Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

CIIR@LiveRAG 2025: Optimizing Multi-Agent Retrieval Augmented Generation through Self-Training

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A self-trained multi-agent RAG system that learns from its own best trajectories outperforms plain retrieve-then-read pipelines and ranked 7th in the SIGIR 2025 LiveRAG competition.

desk verdict A well-engineered competition system whose headline claim over vanilla RAG is undermined by using the same reward for training, checkpoint selection, and evaluation; still worth a referee's time. read the letter →

arxiv 2506.10844 v1 pith:ZSL2TNU5 submitted 2025-06-12 cs.CL cs.IR

Alireza Salemi , Mukta Maddipatla , Hamed Zamani This is my paper
classification cs.CLcs.IR
keywords multi-agentretrieval-augmentedgenerationself-trainingreward-guidedtrajectorysamplingcoordinatoragentLLM-basedrewardmodelsLiveRAG2025DataMorgana
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 multi-agent retrieval-augmented generation system, mRAG, can be optimized by self-training on its own best-scoring interaction trajectories, and that this beats the standard retrieve-then-read RAG pipeline on both correctness and faithfulness. Six specialist agents handle planning, search, summarization, reasoning, validation, and generation, while a coordinator decides which agent to call next. The authors sample eight trajectories per question, score them with two LLM-based reward models, keep the highest-reward trajectories, and fine-tune the trainable agents to reproduce them. On a test set generated with DataMorgana, mRAG outperformed both a no-retrieval LLM and vanilla RAG; in the SIGIR 2025 LiveRAG competition it ranked 7th among the 20 reported teams by correctness. The significance is that inter-agent coordination in RAG can be improved without backpropagating through the agents.

What carries the argument

The central mechanism is the factorized trajectory model: by assuming each agent's action depends only on the input $x$, the joint probability of a trajectory $\tau=a_1\ldots a_n$ simplifies to $\prod_i p(a_i|x)$. This turns end-to-end multi-agent optimization, which would require gradients across cooperating systems, into supervised fine-tuning of each agent's language model on the actions found in high-reward trajectories. The reward signal comes from two LLM-based judges: a recall-oriented nugget correctness score that checks ground-truth atomic aspects against the response, and a faithfulness score that checks response claims against retrieved documents. The coordinator is a Qwen 2.5 7B model that iteratively selects agents, and the searcher uses the Lion sparse retriever to fetch documents.

What would settle it

Take the DataMorgana test set and compare mRAG against vanilla RAG after truncating or length-matching all responses to a fixed token budget before scoring with the same correctness and faithfulness rewards; if mRAG's advantage shrinks or reverses, the claimed gains are largely an artifact of reward models that reward verbosity. A second check is to have the competition's official hidden scorer run on the same outputs.

Watch

Extended reading notes

Core claim

The central claim is that the quality of multi-agent RAG depends on how agents collaborate, and that collaboration can be learned from the system's own successes. mRAG decomposes each question into subtasks executed by a coordinator, planner, searcher, summarizer, reasoner, validator, and generator/reviser. Training samples $T=8$ diverse trajectories per input at high temperature, scores each final response with a correctness reward based on nugget recall and a faithfulness reward based on claim-level support, then keeps only the best trajectory (up to three in ties) and maximizes the factorized log-probability of its action sequence, which reduces to supervised fine-tuning of the agents. With correctness weighted four times more than faithfulness, the trained system outperformed vanilla RAG and a non-retrieval baseline on the DataMorgana-derived test set and placed 7th in the LiveRAG 2025 competition. The paper presents case studies showing the trained coordinator decomposing multi-aspect questions and the searcher persisting with the same query before reformulating after retrieval failures.

Load-bearing premise

Everything rests on the two LLM-based judges of correctness and faithfulness being trustworthy measures of response quality, yet the paper's own verbosity experiment shows the correctness judge can be pushed up by writing longer answers, so the training signal may reward length rather than accuracy.

Editorial extensions

If this is right

  • mRAG, trained by reward-guided self-training, reports higher correctness and faithfulness than vanilla retrieve-then-read RAG and a no-retrieval LLM on a DataMorgana-derived test set.
  • A multi-agent RAG pipeline can be trained end to end without computing gradients across agents, by distilling the behaviors of its own best trajectories.
  • The trained coordinator learns reusable orchestration policies, such as splitting questions into aspects and invoking the searcher per aspect, and the searcher learns to exhaust a query before reformulating it.
  • The reported 7th-place result in the LiveRAG 2025 competition suggests the approach transfers to an external evaluation setting with a hidden scoring function.

