Pith. sign in

REVIEW 4 major objections 4 minor 31 references

C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation

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

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

Pith's one-line read Training a 1.5B proxy between retriever and LLM lifts RAG accuracy to 63.5% without touching either component.

desk verdict A novel multi-agent proxy for RAG with tree-rollout credit assignment, but an unnamed LLM judge supplies both the training reward and every reported accuracy, so the effect sizes are unquantified. read the letter →

arxiv 2502.06205 v2 pith:YDE5R2BO submitted 2025-02-10 cs.CL cs.AIcs.LG

Guoxin Chen , Minpeng Liao , Peiying Yu , Dingmin Wang , Zile Qiao , Chao Yang , Xin Zhao , Kai Fan This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords retrieval-augmentedgenerationmulti-agentreinforcementlearningcreditassignmenttree-structuredrolloutqueryrewritinginformationfilteringLLMalignmentplug-and-playproxy
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

This paper claims that the gap between a frozen retriever and a frozen large language model can be closed by a small trained proxy rather than by fine-tuning either component. The proxy contains three specialized agents—a Reasoning Router that decides whether retrieval is needed and whether the question is simple or complex, an Information Filter that discards irrelevant documents, and a Decision Maker that runs an iterative retrieve-and-filter loop under an LLM-generated plan. The agents are trained jointly with reinforcement learning, using a tree-structured rollout that assigns each agent a credit equal to the average final reward over all reasoning paths through its action. Across six in-domain question-answering datasets, the proxy raises average accuracy to 62.08% at 0.5B parameters and 63.53% at 1.5B parameters, against 52.71% for the strongest full-coverage baseline, with the largest gains on multi-hop questions. The paper also reports that the same proxy generalizes to unseen datasets, a web-search retriever, and LLMs it never saw during training.

What carries the argument

The load-bearing mechanism is the tree-structured rollout with Monte Carlo credit assignment. During training, for each question the Reasoning Router is forced to expand all top-level strategies—[No Retrieval], single-pass [Retrieval], and [Planning]—and then later actions are sampled stochastically with branching factor $K(t)=2$ for depths up to 4 and 1 beyond, building a tree of full trajectories whose leaves carry the system reward ($R=1$ for a correct final answer, $0$ otherwise). Each agent's action node receives the credit reward $r_{credit}(s^i_t, a^i_t) = (1/|L|)\sum_{\ell \in L} R_\ell$, the average of the leaf rewards reachable from that node, and these expectation-based credits are fed into PPO. This converts a single sparse question-level reward into dense per-action supervision and redistributes sampling effort from the question level to the action level.

What would settle it

Run the exact same trained proxies and baselines on the six in-domain datasets, but score answers with exact match after light normalization or with human raters blind to which system produced each answer; if the C-3PO advantage shrinks to near zero or reverses on single-hop sets, then the reported gains come from judge acceptance rather than from the credit-assignment training.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that misalignment between an independently built retriever and LLM is addressable as a learned communication problem: a single small proxy model, prompted with different role instructions, can act as all three go-betweens and be optimized end-to-end against the final answer quality. The proxy chooses among direct answering, single-pass retrieval, and a planning-driven iterative loop, so the LLM is called only once for the plan and once for the final answer. Training uses multi-agent PPO where the retriever and LLM are treated as a fixed environment, and the paper's proposed tree-structured rollout with Monte Carlo credit assignment is what makes joint optimization stable, converting one sparse system-level 0/1 reward into dense per-action expected rewards. With that procedure the trained 0.5B and 1.5B proxies outperform all compared baselines on average, and hold up when the retriever, datasets, and LLM server are swapped for unseen ones.

Load-bearing premise

The reported accuracies all rest on an LLM judge that decides whether the predicted answer is semantically equivalent to the gold answer, and the paper does not name that judge or report how often it agrees with human raters; if that judge only prefers the proxy's longer, structured outputs, the measured gains overstate real correctness.

