Pith. sign in

REVIEW 5 major objections 6 minor 30 references

KunLunBaizeRAG: Reinforcement Learning Driven Inference Performance Leap for Large Language Models

T0 review · 5 major / 6 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 Zero-annotation reinforcement learning, wrapped in pre-alignment and redundancy filtering, lifts a 32B retrieval model by about 15 points on four multi-hop QA benchmarks.

desk verdict Plausible RAG training recipe, but the headline numbers do not match the paper's own table and the evaluation is not verifiable from what is provided. read the letter →

arxiv 2506.19466 v2 pith:TSPV5ARU submitted 2025-06-24 cs.AI

Cheng Li , Jiexiong Liu , Yixuan Chen , Qihang Zhou , KunLun Meta This is my paper
classification cs.AI
keywords reinforcementlearningretrieval-augmentedgenerationmulti-hopquestionansweringretrievaldriftself-reflectionlocal-webroutingprogressivetrainingLLM-as-a-judge
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

The paper claims that a deliberately structured reinforcement-learning training pipeline can turn a plain retrieval-augmented language model into a stronger multi-hop reasoner without any new human labels. On four benchmarks including HotpotQA, the 32B model is reported to gain 14.82 points in exact match and 15.46 points in LLM-judged score over the compared baselines. The gains are attributed to aligning the model's first retrieval step with the question's intent, filtering redundant or low-confidence retrievals across rounds, and learning when to search locally rather than on the web. If the results hold, RAG systems for complex question answering would need no new architecture, only a better training objective and a few control layers around the retriever.

What carries the argument

The load-bearing objects are three control mechanisms plus one training strategy: RDRA (a pre-retrieval "thinking snippet" that turns the raw question into a task decomposition before querying the index), STIE (a memory-filter-confidence loop that computes token-overlap differences against the last three answers and shields answers repeated more than N times), and NLR (a policy network over local versus web retrieval with a reward that assigns +0.42 to local latency savings and +0.35 to web recall gains). The training strategy is a progressive curriculum: 600k samples at a 1:1 noise/clean ratio, then increasing gold-data share and noise perturbation, then reinforcement learning with short/long answer reward functions and masked document-token gradients. Each piece is stated to be lightweight and external to the base model's decoder.

What would settle it

Re-run the comparison on a fixed 500-question sample from HotpotQA with the exact answer parser and judge prompt published ahead of time; if the 32B model's EM margin over the strongest baseline disappears, or if human raters disagree with the LLM judge on 200 answers, the reported gains do not hold.

Watch

Extended reading notes

Core claim

KunLunBaizeRAG is a training-and-inference recipe that makes RAG reasoning self-correcting. Before any answer is generated, the model performs a background retrieval (RDRA) so that its initial reasoning starts from semantics aligned with the question. During multi-round think-retrieve loops, a memory-filter-confidence layer (STIE) blocks repeated answers, replaces low-confidence candidates, and decides when to stop. A routing policy (NLR) learns to balance local and web retrieval using rewards calibrated to latency and recall. A three-stage curriculum over 600k noisy and clean samples ends in reinforcement learning with dual short-answer and long-answer rewards, and document tokens are masked so that only task-relevant gradients propagate. The paper claims this zero-annotation recipe lifts the 32B model's EM and LJ scores by 14.82% and 15.46% across HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle, while making retrieval 42% faster on average when local search is chosen.

Load-bearing premise

The central claim stands or falls on the comparability of the benchmark numbers: all baselines must run under the same retrieval environment, answer parsing, and LLM-judge prompts, and the paper does not release those prompts or evaluation scripts.

Editorial extensions

If this is right

  • If the central claim holds, a RAG system trained this way should degrade less when questions contain ambiguous or domain-specific terms, because the pre-retrieval alignment step rephrases retrieval intent before searching.
  • Multi-round reasoning should stop producing the same wrong answer, because the STIE layer detects near-duplicates and blocks answers whose repeat count exceeds the threshold.
  • With routing learned rather than hand-tuned, a deployment could shift between local and web search depending on the question without losing the 42% latency advantage of local retrieval when it is sufficient.
  • The same 600k-sample curriculum with masked document-token gradients should transfer to other base models and multi-hop QA collections, since none of the control layers modifies the decoder's weights.