Reading between the lines

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

  • Because the training and the main test evaluation both use the same two reward models, the reported superiority may partly reflect optimizing the judge rather than answer quality; a fair test would use length-matched outputs or a judge the system never trained against.
  • The paper's own verbosity experiment supports that caution: lengthening responses raised the correctness reward from 0.528 to 0.584, so part of the self-training gain could come from learning to say more rather than say better.
  • The factorized trajectory assumption discards inter-agent dependencies; a natural extension is trajectory-level credit assignment or rewarding interactions as well as final answers, which could further improve coordination.
  • The same self-training loop could be applied to other agent orchestrators that use a fixed retriever and generator, making the method a general recipe for learning when to call which tool.
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

3 major / 5 minor

Summary. The paper introduces mRAG, a multi-agent retrieval-augmented generation framework composed of a coordinator and task-specialized agents (planner, searcher, reasoner, summarizer, validator, generator/reviser). Agents are trained with a reward-guided self-training procedure: multiple trajectories are sampled per question, scored by two LLM-based reward models (a nugget-based correctness reward and a RAGAS-style faithfulness reward), and the highest-reward trajectories are used as supervised fine-tuning targets for the trainable agents. The system is evaluated on a DataMorgana-derived test set of 1,000 question-answer pairs, where it is reported to outperform a vanilla RAG baseline and a no-retrieval baseline in both correctness and faithfulness. The paper also reports mRAG's 7th-place finish in the SIGIR 2025 LiveRAG competition and includes qualitative case studies of learned coordinator and searcher behaviors.

Significance. If the reported results are valid, the paper demonstrates a practical approach to optimizing multi-agent orchestration for RAG without end-to-end reinforcement learning. The release of code, the detailed prompts and algorithms in the appendices, and the concrete case studies of learned query-reformulation and aspect-decomposition behavior are valuable assets for reproducibility. However, the central comparative claim—that self-trained mRAG outperforms vanilla RAG—rests on an evaluation that uses the same reward models employed for training and for checkpoint selection. The paper's own Section 4 finding that simply prompting longer responses raises the correctness reward from 0.528 to 0.584 indicates that these rewards are sensitive to response length, and thus the reported gains in Figure 2 may partly reflect reward overfitting rather than genuine quality improvements. The LiveRAG competition result is a useful external signal, but it does not include a vanilla RAG baseline, so it cannot validate the specific comparative claim made in the paper. The significance of the work therefore depends on whether the evaluation can be made reward-independent.