Editorial extensions

If this is right

  • A trained proxy can act as a drop-in adapter for closed or commercial RAG stacks: it requires no retriever or LLM fine-tuning and keeps working when the retriever is replaced by Google Search and the LLM by models such as Llama3.3-70B or GPT-4o-mini.
  • Multi-hop reasoning is where the method pays off most, with reported gains of +15.5 points on 2WikiMultiHopQA, +13.2 on HotpotQA, and +12.9 on MuSiQue over the strongest baseline.
  • The tree-structured rollout makes RL training stable enough that, unlike single-trajectory reward assignment, it reliably improves on the supervised warm-up model.
  • The whole adapter can be a 0.5B or 1.5B model adding about 4.8 seconds per question, which the paper argues is practical for edge deployment.

Reading between the lines

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

  • Editorial inference: the same tree-rollout credit assignment should transfer to other multi-stage language-agent pipelines—tool selection, web browsing, or code repair—where the only reward signal arrives at the end.
  • Editorial inference: because all rewards come from an LLM judge, a controlled study that varies only output formatting (forcing baselines to emit the same Thought/Action traces) would separate content gains from presentation gains.
  • Editorial inference: the training dynamics show the [Planning] strategy rising to roughly 60-70% of decisions, so an ablation that removes the LLM-generated roadmap would test how much of the multi-hop gain depends on the plan rather than on the proxy's retrieval decisions.
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 / 4 minor

Summary. This paper proposes C-3PO, a proxy-centric alignment framework for retrieval-augmented generation. A small proxy model implements three agents (Reasoning Router, Information Filter, and Decision Maker) that decide whether retrieval is needed, generate and rewrite queries, filter retrieved documents, and plan multi-step retrieval, while the retriever and the LLM remain frozen. The proxy is trained with a tree-structured rollout and Monte Carlo credit assignment under PPO, with the system-level reward determined by an LLM judge. Experiments on six in-domain datasets and two out-of-distribution datasets with unseen retrievers and LLMs report large gains over baselines (average 62.08% with a 0.5B proxy and 63.53% with a 1.5B proxy, versus 52.71% for the strongest full-coverage baseline), plus an HLE score of 5.79%.

Significance. If the reported results hold under a transparent evaluation, C-3PO is a valuable plug-and-play method: it avoids modifying the retriever or the LLM, uses only a 0.5B/1.5B proxy, and shows consistent gains across datasets and OOD settings, with code released. The Monte Carlo credit assignment in Eq. (5) is a sensible mechanism for distributing sparse system-level rewards, and the ablations support the value of the tree-structured rollout. The main significance risk is that every reported number, including the RL reward itself, depends on an unidentified LLM judge, so the measured gains could partly reflect judge acceptance rather than answer correctness. The HLE claim also lacks a same-server Standard RAG baseline, making the comparison to base and proprietary models difficult to interpret.

