Pith. sign in

REVIEW 5 major objections 5 minor 59 references

ComposeRAG: A Modular and Composable RAG for Corpus-Grounded Multi-Hop Question Answering

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

Pith's one-line read A modular, verification-first RAG pipeline with a self-reflection loop claims up to 15% accuracy gains over fine-tuned baselines and over 10% fewer ungrounded answers in low-quality retrieval settings.

desk verdict A well-engineered modular RAG system with a convincing ablation story, but the headline grounding and accuracy gains are not measured cleanly enough to support the abstract's claims. read the letter →

arxiv 2506.00232 v1 pith:LVVAXLFL submitted 2025-05-30 cs.CL

classification cs.CL
keywords multi-hopquestionansweringretrieval-augmentedgenerationmodularpipelineself-reflectionanswerverificationgroundingfidelitydecompositionqueryrewriting
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 argues that monolithic retrieval-augmented generation pipelines, where query reformulation, retrieval, reasoning, and verification are tightly coupled, can be replaced by a set of atomic, composable modules that each act as a parameterized transformation on structured inputs and outputs. It claims that on four multi-hop QA benchmarks, this modular pipeline consistently outperforms fine-tuning-based and reasoning-specialized baselines under matched retrieval conditions, with accuracy gains up to 15% over fine-tuned methods and up to 5% over agentic reasoning pipelines. The load-bearing design is a verification-first loop: every sub-answer and the final answer must be checked against cited passages, and verification failure triggers self-reflection that analyzes the reasoning trace, re-decomposes the question, and re-executes the pipeline up to three times. The paper reports that this design reduces ungrounded answers by over 10% in low-quality retrieval settings and by about 3% even with strong corpora, at the cost of occasionally abstaining when evidence is insufficient. If true, this supports a modular, upgradeable, interpretable alternative to task-specific fine-tuning or monolithic agentic systems.

What carries the argument

The central object is a formal notion of a well-composed system, defined by two properties: ablation, where removing module $M_i$ yields a non-positive performance change ($\Delta P_{-i} = P(S) - P(S \setminus M_i) \ge 0$), and upgrade, where substituting a better module $M'_i$ yields a positive change ($\Delta P_{+i} = P(S[M_i \leftarrow M'_i]) - P(S) > 0$). The mechanism that carries the argument is the Answer Verification module, a binary LLM judgment $(y, r) = \text{Verify}_\theta(q, a, P)$ that checks both factual correctness and grounding against cited passages, together with the Self-Reflection loop that, on final verification failure, runs an error-analysis prompt to produce guided re-decomposition instructions and re-executes the multi-hop pipeline up to three times. Every other module is defined as a parameterized transformation on structured inputs and outputs, which is what makes independent ablation and independent upgrade meaningful.

What would settle it

Take the set of ComposeRAG answers that passed its own Answer Verification, plus the set it abstained on, and have human annotators or an independent non-LLM grounding metric judge them, on the same 500-question subsets used in the paper; if the verification-pass rate is no better than chance at separating grounded from ungrounded answers, or if the abstained answers are mostly correct, then the grounding-fidelity claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that a retrieval-augmented multi-hop QA system can be decomposed into atomic modules—Question Decomposition, Question Construction, Retrieval Decision, Query Rewriting, Passage Reranking, Answer Generation, Answer Verification, and Final Answering—and that this decomposition is not merely cosmetic: each module contributes measurably, and replacing a module with a stronger version improves the whole system. The paper formalizes this as a well-composed system with two properties: removing a module must not improve performance (P1), and upgrading a module to a better version must improve performance (P2). The mechanism that carries the argument is the verification-first loop: every answer must be explicitly supported by cited passages, and when verification fails, the pipeline diagnoses the reasoning trace, generates guided re-decomposition instructions, and re-executes the multi-hop pipeline. The paper reports accuracy gains up to 15% over fine-tuning-based RQ-RAG and up to 5% over reasoning-specialized Search-o1 under identical retrieval conditions, alongside reductions in ungrounded answers of over 10% in low-quality retrieval settings and about 3% with strong corpora.

Load-bearing premise

The load-bearing premise is that the Answer Verification module and the GPT-4o evaluator produce reliable binary judgments of correctness and grounding; if these LLM verdicts are systematically lenient or biased toward ComposeRAG's own output style, the reported accuracy and grounding improvements are not actually measured.