major comments (3)
  1. [Section 4, Figure 2, Appendix D.2, Section 3.2] The main comparative evaluation is circular with respect to the training objective. The same LLM-based correctness and faithfulness reward models are used for three purposes: selecting high-reward trajectories in Section 3.2, selecting the best checkpoint by evaluating on the test set every 500 steps in Appendix D.2, and computing the headline results in Figure 2. Consequently, the reported superiority of mRAG over vanilla RAG may reflect optimization of these specific reward models rather than a genuine improvement in response quality. This concern is substantiated by the paper's own observation in Section 4 that prompting for longer responses—without any training—raises the recall-oriented correctness reward from 0.528 to 0.584. To support the central claim, the authors should provide an evaluation that is independent of the training rewards, such as human judgment on a sample of responses, or an official competition metric on a held-out set. At minimum, they should control for response length (e.g., by matching length distributions across systems) and show that the gain over vanilla RAG persists.
  2. [Section 3.2, Eq. (2)] The factorization assumption in Eq. (2) that p(τ|x) = ∏ p(ai|x) treats agent actions as conditionally independent given the input question, ignoring the sequential dependencies that define multi-agent collaboration. As a result, the training objective is effectively supervised fine-tuning of each agent on its own action within the selected high-reward trajectory, not an end-to-end optimization of the joint coordination policy. The paper's claim that this method 'optimizes inter-agent collaboration' is therefore overstated. I recommend either tempering the claim or adding an experimental comparison that isolates the benefit of the learned coordination—for example, comparing against a baseline that is trained on trajectories selected at random, or against a single-agent model trained on the same high-reward final answers.
  3. [Appendix D.2, Figure 2] The training protocol evaluates models on the test set every 500 steps and selects the checkpoint with the best performance. This turns the test set into a model selection criterion, which inflates the reported results and breaks the independence of the test set. The absence of error bars or significance tests further weakens the reliability of the comparisons in Figure 2, especially given the small number of runs implied by the setup. The authors should use a held-out validation split for checkpoint selection and report test performance of the selected checkpoint, together with variance estimates across multiple training runs or bootstrapped confidence intervals for the evaluation scores.
minor comments (5)
  1. [Section 4] There is a typo in the sentence about 'the LiveRAG competition prganizers'; it should read 'organizers'.
  2. [Appendix E, Table 2] The case study tables contain several typos and malformed JSON-like fragments (e.g., 'Documnet 1 ID' in Table 2, and mismatched quotation marks in the validator output). Cleaning these up would improve readability.
  3. [Section 3.2] The term 'self-training' is used, but the method is a single round of reward-filtered supervised fine-tuning rather than an iterative self-training loop. Consider clarifying this distinction to avoid confusion with iterative approaches.
  4. [Figure 2] The caption and accompanying text should explicitly state the numeric scores and clearly distinguish the 'mRAG (untrained)' configuration from the trained mRAG, since the untrained configuration is an important control that is only briefly mentioned in Section 4.
  5. [Section 1 and Appendix D.2] The paper should clarify that the test set used for the main comparison is the 1,000 DataMorgana-generated pairs described in Section 2, not the official LiveRAG competition test set. This distinction is important for interpreting the external validity of the reported results.

Circularity Check

3 steps flagged · score 6.0 of 10

The headline comparison is scored with the same LLM reward used to select training trajectories and to pick the test-set checkpoint, so the reported gain over vanilla RAG is not an independent prediction.

  1. fitted input called prediction [Section 4 (Figure 2 caption) and Section 3.2]
    "Evaluation of mRAG and baselines using the introduced reward models in Section 3.2 and Appendix D.3. The findings indicate that mRAG, when optimized using the method described in Section 3.2, outperforms both baselines in terms of correctness and faithfulness."

    Figure 2 evaluates with the exact reward models used for training. Section 3.2 says: 'A reward model RM is then applied to the final response of each trajectory for scoring. We retain only the trajectory with the highest reward score... These selected trajectories are used to train the agents to reproduce the optimal sequence of actions.' Thus the reported correctness/faithfulness advantage is a measurement of the training objective itself, not an independent test of the paper's claim. The concern is concrete: Section 4 reports that lengthening responses alone moved the correctness reward from 0.528 to 0.584, so optimizing this reward can improve the score without improving answer quality.

  2. fitted input called prediction [Appendix D.2 (Training Setup)]
    "Models are evaluated every 500 steps on the test set, and the checkpoint with the best performance is selected."

    The test set is used for model selection, and the same test set produces the Figure 2 numbers that support 'mRAG outperforms both baselines.' The reported result is therefore the best-of-checkpoints reward on that test set, selected to maximize the very metric being reported. This makes the headline comparison partly an artifact of selection rather than an unbiased estimate of generalization, so the 'prediction' is statistically forced.

1 more flagged steps
  1. fitted input called prediction [Section 3.2 (Reward Model)]
    "We observed that mRAG performs worst on the correctness reward. To emphasize improvement in this area, we assign a weight of 4 to the correctness reward and a weight of 1 to faithfulness reward, and compute a weighted average to obtain the final reward."

    The 4:1 weighting is chosen after observing mRAG's correctness weakness, and the same weighted reward is used both to train the agents and as the Figure 2 evaluation metric. Consequently the later-reported correctness improvement is partly manufactured by the post hoc re-weighting of the evaluation/training objective; it is not an independent confirmation that self-training improved correctness.

full rationale

The main circularity is evaluative rather than mathematical: the same LLM-based rewards appear as the trajectory-selection signal, the training objective, the checkpoint-selection criterion on the test set, and the reported test metric in Figure 2. The paper even documents the reward's verbosity bias (0.528 to 0.584 by length alone), which demonstrates that optimizing the reward is not the same as improving quality. The LiveRAG competition result (7th place, Table 1) is a genuinely external benchmark, but it contains no vanilla RAG baseline, so it cannot validate the paper's central comparative claim; therefore it does not neutralize the circularity of Figure 2. Self-citations ([13]–[15], [20]) are present but not load-bearing for the main claim. Overall score 6: the central comparative claim is partially forced by using the training reward as the test metric with test-set checkpoint selection.