major comments (4)
  1. [Appendix D.2; Sections 5.2 and 6.2] The evaluation metric is an LLM judge that compares predicted answers with gold answers for semantic equivalence, but the judge model, its decoding parameters, and any human-agreement study are not reported. The same binary judgment appears to be the system-level reward used in Eq. (5), so the proxy is optimized to maximize this judge's acceptance. If the judge is biased toward verbose or structured outputs, or is from the same model family as the generator, the gains in Tables 1 and 2 could be inflated. Please identify the judge, report agreement against human labels on a sample of each benchmark, and re-score at least a subset with a different judge or with human annotators.
  2. [Section 6.6, Table 4] The HLE evaluation compares C-3PO with Qwen2.5-72B against the base model without retrieval and against proprietary models, but it does not include a Standard RAG baseline using the same LLM server and the same Google Search retriever. Without such a baseline, the improvement from 4.27% to 5.79% cannot be attributed to C-3PO's proxy; part of the gain could come from retrieval itself. Please add Standard RAG and, if feasible, an intermediate-module baseline under the same server/retriever configuration.
  3. [Section 5.1, Eqs. (1), (5), (9)] The deterministic rollout forces the Reasoning Router to take [No Retrieval], [Retrieval], and [Planning] actions that are not sampled from the policy pi_theta, and Section 5.2 states that all nodes are added to a replay buffer for PPO. The PPO loss in Eq. (9) uses importance ratios with respect to pi_theta_old as if the data were on-policy. Please clarify whether forced branches are included in the policy-gradient updates and how the behavior distribution for those branches is accounted for; if they are included without an off-policy correction, the 'on-policy' description in Section 5 is inaccurate.
  4. [Section 6.2, Tables 1-2] No variance or significance testing is reported. Given that the differences on single-hop datasets are relatively small (about +1.5 to +3.5 percentage points) and the evaluation is an automated judge, it would strengthen the paper to report standard errors over at least three independent runs or bootstrap resamples, especially for the headline averages.
minor comments (4)
  1. [Table 1] The 'Average' column is computed over different dataset subsets for different baselines (e.g., Self-RAG 72B averages only 3 datasets, with value 58.33). This makes the claimed comparison against SKR-KNN's full-coverage average of 52.71 misleading. Please report averages over a common subset or clearly mark the number of datasets included in each average.
  2. [Appendix A.3 and Section 6.5] The text alternates between 'out-of-distribution' and 'out-of-generalization'; please unify the terminology throughout the paper.
  3. [Appendix B.1 and Figure 6] There are several typos: 'we details' should be 'we detail', 'fo each agent' should be 'for each agent', and 'displaies' in the Figure 6 caption should be 'displays'.
  4. [Figure 6] The density plots for NaturalQuestions and PopQA have y-axis values above 1, which is possible for a density but can confuse readers; consider using a probability mass function over depths or adding a note explaining that these are density values.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: C-3PO's credit assignment is a standard Monte Carlo estimator, and its self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained with respect to circularity. The central technical component, Monte Carlo credit assignment in Eq. (5), estimates each agent-node reward as the average system-level reward over reachable leaf nodes; this is a standard expectation estimator, not a fitted quantity that is later renamed as a prediction. The PPO objective in Eqs. (6)-(11) is a conventional RL formulation, and the reported in-domain and out-of-distribution accuracies come from held-out test questions and unseen retrievers/LLMs, so the main generalization claim is not forced by the training fit. The self-citations (Chen et al. 2024a, 2024b, 2025) appear only in general background citation lists for RAG and multi-agent systems and are not used as load-bearing evidence for any uniqueness claim or core derivation. One evaluation-validity concern exists: Appendix D.2 describes an LLM-based semantic-equivalence judge, and the same kind of judge supplies the binary system-level reward used in RL training, while its identity and human-agreement rate are not reported. However, this is a potential measurement-bias issue rather than a circularity: optimizing toward a metric and then reporting that metric is standard practice, and the paper does not define its conclusion in terms of the training reward or fit the judge to the test outputs. Therefore, no circular step can be exhibited from the paper's own equations, and the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The framework is empirical; it relies on standard RL machinery, a hand-designed tree rollout, and an LLM judge for computing rewards. No new physical entities are introduced.

free parameters (6)
  • Branching factor K(t) = 2 for depth <=4, 1 for depth >4
    Hand-chosen schedule in Eq. (2) that controls the tree width and thus the credit assignment estimate; no sensitivity analysis is provided.
  • Maximum tree depth = 13
    Table 7 sets maximal depth 13; this truncates the Monte Carlo tree and may bias credit estimates for deeper reasoning chains.
  • Warm-up epochs = 3
    Supervised warm-up uses 3 epochs (Table 6); this is a training hyperparameter that affects initialization.
  • PPO learning rates = policy 5e-7, value 5e-6
    RL hyperparameters from Table 7; standard but user-chosen.
  • KL coefficient = 0.005
    PPO KL penalty from Table 7; chosen without sensitivity analysis.
  • Number of training questions per dataset = 6000
    Only 8% of available data is used; the random sample seed is not given, affecting reproducibility.
assumptions (4)
  • standard math PPO and GAE provide unbiased policy updates
    Used in Section 5.2 and Appendix A.1 without proof.
  • domain assumption Monte Carlo estimate in Eq. (5) is an unbiased estimator of conditional expected reward
    The tree is truncated at depth 13 and uses small branching, so leaf rewards are a sampled subset, not a full expectation.
  • domain assumption The three-agent decomposition (router, filter, decision maker) is sufficient to capture the coordination needed for RAG
    Section 4 assumes human-like search behavior maps to these three roles; no argument that more or fewer roles are needed.
  • domain assumption The LLM judge in Appendix D.2 measures answer correctness faithfully
    All results depend on this judge; its identity and agreement with human judgments are not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/YDE5R2BO

