines in the unseen-entity setting.

Significance. If the central claim were established, this would be a useful contribution to understanding how LLMs use parametric knowledge and to designing retrieval-augmented systems for incomplete knowledge graphs. The paper offers a clear taxonomy of eight partial-knowledge types, a new task definition and dataset construction for unseen-entity KGQA, and a broad experimental matrix spanning model families, sizes, temperatures, and probing variants. The theoretical attempt, while not conclusive, gives a falsifiable formulation of when adding partial knowledge should help. However, the paper's distinctive claim—that the knowledge is already embedded in the model rather than merely provided as new context—is not supported by the current probing protocol. The empirical finding that adding partial facts helps is plausible, but the central conceptual distinction from ordinary in-context learning needs a control that the revision must supply.

major comments (4)
  1. [§3, Task 1; §4.2, Eq. (7)] The definition of awakening in Eq. (1) requires K ∈ M (constraint D1), but the only evidence for membership is the model's own True/False answer in Eq. (7). Because the same model provides both the probe and the downstream QA, a 'True' answer could reflect plausibility or prompt bias rather than stored parametric knowledge. The QA/NLI comparison in Appendix B.4 changes only the task format and remains self-referential. This is load-bearing: if K is not actually stored, the observed improvement after injecting K is ordinary in-context learning from new context, which is the AAG-style baseline the paper distinguishes itself from in Section 2.2. The revision should add a control with facts known to be absent from the model's training or with deliberately corrupted triples, showing that these produce significantly smaller gains, or use an independent external probe (e.g., a second model or a cloze-style test) to select K.
  2. [Appendix A, Eq. (14)] The proof that the path functional is concave and therefore has a global maximum does not establish awakenability. A global maximum on the finite combinatorial domain only shows that some input configuration maximizes the target-token probability; it does not show that this input is 'already embedded' (K ∈ M) or that it improves over the no-K condition, which is the counterfactual in Eq. (1). The proof also assumes the initial hidden state has a nonzero element for K and that an attention path to A exists, i.e., it assumes D1 rather than deriving it. As stated, the argument applies equally to any new fact injected into the prompt, so it cannot distinguish awakening from standard in-context learning.
  3. [Table 2; §6.3] The quantitative support for the knowledge-type ranking and the model-size trend is weak. Table 2 reports differences as small as 0.3–0.6 percentage points on 2Wiki (e.g., T1 vs T5 for Llama3.1-8B: 3.7 vs 3.1 Hits@1), and the comparisons across differing filtered sample counts F are intersection-based but no confidence intervals, standard deviations, or repeated runs are reported. In Figure 5 and Table 10, the 7B→32B jump is clear, but the 32B→72B change is small and would need significance testing. The CWQ differences (25-point gains) are large enough to support the existence of an effect, but RQ1's finer conclusions about which knowledge types awaken best are not statistically substantiated.
  4. [§4.2, Step 1; Table 2] Because only questions that fail all three sampled-temperature runs are retained, the No-RAG Hits@1 figure is 0.0 by construction. This makes the headline hit-rate gains over No RAG a consequence of the selection protocol rather than an intrinsic property of the LLM. The design is defensible as a way to study previously unanswerable questions, but the paper should state explicitly that the gains are conditional on the unanswerable subset and should report results on the full dataset as well, so the reader can assess selection bias.
minor comments (5)
  1. [Section 1, Figure 1 caption] The caption appears truncated and contains garbled notation (e.g., 'T1 is the start entity...'; the right panel is not described). Please make the caption self-contained and consistent with Table 1.
  2. [Section 4.1, Eq. (5)] The order of composition is written inconsistently: the leftmost product notation '(∏_{l=1}^{L} A^{(l)}) H^{(0)}' and the expanded form 'A^{(L)} ... A^{(1)} H^{(0)}' should be aligned with an explicit convention for the product order.
  3. [Appendix B.1 vs Table 2] Table 2 appears to be a subset of the Instruct rows of Table 5; please clarify the relationship between the two tables and why the intersection-based results are used in the main text.
  4. [Eq. (15)] The definition of ACCR appears to have the correctness condition inverted: it returns 0 when all gold tokens are in the prediction list and 1 otherwise, which contradicts its use in Eq. (6), where 1 indicates a correct answer. Please fix the formula.
  5. [Abstract and Introduction] The abstract says 'two Knowledge Graphs QA datasets' and the introduction says 'construct two datasets' for Unseen Entity KGQA, while Table 3 uses three datasets (Mintaka, 2Wiki, and CWQ); align the wording throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the awakening claim is an empirical finding conditional on an operational probe, not a result forced by the paper's definitions or self-citations.