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

The central empirical claim depends on the validity of the LLM-based reward models as proxies for the hidden competition metric, on the hand-chosen DataMorgana data distribution, and on the independence assumption used to justify the self-training objective. No new theoretical entities are postulated.

free parameters (4)
  • Correctness reward weight = 4 (faithfulness weight: 1)
    Chosen post hoc after observing that mRAG 'performs worst on the correctness reward' (Section 3.2, Reward Model paragraph). This weighting shapes which trajectories are selected for training and therefore what the system is optimized to produce.
  • DataMorgana category mixture = 10 categories grouped into 5 combinations, with per-category probabilities such as 0.5/0.5, 0.35/0.35/0.1/0.1/0.05/0.05
    The training and test distributions are defined by hand-chosen schemas in Appendix A and Figures 3 through 12. The claimed superiority of mRAG is only demonstrated on this self-selected synthetic distribution.
  • Trajectory sampling temperature = 0.7 (nucleus sampling) for trainable agents; 0.1 for the generator
    A hyperparameter choice in Appendix D.2 that determines trajectory diversity. It is hand-set rather than fitted, but it directly affects which trajectories are seen and selected.
  • Number of sampled trajectories T = 8
    Set in Appendix D.2. Larger T would increase diversity at higher inference cost; the retained best trajectories become the training labels.
assumptions (4)
  • domain assumption Agent actions are conditionally independent given the input; the trajectory probability factorizes as p(a1|x)p(a2|x)...p(an|x).
    Section 3.2, 'Optimization through Self-Training': this factorization replaces the joint distribution and justifies the supervised objective. If agent actions genuinely interact, the training objective is misspecified.
  • domain assumption The LLM-based correctness and faithfulness rewards approximate the competition's hidden scoring function.
    Sections 3.2 and D.3 state 'The competition does not provide a publicly available scoring function.' The entire self-training loop and the in-house evaluation depend on this proxy being valid.
  • domain assumption DataMorgana-generated QA pairs are representative of complex, real-world RAG tasks.
    Section 2: training and test data are generated with DataMorgana using hand-defined category schemas. The conclusion that mRAG is effective for 'real-world settings' generalizes from this synthetic distribution.
  • domain assumption The preprocessed FineWeb corpus used for indexing matches the competition's target corpus.
    Appendix C: retrieval indexes a chunked FineWeb corpus in a 571 GB inverted index. The LiveRAG competition results assume this index is the appropriate target collection for the competition's queries.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CIIR@LiveRAG 2025: Optimizing Multi-Agent Retrieval Augmented Generation through Self-Training." pith.science (2026). https://pith.science/paper/ZSL2TNU5

@misc{pith2026250610844,
  author       = {Pith},
  title        = {Pith review of: CIIR@LiveRAG 2025: Optimizing Multi-Agent Retrieval Augmented Generation through Self-Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSL2TNU5}},
  note         = {Machine review of arXiv:2506.10844}
}
read the original abstract

This paper presents mRAG, a multi-agent retrieval-augmented generation (RAG) framework composed of specialized agents for subtasks such as planning, searching, reasoning, and coordination. Our system uses a self-training paradigm with reward-guided trajectory sampling to optimize inter-agent collaboration and enhance response generation. Evaluated on DataMorgana-derived datasets during the SIGIR 2025 LiveRAG competition, mRAG outperforms conventional RAG baselines. We further analyze competition outcomes and showcase the framework's strengths with case studies, demonstrating its efficacy for complex, real-world RAG tasks.

Figures

Figures reproduced from arXiv: 2506.10844 by the authors.