@misc{pith2026250206205,
  author       = {Pith},
  title        = {Pith review of: C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDE5R2BO}},
  note         = {Machine review of arXiv:2502.06205}
}
read the original abstract

Retrieval-augmented generation (RAG) systems face a fundamental challenge in aligning independently developed retrievers and large language models (LLMs). Existing approaches typically involve modifying either component or introducing simple intermediate modules, resulting in practical limitations and sub-optimal performance. Inspired by human search behavior -- typically involving a back-and-forth process of proposing search queries and reviewing documents, we propose C-3PO, a proxy-centric framework that facilitates communication between retrievers and LLMs through a lightweight multi-agent system. Our framework implements three specialized agents that collaboratively optimize the entire RAG pipeline without altering the retriever and LLMs. These agents work together to assess the need for retrieval, generate effective queries, and select information suitable for the LLMs. To enable effective multi-agent coordination, we develop a tree-structured rollout approach for reward credit assignment in reinforcement learning. Extensive experiments in both in-domain and out-of-distribution scenarios demonstrate that C-3PO significantly enhances RAG performance while maintaining plug-and-play flexibility and superior generalization capabilities.

Figures

Figures reproduced from arXiv: 2502.06205 by the authors.

Figure 1
Figure 1. Overall framework of C-3PO. (Upper left) Essential cognitive capabilities required for effective RAG system interaction in human-guided alignment. (Upper right) Our proxy-centric alignment simulates these human-like interaction through a lightweight multi-agent system with collaborative strategies. (Bottom) The end-to-end optimization pipeline for our multi-agent system. update given their joint action a ∈ A. • R : … view at source ↗
Figure 2
Figure 2. Ablation Study. 70B, and GPT-4o-mini), with consistent improvements rang￾ing from 1.7% to 5.6%. This platform-agnostic performance demonstrates the plug-and-play capability of our method, enabling seamless integration with various retrievers and LLM servers without requiring any modifications. 6.4. Ablation Study Ablation on Training Paradigm. To thoroughly evaluate the effectiveness of different components in our t… view at source ↗
Figure 3
Figure 3. Performance and Efficiency Comparison. 0 50 100 150 200 250 300 Steps 57 58 59 60 61 62 63 64 Average Accuracy (%) C-3PO-1.5B-SFT C-3PO-0.5B-SFT C-3PO-1.5B-RL C-3PO-0.5B-RL [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Average Accuracy of C-3PO during RL training. method (3.6s), C-3PO yields significant performance gains across both in-domain and out-of-generation evaluations. Furthermore, C-3PO outperforms most methods, such as AutoRAG (Yu et al., 2024a) and SlimPLM (Tan et al., 202…
Figure 5
Figure 5. Figure 5: Strategy Ratio in RL training process. Strategy Ratio during RL Training Process. As introduced in the Section 4.2, our C-3PO incorporates three distinct strategies: Direct Answering Strategy ([No Retrieval]), Single-pass Strategy ([Retrieval]<query content>), and Mult…
Figure 6
Figure 6. Figure 6: Depth Distribution in Test set. Depth Distribution [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

  1. [1]

    Consider the specificity, complexity, and clarity of the question

    **Evaluate the Question**: Assess whether a precise answer can be provided based on the existing knowledge of LLM. Consider the specificity, complexity, and clarity of the question

    work page

  2. [2]