full rationale

The paper's derivation chain is self-contained rather than circular. Task1 defines awakening operationally in Eq. 1 as the existence of partially relevant knowledge K, with K∈M established by the knowledge-probing procedure and D2 requiring the question to be initially unanswerable. The probe in Eq. 7 is a True/False verification task, while the awakening evaluation measures whether the probed K changes the answer to a different question; these are distinct behaviors, so the observed improvement is not the probe output by construction. The theoretical analysis in Section 4.1 and Appendix A is a concavity/maximum argument that is vacuous as a proof of awakening, but it does not assume the conclusion and is therefore not circular. The Task2 application transfers empirically observed awakening rules to an incomplete-KG setting, which is a standard application rather than a derivation from the conclusion. Self-citations in the reference list are background citations and are not load-bearing for the central claim, which is instead supported by experiments on external benchmarks (2Wiki, CWQ, Mintaka) and baselines such as DiFaR. The main weakness, namely that the probe uses the same LLM to certify parametric storage and thus may not rule out standard in-context learning from plausible context, is a construct-validity concern rather than a formal circularity under the rubric, which requires an exhibited equation-level reduction or a fitted parameter renamed as prediction. No such reduction is present.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The central claim rests on a self-assessment loop: the model decides which facts are stored in itself, and the same model then demonstrates the awakening effect. The theoretical section is a sketch (Markov chain metaphor plus a concavity existence statement) rather than a proof that the constructed T1-T8 knowledge actually raises answer probability. The free parameters are mostly hand-set thresholds and post-hoc exclusions that shape which results are reported.

free parameters (4)
  • Unanswerable selection threshold = three sampling passes, temperature 0.1 to 1.0
    Hand-set criterion in Section 4.2 Step 1 that defines the evaluation set; any question the model answers correctly in one of three attempts is removed, so the No RAG baseline is 0 by construction.
  • Knowledge probing rounds = three consecutive True responses
    Hand-set threshold in Section 4.2 Stage 3 (Eq. 7) for deciding K in M; changing it changes which facts count as stored and thus the measured awakening effect.
  • Knowledge type taxonomy T1-T8
    The set of partial knowledge types is chosen by the authors; the main results exclude T3 and T8 after observing their performance (Appendix B.5), making the reported awakening rates conditional on this post-hoc selection.
  • Subgraph triple cap = 40,000
    Implementation limit during incomplete KG construction (Appendix C); may affect which triples are available for retrieval.
assumptions (3)
  • ad hoc to paper The attention matrix in each transformer layer can be interpreted as a context-conditioned Markov transition matrix.
    Invoked in Section 4.1; residual connections and multi-head aggregation are abstracted as identity matrices, so the Markov chain is a modeling metaphor, not an exact equivalence.
  • ad hoc to paper If the log-softmax path functional is concave on a finite combinatorial domain, there exists an external knowledge K that increases the probability of the answer.
    Appendix A proves concavity and existence of a maximum, but the maximum is over paths already present; the leap to 'there may exist K that increases probability' is not proven, and no link is made to the eight constructed knowledge types.
  • domain assumption The LLM's answers to True/False probing questions accurately reflect whether a fact is stored in its parameters.
    Used in Section 4.2 Stage 3 and in Task1 definition; no external ground truth for parameter storage is given.
invented entities (1)
  • structured latent activation K'
    purpose: Postulated internal state activated by partial knowledge K, claimed to mediate the awakening effect.
    Introduced in Section 4.1 as the hidden-state path from Q and K to A; the paper provides no direct measurement, only the behavioral improvement, so the entity is unfalsifiable on its own.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prompting Large Language Models with Partial Knowledge for Answering Questions with Unseen Entities." pith.science (2026). https://pith.science/paper/PJ7BBE2A