Figure 1
Figure 1. Overview of multi-agent RAG. After interacting with DataMorgana, we find that using all cate￾gories simultaneously does not yield high-quality outputs. Conse￾quently, we grouped them into five combinations when invoking DataMorgana for data generation. All combinations include the following core dimensions: User Expertise, Question Type, Answer Type, Document Granularity, Interaction Type, and Aspect Granu￾larity. T… view at source ↗
Figure 2
Figure 2. Evaluation of mRAG and baselines using the intro [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 7
Figure 7. • Answer Intent: the definition of each subcategory and its proba￾bility of generating data with DataMorgana are shown in [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figures from the paper (20 more)
Figure 3
Figure 3. Figure 3: User expertise categorization schema and probabil [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Question type categorization schema and probabili [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Answer Intent categorization schema and probabil [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Question Intent categorization schema and proba [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 6
Figure 6. Figure 6: Premise Inclusion categorization schema and prob [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 9
Figure 9. Figure 9: Interaction type categorization schema and proba [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Answer type categorization schema and probabili [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 13
Figure 13. Figure 13: Prompt used with the Non-RAG baseline. You are a helpful assistant. You will be given a question and a set of supporting documents. You can use the documents to answer the question accurately. You should generate a response that is relevant to the question and support…
Figure 14
Figure 14. Figure 14: Prompt used with the RAG baseline that uses re [PITH_FULL_IMAGE:figures/full_fig_p009_14.png]
Figure 15
Figure 15. Figure 15: Prompt used for the Coordinator agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 16
Figure 16. Figure 16: Prompt used for the Searcher agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]
Figure 17
Figure 17. Figure 17: Prompt used for the Planner agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]
Figure 18
Figure 18. Figure 18: Prompt used for the Summarizer agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]
Figure 19
Figure 19. Figure 19: Prompt used for the Reasoner agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p013_19.png]
Figure 20
Figure 20. Figure 20: Prompt used for the Validator agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p014_20.png]
Figure 21
Figure 21. Figure 21: Prompt used for the Generator/Reviser agent in the mRAG framework. [PITH_FULL_IMAGE:figures/full_fig_p015_21.png]
Figure 22
Figure 22. Figure 22: Prompt used for extracting atomic aspects from the expected output for recall-oriented nugget correctness reward [PITH_FULL_IMAGE:figures/full_fig_p016_22.png]
Figure 23
Figure 23. Figure 23: Prompt used for matching atomic aspects between the expected output and generated response for recall-oriented [PITH_FULL_IMAGE:figures/full_fig_p016_23.png]
Figure 24
Figure 24. Figure 24: Prompt used for extracting atomic aspects from the generated output for faithfulness reward. [PITH_FULL_IMAGE:figures/full_fig_p017_24.png]
Figure 25
Figure 25. Figure 25: Prompt used for matching aspects between the generated output and retrieved documents for faithfulness reward. [PITH_FULL_IMAGE:figures/full_fig_p017_25.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Uncertainty Quantification for Retrieval-Augmented Reasoning

    cs.IR 2025-10 conditional novelty 6.0 of 10

    R2C perturbs reasoning states (paraphrasing, rethinking, validating) to score consistency, improving UQ AUROC by over 5% on average for retrieval-augmented reasoning.

  2. SIGIR 2025 -- LiveRAG Challenge Report

    cs.CL 2025-07 conditional novelty 3.0 of 10

    In the SIGIR 2025 LiveRAG Challenge, all 25 active RAG teams beat the no-RAG baseline on LLM-judged correctness, and LLM scores correlated with human scores at r=0.88.

Reference graph

Works this paper leans on

23 extracted references · 6 canonical work pages · cited by 2 Pith papers

  1. [1]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. arXiv:2310.11511 [cs.CL] https://arxiv.org/abs/2310.11511

    arXiv 2023

  2. [2]

    Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. 2024. RAGAs: Automated Evaluation of Retrieval Augmented Generation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, Nikolaos Aletras and Orphee De Clercq (Eds.). Association for Computational Linguisti...

    work page 2024

  3. [3]

    Simone Filice, Guy Horowitz, David Carmel, Zohar Karnin, Liane Lewin-Eytan, and Yoelle Maarek. 2025. Generating Diverse Q&A Benchmarks for RAG Evalu- ation with DataMorgana. arXiv:2501.12789 [cs.CL] https://arxiv.org/abs/2501. 12789

    arXiv 2025

  4. [4]

    Kailash Gogineni, Yongsheng Mei, Karthikeya Gogineni, Peng Wei, Tian-Shing Lan, and Guru Venkataramani. 2024. Characterizing and Optimizing the End-to- End Performance of Multi-Agent Reinforcement Learning Systems. 2024 IEEE International Symposium on Workload Characterization (IISWC) (2024), 224–235. https://api.semanticscholar.org/CorpusID:272602792

    work page 2024

  5. [5]