    Guo, T., Chen, X., Wang, Y ., Chang, R., Pei, S., Chawla, N

    URL https://openreview.net/forum?id= 4jqOV6NlUz. Guo, T., Chen, X., Wang, Y ., Chang, R., Pei, S., Chawla, N. V ., Wiest, O., and Zhang, X. Large language model based multi-agents: A survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI 2024, Jeju, South Korea, August 3-9, 20...

    work page 2024

  3. [3]

    Ho, X., Duong Nguyen, A.-K., Sugawara, S., and Aizawa, A

    URL https://www.ijcai.org/proceedings/ 2024/890. Ho, X., Duong Nguyen, A.-K., Sugawara, S., and Aizawa, A. Constructing a multi-hop QA dataset for compre- hensive evaluation of reasoning steps. In Scott, D., Bel, N., and Zong, C. (eds.), Proceedings of the 28th Inter- national Conference on Computational Linguistics, pp. 6609–6625, Barcelona, Spain (Onlin...

    arXiv 2024

  4. [4]

    URL https: //doi.org/10.48550/arXiv.2410.21276

    doi: 10.48550/ARXIV .2410.21276. URL https: //doi.org/10.48550/arXiv.2410.21276. Izacard, G., Caron, M., Hosseini, L., Riedel, S., Bojanowski, P., Joulin, A., and Grave, E. Unsupervised dense infor- mation retrieval with contrastive learning. Trans. Mach. Learn. Res., 2022, 2022. URL https://openreview. net/forum?id=jKN1pXi7b0. Joshi, M., Choi, E., Weld, ...

    arXiv doi:10.48550/arxiv.2410.21276

  5. [5]

    The predicted answer is considered correct if it matches ANY ONE of the provided correct answers

    work page

  6. [6]

    The matching can be exact or semantically equivalent to any of the correct answers

    work page

  7. [7]

    URL https: //doi.org/10.48550/arXiv.2403.08295

    doi: 10.48550/ARXIV .2403.08295. URL https: //doi.org/10.48550/arXiv.2403.08295. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V ., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., and Schulman, J. Webgpt: Browser-assisted question-answering with hu- man fee...

    arXiv doi:10.48550/arxiv.2403.08295

  8. [13]

    **Decision Categories:** - If the question is complex and requires a planning phase before retrieval, your response should be: [Planning] - If the question requests specific information that you believe the LLM does not possess or pertains to recent events or niche topics outside LLM’s knowledge scope, format your response as follows: [Retrieval] ‘YOUR QU...

    work page

Show all 31 references
  1. [14]

    Concentrate solely on determining the necessity for retrieval or planning

    **Focus on Assessment**: Avoid providing direct answers to the questions. Concentrate solely on determining the necessity for retrieval or planning. State of Reasoning Router Now, process the following question: Question: {question} Output (All possible Actions) of Reasoning R...

  2. [15]

    **Analyze Relevance**: Evaluate each document whether it aligns with the objectives of the current retrieval step and contains a direct answer to the question

  3. [16]

    **Thought Process**: Provide a brief analysis for each document, considering both the answer content and the retrieval objectives

  4. [17]

    **Filter Documents**: After your thought process, generate a list of document indices indicating which documents to retain. State of Information Filter Now, process the following question: Current step’s objectives: {objective} (only for [Planning] mode) Question: {question} D...

  5. [18]

    **Evaluate Existing Documents**: Assess the existing documents to determine if it is sufficient to answer the question

  6. [19]

    **Follow the Plan**: Understand the next steps outlined in the plan

  7. [20]

    **Decision Categories:** - If the existing documents is insufficient and requires additional retrieval, respond with: [Retrieval] ‘YOUR QUERY HERE‘ - If the existing documents is adequate to answer the question, respond with: [LLM]

  8. [21]

    **Focus on Action**: Do not answer the question directly; concentrate on identifying the next appropriate action based on the existing documents, plan, and question. State of Decision Maker Now, process the following question: Existing Documents: {accumulated documents} Roadma...