Editorial extensions

If this is right

  • If the claim holds, task-specific fine-tuning of retrieval modules becomes unnecessary: a general instruction-tuned LLM in a modular prompt pipeline can match or beat a fine-tuned system across all four tested benchmarks.
  • If verification-first abstention is the source of the grounding gains, then adding a verification loop to existing monolithic RAG systems may reproduce part of the improvement without redesigning the whole pipeline.
  • If P1 and P2 hold, then upgrading a single module (for example, swapping GPT-4o-mini for GPT-4o in Question Decomposition) yields measurable, additive gains, so progress in individual LLMs transfers directly to the full system.
  • If self-reflection's benefit is monotone in the number of reflection steps (Cover-EM rising from 0.340 to 0.375 and LLM Eval from 0.375 to 0.415 on the MuSiQue-200 subset), then bounded re-decomposition is a cheap robustness mechanism.
  • If the Simple QA pipeline and Retrieval Decision maintain accuracy while cutting average token use by about 32%, then modular routing can improve efficiency without sacrificing answer quality.

Reading between the lines

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

  • A testable extension the paper does not run: apply ComposeRAG's Answer Verification module as a post-hoc filter on Search-o1's outputs, then measure how much accuracy survives when ungrounded answers are removed; this would isolate how much of the grounding difference is due to verification alone rather than to the full modular pipeline.
  • Because the modules are defined entirely by prompts and structured input/output formats, the same decomposition, verification, and self-reflection loop should transfer to other retrieval corpora or tool-augmented settings, such as question answering over private documents; the paper only evaluates Wikipedia-based open-domain benchmarks.
  • An honest reading should treat the reported module-level effect sizes as coarse: the ablations use 200-example subsets and a single GPT-4o evaluator, so the stable signal is likely the ordering and general direction of each module's contribution rather than the exact percentage points.
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 / 5 minor

Summary. The paper introduces ComposeRAG, a modular RAG pipeline for multi-hop question answering in which the reasoning process is decomposed into atomic modules such as Question Decomposition, Question Construction, Query Rewriting, Retrieval Decision, Passage Reranking, Answer Generation, Answer Verification, and Final Answering. A self-reflection loop re-decomposes and re-executes the pipeline when final-answer verification fails. The authors evaluate ComposeRAG on HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle, comparing against RQ-RAG and Search-o1, and report accuracy gains of up to 15% over fine-tuning-based methods and up to 5% over reasoning-specialized pipelines, plus reductions in ungrounded answers of over 10% in low-quality retrieval settings. Ablation and module-upgrade experiments are presented to support the claim that the architecture is well-composed. The code is stated to be released.

Significance. If the empirical claims were fully supported, ComposeRAG would be a useful demonstration that modular, prompt-based pipelines can match or exceed task-specialized models while offering transparency and component-level upgradability. The paper's formal criteria for a well-composed system (P1 and P2), the four-dataset evaluation, and the claimed code release are assets. However, the headline comparisons and the grounding claim are not currently established by controlled evidence, so the significance of the contribution is contingent on substantial revision.