    Kailash Gogineni, Peng Wei, Tian Lan, and Guru Venkataramani. 2023. Towards Efficient Multi-Agent Learning Systems. arXiv:2305.13411 [cs.MA] https://arxiv. org/abs/2305.13411

    arXiv 2023

  6. [6]

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The Curious Case of Neural Text Degeneration. arXiv:1904.09751 [cs.CL] https: //arxiv.org/abs/1904.09751

    arXiv 2020

  7. [7]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations . https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  8. [8]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning. arXiv:2503.09516 [cs.CL] https://arxiv.org/abs/2503.09516

    arXiv 2025

Show all 23 references
  1. [9]

    To Eun Kim, Alireza Salemi, Andrew Drozdov, Fernando Diaz, and Hamed Za- mani. 2024. Retrieval-Enhanced Machine Learning: Synthesis and Opportunities. arXiv:2407.12982 [cs.LG] https://arxiv.org/abs/2407.12982

    2024 arXiv

  2. [10]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Opti- mization. CoRR abs/1412.6980 (2014). https://api.semanticscholar.org/CorpusID: 6628106

    2014 arXiv

  3. [11]

    Ronak Pradeep, Nandan Thakur, Shivani Upadhyay, Daniel Campos, Nick Craswell, and Jimmy Lin. 2025. The Great Nugget Recall: Automating Fact Extrac- tion and RAG Evaluation with Large Language Models. arXiv:2504.15068 [cs.IR] https://arxiv.org/abs/2504.15068

    2025 arXiv

  4. [12]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  5. [13]

    Alireza Salemi and Hamed Zamani. 2024. Evaluating Retrieval Quality in Retrieval-Augmented Generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (Wash- ington DC, USA) (SIGIR ’24). Association for Computi...

    2024

  6. [14]

    Alireza Salemi and Hamed Zamani. 2024. Learning to Rank for Mul- tiple Retrieval-Augmented Models through Iterative Utility Maximization. arXiv:2410.09942 [cs.CL] https://arxiv.org/abs/2410.09942

    2024 arXiv

  7. [15]

    Alireza Salemi and Hamed Zamani. 2024. Towards a Search Engine for Ma- chines: Unified Ranking for Multiple Retrieval-Augmented Large Language Models. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval (Washingt...

    2024

  8. [16]

    Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J

    Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Elsayed, Hanie Sedghi, Igor ...

  9. [17]

    Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv:2501.09136 [cs.AI] https://arxiv.org/abs/2501.09136

    2025 arXiv

  10. [18]

    Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. InProceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2 (Montreal, Canada) (NIPS’14). MIT Press, Cambridge, MA, USA, 3104–3112

    2014

  11. [19]

    Falcon-LLM Team. 2024. The Falcon 3 Family of Open Models. https: //huggingface.co/blog/falcon3

    2024

  12. [20]

    categorization_name

    Hansi Zeng, Julian Killingback, and Hamed Zamani. 2025. Scaling Sparse and Dense Retrieval in Decoder-Only LLMs. arXiv:2502.15526 [cs.IR] https://arxiv. org/abs/2502.15526 A DETAILS OF CREATING DATA WITH DATAMORGANA This section provides verbatim definitions of all categorizat...

    2025 arXiv

  13. [22]

    The lower temperature reduces randomness, leading to more deterministic outputs, which allows the trainable agents to better adapt to the fixed behavior of the generator/reviser

    with a temperature of 0.1, as it is not trainable due to com- petition constraints. The lower temperature reduces randomness, leading to more deterministic outputs, which allows the trainable agents to better adapt to the fixed behavior of the generator/reviser. To train the a...

  14. [128]

    agent": the name of the agent you want to choose. -

    All training is conducted on a single NVIDIA A100 GPU. D.3 Reward Models The competition does not provide an official scoring function, so we define and implement two complementary reward signals— correctness and faithfulness—based on the provided evaluation guidelines. correc...

    2025

  15. [2024]

    arXiv:2312.06585 [cs.LG] https://arxiv.org/abs/2312.06585

    Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models. arXiv:2312.06585 [cs.LG] https://arxiv.org/abs/2312.06585

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.