@misc{pith2026250801290,
  author       = {Pith},
  title        = {Pith review of: Prompting Large Language Models with Partial Knowledge for Answering Questions with Unseen Entities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJ7BBE2A}},
  note         = {Machine review of arXiv:2508.01290}
}
read the original abstract

Retrieval-Augmented Generation (RAG) shows impressive performance by supplementing and substituting parametric knowledge in Large Language Models (LLMs). Retrieved knowledge can be divided into three types: explicit answer evidence, implicit answer clue, and insufficient answer context which can be further categorized into totally irrelevant and partially relevant information. Effectively utilizing partially relevant knowledge remains a key challenge for RAG systems, especially in incomplete knowledge base retrieval. Contrary to the conventional view, we propose a new perspective: LLMs can be awakened via partially relevant knowledge already embedded in LLMs. To comprehensively investigate this phenomenon, the triplets located in the gold reasoning path and their variants are used to construct partially relevant knowledge by removing the path that contains the answer. We provide theoretical analysis of the awakening effect in LLMs and support our hypothesis with experiments on two Knowledge Graphs (KGs) Question Answering (QA) datasets. Furthermore, we present a new task, Unseen Entity KGQA, simulating real-world challenges where entity linking fails due to KG incompleteness. Our awakening-based approach demonstrates greater efficacy in practical applications, outperforms traditional methods that rely on embedding-based similarity which are prone to returning noisy information.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. CAGE: Cognitive Attribution Graphs for Faithful Inline Citation Generation in Long-Form Question Answering

    cs.CL 2026-07 conditional novelty 5.0 of 10

    Explicit cognitive attribution graphs before generation contract claim–document assignment space and yield SOTA faithful inline citations on long-form QA benchmarks.

Reference graph

Works this paper leans on

48 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

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

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

    2024

  2. [2]

    Language models are few-shot learners.Advances in neural infor- mation processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural infor- mation processing systems, 33:1877–1901, 2020

    1901

  3. [3]

    Large language models-guided dynamic adaptation for temporal knowledge graph reasoning.Advances in Neural Information Processing Systems, 37:8384–8410, 2024

    Jiapu Wang, Sun Kai, Linhao Luo, Wei Wei, Yongli Hu, Alan Wee-Chung Liew, Shirui Pan, and Baocai Yin. Large language models-guided dynamic adaptation for temporal knowledge graph reasoning.Advances in Neural Information Processing Systems, 37:8384–8410, 2024

    2024

  4. [4]

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

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

    2020

  5. [5]

    Retrieval-augmented generation for large language models: A survey, 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey, 2023

    work page 2023

  6. [6]

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

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

    work page 2024

  7. [7]

    Enabling large language models to generate text with citations

    Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. Enabling large language models to generate text with citations. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6465–6488, 2023

    2023

  8. [8]

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

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self- rag: Learning to retrieve, generate, and critique through self-reflection. InThe Twelfth International Conference on Learning Representations, 2023

    work page 2023

Show all 48 references
  1. [9]

    Fine tuning vs

    Heydar Soudani, Evangelos Kanoulas, and Faegheh Hasibi. Fine tuning vs. re- trieval augmented generation for less popular knowledge. InProceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Reg...

    2024

  2. [10]

    Sufficient context: A new lens on retrieval augmented genera- tion systems

    Hailey Joren, Jianyi Zhang, Chun-Sung Ferng, Da-Cheng Juan, Ankur Taly, and Cyrus Rashtchian. Sufficient context: A new lens on retrieval augmented genera- tion systems. InThe Thirteenth International Conference on Learning Representa- tions, 2025

    2025

  3. [11]

    Making retrieval- augmented language models robust to irrelevant context

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. Making retrieval- augmented language models robust to irrelevant context. InThe Twelfth Interna- tional Conference on Learning Representations, 2024

    2024

  4. [12]

    The power of noise: Redefining retrieval for rag systems

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. The power of noise: Redefining retrieval for rag systems. InProceedings of the 47th International ACM SIGIR Conference on Resea...