major comments (5)
  1. [Section 5.2.1, Table 1, Appendix B] The RQ-RAG comparison is not a controlled test of ComposeRAG. The comparison changes the LLM (fine-tuned llama2-7b to Llama3.1-8b), the retriever (e5-base-v2 to Cortex Search), and the corpus preprocessing (FlashRAG wiki2018 chunks vs. the authors' own KILT-based preprocessing) simultaneously. The abstract's claim of up to a 15% accuracy improvement over fine-tuning-based methods therefore cannot be attributed to ComposeRAG's modular architecture. Please provide a same-backbone, same-retriever, same-corpus comparison, or explicitly rephrase the claim as a combined system comparison.
  2. [Section 5.2.2, Appendix C] The grounding improvement claim is not directly measured. The abstract states that ComposeRAG reduces ungrounded answers by over 10% in low-quality retrieval settings, but Section 5.2.2 only reports that 18% of Search-o1-only correct answers lack direct support, and Appendix C.2 categorizes Search-o1's failure modes (totaling 10.4% in Table 8). No external groundedness metric, such as NLI-based faithfulness, citation-precision scoring, or human annotation, is applied to ComposeRAG outputs. Please measure grounding and abstention rates directly for ComposeRAG and for each baseline, under both low-quality and standard retrieval conditions.
  3. [Section 4.1.3, Section 5.1] There is a potential circularity in the evaluation. Answer Verification is defined in Section 4.1.3 as a binary LLM judgment, and the main accuracy metric in Section 5.1 uses GPT-4o as the LLM evaluator, while GPT-4o and GPT-4o-mini are also backbones of ComposeRAG. No calibration, human-agreement check, or independent metric is reported for either the verifier or the evaluator. This is especially consequential for the grounding claim, because the same verification module decides which answers are retained and which are abstained on.
  4. [Table 3, Section 6.2] The claim that modules provide distinct and additive contributions is not fully supported by the ablation table. On Hotpot-200, Cover-EM drops from 0.510 with QD+QC+PR to 0.505 when all modules are active, and no significance tests or confidence intervals are provided for the 200-example subsets. The non-monotonic behavior should be reconciled with property P1, and the results should be accompanied by variance estimates or repeated sampling.
  5. [Table 1, Section 5.2.2] The Search-o1 comparison also confounds the architecture with the choice of backbone: ComposeRAG uses Qwen2.5-72B-Instruct while Search-o1 uses QwQ-32B, even though the retriever and corpus are matched. Moreover, on 2WikiMultiHopQA Search-o1 outperforms ComposeRAG (average 0.789 vs. 0.723). The 'up to 5% gain over reasoning-specialized pipelines' should therefore be characterized as a system-level comparison rather than a property of ComposeRAG alone, and the datasets where ComposeRAG does not win should be discussed.
minor comments (5)
  1. [Abstract, Section 1] The phrase '15% accuracy improvement' is ambiguous between absolute percentage-point differences and relative improvements; please state the comparison convention explicitly.
  2. [Appendix D] The prompt examples contain typos, such as 'captital city' in the Question Decomposition example and 'rertieval' in the Retrieval Decision section; these should be corrected before publication.
  3. [Section 1, Code Release] The code release is referenced only as 'Arctic Agentic RAG' without a URL or repository identifier; please provide a complete link or anonymized repository for reproducibility.
  4. [Table 2] The comparison with ReARTeR and R1-Searcher uses numbers reported in other papers under potentially different evaluation settings; the caveat in the caption is appropriate, but the main text should avoid drawing strong conclusions from this comparison.
  5. [Section 5.1, Cover EM] Cover EM rewards verbose predictions that contain a ground-truth span as a contiguous subsequence; consider also reporting exact match or span precision to avoid inflating scores for over-generated answers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the grounding claim is under-measured but not derived from the paper's own inputs.

full rationale

ComposeRAG is an empirical systems paper: the modular abstraction is defined independently in Section 4.1, the answer generation and verification steps are LLM calls rather than fitted parameters, and the headline accuracy numbers come from external multi-hop QA benchmarks against reproduced baselines. No equation in the paper defines a predicted quantity in terms of the input being explained, and no load-bearing claim is justified by a self-citation or by a uniqueness theorem. The one area that invites suspicion is the grounding claim: the abstract says the verification-first design 'reduces ungrounded answers by over 10% in low-quality retrieval settings,' and Section 5.2.2 supports this only by analyzing Search-o1's ungrounded successes (18% of Search-o1-only correct answers; 2.8% overall) plus a manual taxonomy in Appendix C that again concerns the baseline, not ComposeRAG's own ungrounded-answer rate. If 'ungrounded' were operationalized as the Answer Verification module's own binary label, the reduction would be close to a restatement of the module's filter. But the paper never states that operationalization or presents an equation equating the claimed reduction to verifier abstentions; the deficit is a missing external grounding measurement, which is a correctness and evaluation-validity concern rather than a circular derivation. The benchmark accuracy comparisons and ablations are externally grounded, so no specific circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no fitted numerical parameters for a derivation; the main load-bearing assumptions are about LLM judgment reliability and sample representativeness. Design choices (max reflection steps, top-k) affect results but are not fitted to the evaluation data.