Reading between the lines

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

  • One implication the paper leaves implicit is that the RDRA pre-retrieval step is a general query-intent reformulator: it could be attached to any search-augmented agent without the 600k-sample curriculum.
  • The routing rewards (+0.42 latency, +0.35 recall) are hard-coded constants; a natural extension is to re-estimate them on the deployment corpus so the policy tracks changing network and index conditions.
  • The paper's headline gain averages over a mixed baseline set; its own table shows the margins over the strongest baseline (instructRAG) are a few EM points per benchmark, so readers who care about deployment should look at that per-benchmark margin rather than the 15-point average.
  • Because the control layers sit outside the decoder, the recipe should transfer to any base model that can emit think/search/answer tags; an ablation that turns off STIE's shielding threshold would reveal how much of the gain comes from repetition blocking versus retrieval quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces KunLunBaizeRAG, a reinforcement-learning-driven framework for improving multi-hop question answering in large language models. The claimed contributions are four mechanisms: RAG-driven Reasoning Alignment (RDRA), Search-Think Iterative Enhancement (STIE), Network-Local Intelligent Routing (NLR), and a progressive hybrid training strategy based on a 600k-sample dataset and the DAPO framework. The central quantitative claim is that the 32B variant achieves exact-match (EM) and LLM-judged (LJ) improvements of 14.82% and 15.46% across four benchmarks, including HotpotQA, with zero-annotation training. The paper reports experimental results in Table 2 and argues for strong generalization and self-correction ability.

Significance. If the reported results were reproducible, the framework would constitute a substantial engineering contribution to RAG-based reasoning: it combines retrieval routing, iterative refinement, and RL training at a meaningful scale, and it addresses practically important issues of retrieval drift and redundancy. The paper also includes concrete engineering details such as a clustering retriever with latency measurements, and it evaluates on four standard multi-hop QA benchmarks. However, the current significance cannot be assessed because the experimental apparatus is not described at the level required for independent verification and because the paper's own internal numbers do not support the headline improvements. No code, checkpoints, evaluation scripts, judge prompts, or answer-parsing rules are provided, and key baseline rows in Table 2 are undefined. The claimed scientific contribution therefore rests on evidence that is not currently available to the reader.