    2024

  5. [13]

    The distracting effect: Understanding irrelevant passages in RAG

    Chen Amiraz, Florin Cuconasu, Simone Filice, and Zohar Karnin. The distracting effect: Understanding irrelevant passages in RAG. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18228–18258. Association fo...

    2025

  6. [14]

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

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

    1983

  7. [15]

    Knowl- edge neurons in pretrained transformers

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowl- edge neurons in pretrained transformers. InProceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493–8502, 2022

    2022

  8. [16]

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

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

    2024

  9. [17]

    Psychology Press, 2013

    John R Anderson.The architecture of cognition. Psychology Press, 2013

    2013

  10. [18]

    Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training

    Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    2024

  11. [19]

    How easily do irrelevant inputs skew the responses of large language models? InFirst Conference on Language Modeling, 2024

    Siye Wu, Jian Xie, Jiangjie Chen, Tinghui Zhu, Kai Zhang, and Yanghua Xiao. How easily do irrelevant inputs skew the responses of large language models? InFirst Conference on Language Modeling, 2024

    2024

  12. [20]

    Robust information retrieval

    Yu-An Liu, Ruqing Zhang, Jiafeng Guo, and Maarten de Rijke. Robust information retrieval. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining, pages 1008–1011, 2025

    2025

  13. [21]

    Query2doc: Query expansion with large language models

    Liang Wang, Nan Yang, and Furu Wei. Query2doc: Query expansion with large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9414–9423, 2023

    2023

  14. [22]

    Sketching without worrying: Noise-tolerant sketch-based image retrieval

    Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Anee- shan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. InProceedings of the IEEE/CVF conference on computer vision and patt...

    2022

  15. [23]

    Retrieval, re- ranking and multi-task learning for knowledge-base question answering

    Zhiguo Wang, Patrick Ng, Ramesh Nallapati, and Bing Xiang. Retrieval, re- ranking and multi-task learning for knowledge-base question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3...

    2021

  16. [24]

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering.Advances in Neural Information Processing Systems, 37:132876–132907, 2024

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

    2024

  17. [25]

    Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph

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

    2024

  18. [26]

    Reasoning on graphs: Faithful and interpretable large language model reasoning

    LINHAO LUO, Yuan-Fang Li, Reza Haf, and Shirui Pan. Reasoning on graphs: Faithful and interpretable large language model reasoning. InThe Twelfth Inter- national Conference on Learning Representations, 2024

    2024

  19. [27]

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

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

    2024 arXiv

  20. [28]

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

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

    2024 arXiv

  21. [29]

    Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025. �������������� ���� ����� ����������� ��� ��� ��� �� ���

    2025 arXiv

  22. [30]

    Rethinking reflection in pre-training.arXiv preprint arXiv:2504.04022, 2025

    Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Parmar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, et al. Rethinking reflection in pre-training.arXiv preprint arXiv:2504.04022, 2025

    2025 arXiv

  23. [31]

    Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars.arXiv preprint arXiv:2503.01307, 2025

    Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars.arXiv preprint arXiv:2503.01307, 2025

    2025 arXiv

  24. [32]

    R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592, 2025

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592, 2025

    2025 arXiv

  25. [33]

    Search-r1: Training llms to reason and leverage search engines with rein- forcement learning.arXiv preprint arXiv:2503.09516, 2025

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with rein- forcement learning.arXiv preprint arXiv:2503.09516, 2025

    2025 arXiv

  26. [34]

    Learning to reason with search for llms via reinforcement learning.arXiv preprint arXiv:2503.19470, 2025

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Fan Yang, Zenan Zhou, Weipeng Chen, Haofen Wang, Jeff Z Pan, et al. Learning to reason with search for llms via reinforcement learning.arXiv preprint arXiv:2503.19470, 2025

    2025 arXiv

  27. [35]

    Awakening augmented generation: Learning to awaken internal knowledge of large language models for question answering

    Huanxuan Liao, Shizhu He, Yao Xu, Yuanzhe Zhang, Shengping Liu, Kang Liu, and Jun Zhao. Awakening augmented generation: Learning to awaken internal knowledge of large language models for question answering. InProceedings of the 31st International Conference on Computational Li...