free parameters (2)
  • max_reflection_steps = 3
    Manual cap on self-reflection iterations; Figure 5 shows performance keeps rising with more steps, so the chosen value affects the reported self-reflection gains.
  • passage_reranking_top_k = 10
    Inferred from the reranking prompt in Appendix D ('top-10 most relevant passages'); retrieval quality and downstream accuracy depend on this hand-set number, which is not varied or justified.
assumptions (4)
  • domain assumption The Answer Verification module's LLM judgments reliably separate grounded from ungrounded answers.
    Section 4.1.3; the grounding claims hinge on this without independent validation or human agreement data.
  • domain assumption GPT-4o as evaluator produces correctness labels consistent with human judgment.
    Section 5.1 metrics; no human agreement study is reported.
  • domain assumption The Wikipedia corpora (wiki2018 truncated FlashRAG; wiki2019 KILT preprocessed) contain the evidence needed for the sampled questions.
    Section 5.1; unrecoverable samples make this untestable from the paper.
  • domain assumption Random 500/200-question samples are representative of the full dev sets.
    Section 5.1 and Section 6; no seeds or question IDs are released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ComposeRAG: A Modular and Composable RAG for Corpus-Grounded Multi-Hop Question Answering." pith.science (2026). https://pith.science/paper/LVVAXLFL

@misc{pith2026250600232,
  author       = {Pith},
  title        = {Pith review of: ComposeRAG: A Modular and Composable RAG for Corpus-Grounded Multi-Hop Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LVVAXLFL}},
  note         = {Machine review of arXiv:2506.00232}
}
read the original abstract

Retrieval-Augmented Generation (RAG) systems are increasingly diverse, yet many suffer from monolithic designs that tightly couple core functions like query reformulation, retrieval, reasoning, and verification. This limits their interpretability, systematic evaluation, and targeted improvement, especially for complex multi-hop question answering. We introduce ComposeRAG, a novel modular abstraction that decomposes RAG pipelines into atomic, composable modules. Each module, such as Question Decomposition, Query Rewriting, Retrieval Decision, and Answer Verification, acts as a parameterized transformation on structured inputs/outputs, allowing independent implementation, upgrade, and analysis. To enhance robustness against errors in multi-step reasoning, ComposeRAG incorporates a self-reflection mechanism that iteratively revisits and refines earlier steps upon verification failure. Evaluated on four challenging multi-hop QA benchmarks, ComposeRAG consistently outperforms strong baselines in both accuracy and grounding fidelity. Specifically, it achieves up to a 15% accuracy improvement over fine-tuning-based methods and up to a 5% gain over reasoning-specialized pipelines under identical retrieval conditions. Crucially, ComposeRAG significantly enhances grounding: its verification-first design reduces ungrounded answers by over 10% in low-quality retrieval settings, and by approximately 3% even with strong corpora. Comprehensive ablation studies validate the modular architecture, demonstrating distinct and additive contributions from each component. These findings underscore ComposeRAG's capacity to deliver flexible, transparent, scalable, and high-performing multi-hop reasoning with improved grounding and interpretability.

Figures

Figures reproduced from arXiv: 2506.00232 by the authors.

Figure 1
Figure 1. An overview of our modular Multi-Hop QA pipeline, h [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Question Decomposition and Construction Example [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Retrieval Decision Example 4.1.3 Answering with Evidence-Based Verification Ultimately, the goal is to generate faithful, well-grounded answers. The Answer Generation mod￾ule synthesizes a candidate answer for the current sub-question, utilizing the prioritized retrieved passages and, in multi-hop scenarios, the accumulated history of prior sub-questions and their an￾swers. A core tenet is to produce answers that ar… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Iterative Refinement Example: Answer Verification [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Accuracy of the self-reflection mechanism on the Mu [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Accuracy across different generation length inte [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 31 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad , Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shy amal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023

  2. [2]

    Ttqa-rs -a break-down prompting approach for multi-hop table-text question answering with reasonin g and summarization

    Jayetri Bardhan, Bushi Xiao, and Daisy Zhe Wang. Ttqa-rs -a break-down prompting approach for multi-hop table-text question answering with reasonin g and summarization. arXiv preprint arXiv:2406.14732, 2024

  3. [3]

    Improving language models by retrieving from trillio ns of tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Tr evor Cai, Eliza Rutherford, Katie Millican, George Bm V an Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillio ns of tokens. In International conference on machine learning , pages 2206–2240. PMLR, 2022

  4. [4]