  9. [22]

    Determine what key information we currently know and what additional information is needed through retrieval

    **Question Analysis**: Identifying the core components of the question. Determine what key information we currently know and what additional information is needed through retrieval

  10. [23]

    Focus on the planning process rather than providing direct answers

    **Step By Step Planning**: Develop a detailed plan step by step. Focus on the planning process rather than providing direct answers

  11. [24]

    **Focus on Planning**: Keep your response clear and structured, concentrating solely on the analysis and planning aspects. Now, process the following question: Question: {question} Example of generated roadmap (Take What nationality is the director of film The Caper Of The Gol...

    2023

  12. [25]

    Consider semantic equivalence, not just exact string matching

  13. [26]

    Ignore minor differences in formatting, spacing, or capitalization

  14. [27]

    For numerical answers, consider acceptable margin of error if applicable

  15. [28]

    For text answers, focus on the core meaning rather than exact wording

  16. [31]

    True” if the predicted answer is correct, or “False

    Return only “True” if the predicted answer is correct, or “False” if it is incorrect. Now, process the following question: Question: {question} Correct Answer: {true answer} Predicted Answer: {long answer} 22

  17. [463]

    naacl-long.463

    URL https://doi.org/10.18653/v1/2024. naacl-long.463. Sun, H.-L., Sun, Z., Peng, H., and Ye, H.-J. Mitigating visual forgetting via take-along visual conditioning for multi-modal long cot reasoning. In ACL, 2025a. Sun, H.-L., Zhou, D.-W., Li, Y ., Lu, S., Yi, C., Chen, Q.-G., ...

    2024 doi

  18. [691]

    findings-emnlp.691

    URL https://doi.org/10.18653/v1/2023. findings-emnlp.691. Wei, Z., Chen, W.-L., and Meng, Y . InstructRAG: Instruct- ing retrieval-augmented generation via self-synthesized rationales. In The Thirteenth International Conference on Learning Representations , 2025. URL https:// ...

    2023 doi arXiv

  19. [2017]

    doi: 10.18653/v1/P17-1147

    Association for Computational Linguistics. doi: 10.18653/v1/P17-1147. URL https://aclanthology. org/P17-1147/. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A. P., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., Toutanova, K., Jones, L., Kelc...

    arXiv

  20. [2018]

    doi: 10.18653/v1/D18-1259

    Association for Computational Linguistics. doi: 10.18653/v1/D18-1259. URL https://aclanthology. org/D18-1259/. Yu, T., Zhang, S., and Feng, Y . Auto-rag: Au- tonomous retrieval-augmented generation for large lan- guage models. CoRR, abs/2411.19443, 2024a. doi: 10. 48550/ARXIV ...

    doi arXiv

  21. [2023]

    URL https: //doi.org/10.48550/arXiv.2308.03281

    doi: 10.48550/ARXIV .2308.03281. URL https: //doi.org/10.48550/arXiv.2308.03281. Loshchilov, I. and Hutter, F. Decoupled weight decay regu- larization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net,...

    2019 doi arXiv

  22. [2024]

    URL https: //doi.org/10.48550/arXiv.2407.21783

    doi: 10.48550/ARXIV .2407.21783. URL https: //doi.org/10.48550/arXiv.2407.21783. Fan, W., Ding, Y ., Ning, L., Wang, S., Li, H., Yin, D., Chua, T., and Li, Q. A survey on RAG meeting llms: Towards retrieval-augmented large language models. In Baeza- Yates, R. and Bonchi, F. (e...

    doi arXiv

  23. [4436]

    doi: 10.18653/V1/2024.ACL-LONG.242

    Association for Computational Linguistics, 2024. doi: 10.18653/V1/2024.ACL-LONG.242. URL https: //doi.org/10.18653/v1/2024.acl-long.242. 12 C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation Tang, Y . and Yang, Y . Multihop-rag...

    2024 doi arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.