major comments (5)
  1. [Section 3.1 / Table 2] The baseline list in Section 3.1 names Direct comparison, NaiveRAG, Rewrite-Retrieve-Read, Iter-RetGen, ActiveRAG, SelfRAG, and SearChain, but Table 2 reports rows named IRCoT, R1-Searcher, Search-o1, Research, and instructRAG. The two rows labeled Research and instructRAG are never defined or cited anywhere in the paper, and none of Rewrite-Retrieve-Read, ActiveRAG, SelfRAG, or SearChain appears in the table. As a result, the comparison described in the setup is not the comparison presented in the results, and the reader cannot determine what methods the proposed approach is actually being compared against.
  2. [Abstract / Table 2] The abstract claims EM and LJ improvements of 14.82% and 15.46% for the 32B model across four benchmarks, but no comparison in Table 2 yields those aggregates. For example, KLBRAG-Baize-32B-Ins versus Search-o1 gives average deltas of approximately 4.47 EM and 5.43 LJ across the four datasets, while the same model versus instructRAG gives average deltas of approximately 2.25 EM and 2.92 LJ. The headline numbers therefore cannot be reconstructed from the paper's own reported results, so the paper's central quantitative claim is not supported by its own evidence.
  3. [Section 2.3 / Abstract] The abstract and the introduction state that training is 'zero-annotation' within the DAPO framework, but Section 2.3 describes a 600k-sample dataset composed of Dgold ∪ Dnoise, where Dgold is characterized as 'high-quality labeled data', and the reward functions in Eqs. (17) and (18) include an accuracy reward Racc. The paper never explains how gold labels and accuracy rewards are obtained without annotation, so the zero-annotation claim is internally inconsistent with the described training procedure.
  4. [Section 2.4, Eqs. (21)-(22)] The NLR mechanism's reward constants Reff(alocal)=+0.42 and Rinfo(aweb)=+0.35 are presented as the empirical facts that local retrieval reduces latency by 42% and web retrieval improves recall by 35%. No source, measurement procedure, or dataset for these numbers is given in Section 2.4 or elsewhere. If they are measured values, the protocol is absent; if they are manually chosen, the claimed contribution of the NLR mechanism is untested because no sensitivity analysis or ablation over these constants is reported. The same concern applies to the STIE thresholds δ1, δ2, δ3 in Eq. (9), the repetition cap N in Eq. (13), and the λ weights in Eqs. (17)-(18).
  5. [Section 3.1, Evaluation Metrics] The LJ metric is defined by 'GPT-40-mini and our defined judge prompts', but the judge prompts are not included, no human agreement or calibration is reported, and no error analysis is provided. Given that the paper's generalization claims in Section 3.2 rest heavily on the LJ columns, the reliability of the judge is unverified. In addition, the answer format shown in Section 3.4 places final answers in boxed form inside tags, whereas baselines are presumably free-form; without exact parsing rules for both the proposed model and the baselines, the reported EM margins cannot be independently confirmed.
minor comments (6)
  1. [Section 3.1] The text says 'leveraging GPT-40-mini'; this should be 'GPT-4o-mini'.
  2. [Section 3.1.2] The sentence 'with a throughput improvement of approximately 68' is incomplete, and 'Retrieval accuracy: Recall@10 metric keep 0.82' is ungrammatical; both should be rewritten.
  3. [Section 3.1] The benchmark enumeration is missing separators and a correct name: 'HotpotQA,2WikiMulti,HopQAMuSiQue and Bamboogle' should be 'HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle'.
  4. [References] Reference [27] is cited as the 'Veri' RL framework, but the cited paper is 'Learning to plan for retrieval-augmented large language models from knowledge graphs'; the citation does not match the claimed software framework.
  5. [Table 1 / Section 3.1.2] The latency figures in Table 1 (full retrieval 820 ms, cluster retrieval 85 ms, baseline scheme 120 ms) appear inconsistent with the numbers in Section 3.1.2 (6500 ms reduced to 700 ms, with a secondary similarity calculation reduction from 320 ms to 32 ms); the relationship between the two measurement conditions should be clarified.
  6. [Section 3.4] The case study contains corrupted markup such as '¡background¿' and mismatched tags, and the system-prompt example question about movie titles is unrelated to the actual question in the case; the excerpt should be proofread and aligned with the described tags.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claim is evaluated against Table 2 and does not reduce to fitted inputs or self-citations.

full rationale

KunLunBaizeRAG is an empirical systems paper; its central claim is that RL-driven training improves EM and LJ scores on four multi-hop QA benchmarks. That claim is supported by comparisons in Table 2 and does not reduce by construction to any fitted parameter, reward constant, or self-citation. The NLR reward constants (0.42 for local latency, 0.35 for web recall) are hand-set inputs to the training objective, but the paper does not present the reported score gains as derived from those constants, so this is mechanism design rather than a predictive claim that collapses into its own definition. The abstract's aggregate 14.82/15.46 improvements are not obviously reconstructible from Table 2, and the baselines listed in Section 3.1 do not match Table 2; however, these are reproducibility and consistency concerns, not circularity. No load-bearing self-citation chain, imported uniqueness theorem, or renaming of a known result as a derivation is present. Accordingly, the appropriate circularity finding is none.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

This systems paper rests on empirical assumptions rather than free derivations. Its headline numbers depend on hand-set thresholds, reward constants taken from the authors' own measurements, an unreleased training set, and unverified comparability of baselines and LLM-judged metrics. The central claim is not a mathematical derivation, so no invented entities are introduced.