    2025

  28. [36]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

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

    2022

  29. [37]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809– 11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809– 11822, 2023

    2023

  30. [38]

    Chain-of-note: Enhancing robustness in retrieval- augmented language models

    Wenhao Yu, Hongming Zhang, Xiaoman Pan, Peixin Cao, Kaixin Ma, Jian Li, Hongwei Wang, and Dong Yu. Chain-of-note: Enhancing robustness in retrieval- augmented language models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14672...

    2024

  31. [39]

    Struc- tured prompting: Scaling in-context learning to 1,000 examples.arXiv preprint arXiv:2212.06713, 2022

    Yaru Hao, Yutao Sun, Li Dong, Zhixiong Han, Yuxian Gu, and Furu Wei. Struc- tured prompting: Scaling in-context learning to 1,000 examples.arXiv preprint arXiv:2212.06713, 2022

    2022 arXiv

  32. [40]

    Not all demonstration examples are equally beneficial: Reweighting demonstration examples for in- context learning

    Zhe Yang, Damai Dai, Peiyi Wang, and Zhifang Sui. Not all demonstration examples are equally beneficial: Reweighting demonstration examples for in- context learning. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 13209–13221, 2023

    2023

  33. [41]

    Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4862–4876, 2023

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4862–4876, 2023

    2023

  34. [42]

    Prompting as probing: Using language models for knowledge base construction

    Dimitrios Alivanistos, Selene Báez Santamaría, Michael Cochez, Jan Christoph Kalo, Emile van Krieken, and Thiviyan Thanapalasingam. Prompting as probing: Using language models for knowledge base construction. In2022 Semantic Web Challenge on Knowledge Base Construction from Pr...

    2022

  35. [43]

    From self-attention to markov models: Unveiling the dynamics of generative transformers

    Muhammed Emrullah Ildiz, Yixiao Huang, Yingcong Li, Ankit Singh Rawat, and Samet Oymak. From self-attention to markov models: Unveiling the dynamics of generative transformers. InInternational Conference on Machine Learning, pages 20955–20982. PMLR, 2024

    2024

  36. [44]

    Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation, 2024

    2024

  37. [45]

    Mintaka: A complex, natural, and multilingual dataset for end-to-end question answering

    Priyanka Sen, Alham Fikri Aji, and Amir Saffari. Mintaka: A complex, natural, and multilingual dataset for end-to-end question answering. InProceedings of the 29th International Conference on Computational Linguistics, pages 1604–1619, 2022

    2022

  38. [46]

    Rearter: Retrieval-augmented reasoning with trustworthy process rewarding.arXiv preprint arXiv:2501.07861, 2025

    Zhongxiang Sun, Qipeng Wang, Weijie Yu, Xiaoxue Zang, Kai Zheng, Jun Xu, Xiao Zhang, Song Yang, and Han Li. Rearter: Retrieval-augmented reasoning with trustworthy process rewarding.arXiv preprint arXiv:2501.07861, 2025

    2025 arXiv

  39. [47]

    Direct fact retrieval from knowledge graphs without entity linking

    Jinheon Baek, Alham Fikri Aji, Jens Lehmann, and Sung Ju Hwang. Direct fact retrieval from knowledge graphs without entity linking. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10038–10055, 2023

    2023

  40. [48]

    T” represents the knowledge types (e.g., No RAG acts the knowledge �−�−� ), “F

    Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. Large language models are versatile decomposers: Decomposing evidence and questions for table-based reasoning. InProceedings of the 46th international ACM SIGIR conference on research and development in inf...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.