free parameters (5)
  • STIE difference thresholds = delta1=0.25, delta2=0.5, delta3=0.75
    Hand-chosen thresholds for redundancy detection in Eq. (9); no sensitivity analysis is provided.
  • STIE repetition cap = Nmax=4
    Chosen to block repeated answers in Eq. (13); no analysis of this value is given.
  • NLR reward constants = R_eff(a_local)=0.42, R_info(a_web)=0.35
    Used in the dual-objective reward, Eqs. (20) to (22); these numbers are stated as measured latency and recall improvements, likely from the authors' own system, and are treated as fixed without error analysis.
  • Clustering retriever parameters = nclusters=5000, mindoc=150, alpha=0.9, tau0=1.2, tau_min=0.3, N=1000
    Hand-tuned for the FAISS IVF-PQ index and dynamic sampling in Section 3.1.1; no ablations are reported for these choices.
  • Reward weights = lambda_i with sum 1
    Weights in R_short and R_long, Eqs. (17) and (18); the specific values are never specified in the paper.
assumptions (4)
  • domain assumption The self-constructed 600k-sample hybrid dataset is suitable for training multi-hop QA models.
    Section 2.3 describes dataset construction from ReSearch-Qwen and ZeroSearch, but these sources are not cited and the dataset is not released.
  • domain assumption GPT-4o-mini LLM-judged scores are a valid and consistent measure of answer correctness across all models.
    Section 3.1 defines LJ using GPT-4o-mini and custom prompts; no validation of judge reliability or correlation with human judgments is given.
  • domain assumption Baseline results (including Search-o1, Research, instructRAG) were obtained under settings comparable to the proposed method.
    Table 2 lists baselines without describing their retrieval environments, prompts, or fine-tuning status; 'Research' and 'instructRAG' are undefined.
  • domain assumption The Baize-7B and Baize-32B base models exist and are appropriate starting points.
    No citation or description of the Baize models is provided; the paper assumes readers accept these as known.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KunLunBaizeRAG: Reinforcement Learning Driven Inference Performance Leap for Large Language Models." pith.science (2026). https://pith.science/paper/TSPV5ARU

@misc{pith2026250619466,
  author       = {Pith},
  title        = {Pith review of: KunLunBaizeRAG: Reinforcement Learning Driven Inference Performance Leap for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSPV5ARU}},
  note         = {Machine review of arXiv:2506.19466}
}
read the original abstract

This paper introduces KunLunBaizeRAG, a reinforcement learning-driven reasoning framework designed to enhance the reasoning capabilities of large language models (LLMs) in complex multi-hop question-answering tasks. The framework addresses key limitations of traditional RAG, such as retrieval drift, information redundancy, and strategy rigidity. Key innovations include the RAG-driven Reasoning Alignment (RDRA) mechanism, the Search-Think Iterative Enhancement (STIE) mechanism, the Network-Local Intelligent Routing (NLR) mechanism, and a progressive hybrid training strategy. Experimental results demonstrate significant improvements in exact match (EM) and LLM-judged score (LJ) across four benchmarks, highlighting the framework's robustness and effectiveness in complex reasoning scenarios.

Figures

Figures reproduced from arXiv: 2506.19466 by the authors.

Figure 1
Figure 1. The training overview of KunLunBaizeRAG. (a) The pipeline. (b) The details of the rollout [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Response length and number of search operations during training. [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Training and validation reward during training. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages

  1. [1]

    Dense passage retrieval for open-domain question answering,

    V . Karpukhin, B. Oguz, S. Min, P. S. H. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih, “Dense passage retrieval for open-domain question answering,” in EMNLP (1). Association for Computational Linguistics, 2020, pp. 6769–6781

    work page 2020

  2. [2]

    Text embeddings by weakly-supervised contrastive pre-training,

    L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, and F. Wei, “Text embeddings by weakly-supervised contrastive pre-training,” CoRR, vol. abs/2212.03533, 2022

    arXiv 2022

  3. [3]

    Flashrag: A modular toolkit for efficient retrieval-augmented generation research,

    J. Jin, Y . Zhu, X. Yang, C. Zhang, and Z. Dou, “Flashrag: A modular toolkit for efficient retrieval-augmented generation research,” CoRR, vol. abs/2405.13576, 2024

    arXiv 2024

  4. [4]

    Hybrid- flow: A flexible and efficient RLHF framework,

    G. Sheng, C. Zhang, Z. Ye, X. Wu, W. Zhang, R. Zhang, Y . Peng, H. Lin, and C. Wu, “Hybrid- flow: A flexible and efficient RLHF framework,” CoRR, vol. abs/2409.19256, 2024

    arXiv 2024

  5. [5]

    Wikidata: a free collaborative knowledgebase,

    D. Vrandecic and M. Kr ¨otzsch, “Wikidata: a free collaborative knowledgebase,” Commun. ACM, vol. 57, no. 10, pp. 78–85, 2014

    work page 2014

  6. [6]

    Constructing A multi-hop QA dataset for comprehensive evaluation of reasoning steps,

    X. Ho, A. D. Nguyen, S. Sugawara, and A. Aizawa, “Constructing A multi-hop QA dataset for comprehensive evaluation of reasoning steps,” inCOLING. International Committee on Computational Linguistics, 2020, pp. 6609–6625

    work page 2020

  7. [7]

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

    Z. Yang, P. Qi, S. Zhang, Y . Bengio, W. W. Cohen, R. Salakhutdinov, and C. D. Manning, “Hotpotqa: A dataset for diverse, explainable multi-hop question answering,” in EMNLP. Association for Computational Linguistics, 2018, pp. 2369–2380

    work page 2018

  8. [8]

    Musique: Multihop questions via single-hop question composition,

    H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal, “Musique: Multihop questions via single-hop question composition,” Trans. Assoc. Comput. Linguistics, vol. 10, pp. 539–554, 2022

    work page 2022

Show all 30 references
  1. [9]

    Dense text retrieval based on pretrained language models: A survey,

    W. X. Zhao, J. Liu, R. Ren, and J. Wen, “Dense text retrieval based on pretrained language models: A survey,” ACM Trans. Inf. Syst., vol. 42, no. 4, pp. 89:1–89:60, 2024

    2024

  2. [10]

    Deepseek- math: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “Deepseek- math: Pushing the limits of mathematical reasoning in open language models,” CoRR, vol. abs/2402.03300, 2024. 14

    2024 arXiv

  3. [11]

    Measuring and narrow- ing the compositionality gap in language models,

    O. Press, M. Zhang, S. Min, L. Schmidt, N. A. Smith, and M. Lewis, “Measuring and narrow- ing the compositionality gap in language models,” in EMNLP (Findings). Association for Computational Linguistics, 2023, pp. 5687–5711

    2023

  4. [12]

    Scaling LLM test-time compute optimally can be more effective than scaling model parameters,

    C. Snell, J. Lee, K. Xu, and A. Kumar, “Scaling LLM test-time compute optimally can be more effective than scaling model parameters,” CoRR, vol. abs/2408.03314, 2024

    2024 arXiv

  5. [13]

    s1: Simple test-time scaling,

    N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. J. Cand`es, and T. Hashimoto, “s1: Simple test-time scaling,” CoRR, vol. abs/2501.19393, 2025

    2025 arXiv

  6. [14]

    Star: Bootstrapping reasoning with reasoning,

    E. Zelikman, Y . Wu, J. Mu, and N. D. Goodman, “Star: Bootstrapping reasoning with reasoning,” in NeurIPS, 2022

    2022

  7. [15]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in NeurIPS, 2022

    2022

  8. [16]

    Claude 3.7 sonnet and claude code,

    Anthropic, “Claude 3.7 sonnet and claude code,” 2025. [Online]. Available: https: //www.anthropic.com/news/claude-3-7-sonnet

    2025

  9. [17]

    Learning to reason with LLMs,

    OpenAI, “Learning to reason with LLMs,” 2024. [Online]. Available: https://openai.com/index/ learning-to-reason-with-llms

    2024

  10. [18]

    Corrective retrieval augmented generation,

    S. Yan, J. Gu, Y . Zhu, and Z. Ling, “Corrective retrieval augmented generation,”CoRR, vol. abs/2401.15884, 2024

    2024 arXiv

  11. [19]

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

    A. Asai, Z. Wu, Y . Wang, A. Sil, and H. Hajishirzi, “Self-rag: Learning to retrieve, generate, and critique through self-reflection,” in ICLR. OpenReview.net, 2024

    2024

  12. [20]

    Interleaving retrieval with chain- of-thought reasoning for knowledge-intensive multi-step questions,

    H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal, “Interleaving retrieval with chain- of-thought reasoning for knowledge-intensive multi-step questions,” in ACL (1). Association for Computational Linguistics, 2023, pp. 10 014–10 037

    2023

  13. [21]

    Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy,

    Z. Shao, Y . Gong, Y . Shen, M. Huang, N. Duan, and W. Chen, “Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy,” in EMNLP (Findings) . Association for Computational Linguistics, 2023, pp. 9248–9274

    2023

  14. [22]

    Hugginggpt: Solving AI tasks with chatgpt and its friends in hugging face,

    Y . Shen, K. Song, X. Tan, D. Li, W. Lu, and Y . Zhuang, “Hugginggpt: Solving AI tasks with chatgpt and its friends in hugging face,” in NeurIPS, 2023

    2023

  15. [23]

    Agentboard: An analytical evaluation board of multi-turn LLM agents,

    C. Ma, J. Zhang, Z. Zhu, C. Yang, Y . Yang, Y . Jin, Z. Lan, L. Kong, and J. He, “Agentboard: An analytical evaluation board of multi-turn LLM agents,” in NeurIPS, 2024

    2024

  16. [24]

    Facilitating multi-turn function calling for llms via compositional instruction tuning,

    M. Chen, H. Sun, T. Li, F. Yang, H. Liang, K. Lu, B. Cui, W. Zhang, Z. Zhou, and W. Chen, “Facilitating multi-turn function calling for llms via compositional instruction tuning,” CoRR, vol. abs/2410.12952, 2024

    2024 arXiv

  17. [25]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick, J. Dwivedi-Yu, R. Dess `ı, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,” in NeurIPS, 2023

    2023

  18. [26]

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

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, Q. Guo, M. Wang, and H. Wang, “Retrieval-augmented generation for large language models: A survey,”CoRR, vol. abs/2312.10997, 2023

    2023 arXiv

  19. [27]

    Learning to plan for retrieval-augmented large language models from knowledge graphs,

    J. Wang, M. Chen, B. Hu, D. Yang, Z. Liu, Y . Shen, P. Wei, Z. Zhang, J. Gu, J. Zhou, J. Z. Pan, W. Zhang, and H. Chen, “Learning to plan for retrieval-augmented large language models from knowledge graphs,” in EMNLP (Findings). Association for Computational Linguistics, 2024,...

    2024

  20. [28]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    DeepSeek-AI, D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, X. Zhang, X. Yu, Y . Wu, Z. F. Wu, Z. Gou, Z. Shao, Z. Li, Z. Gao, A. Liu, B. Xue, B. Wang, B. Wu, B. Feng, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, D. Dai, D. Chen, D. Ji, E. L...

    2025 arXiv

  21. [29]

    Baichuan alignment technical report,

    M. Lin, F. Yang, Y . Shen, H. Sun, T. Li, T. Zhang, C. Zhu, T. Zhang, M. Zheng, X. Li, Y . Zhou, M. Chen, Y . Qin, Y . Li, H. Liang, F. Li, Y . Li, M. Wang, G. Dong, K. Fang, J. Xu, B. Cui, W. Zhang, Z. Zhou, and W. Chen, “Baichuan alignment technical report,”CoRR, vol. abs/24...

  22. [30]

    Qwen2.5 technical report,

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Xia, X. Ren, X. Ren, Y ...

    2024 arXiv

Pith tools

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