    Language mod- els are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, J ared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Aman da Askell, et al. Language mod- els are few-shot learners. Advances in neural information processing systems , 33:1877–1901, 2020

  5. [5]

    Rq-rag: Learning to refine queries for retrieval augment ed generation

    Chi-Min Chan, Chunpu Xu, Ruibin Y uan, Hongyin Luo, Wei Xu e, Yike Guo, and Jie Fu. Rq-rag: Learning to refine queries for retrieval augment ed generation. arXiv preprint arXiv:2404.00610, 2024

  6. [6]

    The faiss library

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Joh nson, Gergely Szilvasy, Pierre- Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jé gou. The faiss library. arXiv preprint arXiv:2401.08281, 2024

  7. [7]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abh inav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten , Alex V aughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. 16

  8. [8]

    A survey on llm-as- a-judge

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Ying- han Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as- a-judge. arXiv preprint arXiv:2411.15594, 2024

Show all 59 references
  1. [9]

    Retrieval aug- mented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval aug- mented language model pre-training. In International conference on machine learning , pages 3929–3938. PMLR, 2020

  2. [10]

    Constructing a multi-hop qa dataset for comprehensive evaluation of reas oning steps

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akik o Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reas oning steps. arXiv preprint arXiv:2011.01060, 2020

  3. [11]

    Atlas: Few-shot learning with retrieval augmented language models

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Ho sseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Y u, Armand Joulin, Sebastian Riedel, and Edou ard Grave. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research , 24(251):1...

  4. [12]

    Survey of hallucination i n natural language generation

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Y u, Dan Su, Y an Xu, Etsuko Ishii, Y e Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination i n natural language generation. ACM computing surveys, 55(12):1–38, 2023

  5. [13]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning

    Bowen Jin, Hansi Zeng, Zhenrui Y ue, Jinsung Y oon, Serca n Arik, Dong Wang, Hamed Za- mani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516 , 2025

  6. [14]

    Flashrag: A m odular toolkit for efficient retrieval-augmented generation research

    Jiajie Jin, Y utao Zhu, Guanting Dong, Y uyao Zhang, Xiny u Y ang, Chenghao Zhang, Tong Zhao, Zhao Y ang, Zhicheng Dou, and Ji-Rong Wen. Flashrag: A m odular toolkit for efficient retrieval-augmented generation research. arXiv preprint arXiv:2405.13576 , 2024

  7. [15]

    Sure: Summarizing retrievals using answer candidates for open- domain qa of llms

    Jaehyung Kim, Jaehyun Nam, Sangwoo Mo, Jongjin Park, Sa ng-Woo Lee, Minjoon Seo, Jung- Woo Ha, and Jinwoo Shin. Sure: Summarizing retrievals using answer candidates for open- domain qa of llms. arXiv preprint arXiv:2404.13081 , 2024

  8. [16]

    Re-rag: Improving open-d omain qa performance and interpretability with relevance estimator in retrieval-a ugmented generation

    Kiseung Kim and Jay-Y oon Lee. Re-rag: Improving open-d omain qa performance and interpretability with relevance estimator in retrieval-a ugmented generation. arXiv preprint arXiv:2406.05794, 2024

  9. [17]

    Retrieval- augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio P etroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rockt äschel, et al. Retrieval- augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing syst...

  10. [18]

    Search-o1: Agentic search-enhanced large re asoning models

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Y uyao Zhang, Y ujia Zhou, Y utao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large re asoning models. arXiv preprint arXiv:2501.05366, 2025

  11. [19]

    Kilt: a benchmark for knowledge intensive language tasks

    Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Y azdani, Nicola De Cao, James Thorne, Y acine Jernite, Vladimir Karpukhin, Jean Maillard, et al. Kilt: a benchmark for knowledge intensive language tasks. arXiv preprint arXiv:2009.02252 , 2020

  12. [20]

    Measuring and narrowing the compositionality gap in langua ge models

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in langua ge models. arXiv preprint arXiv:2210.03350, 2022

  13. [21]

    R1-searcher: Incentivizing the sear ch capability in llms via reinforce- ment learning

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhi peng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the sear ch capability in llms via reinforce- ment learning. arXiv preprint arXiv:2503.05592 , 2025

  14. [22]

    Is chatgpt good at search? investigat ing large language models as re-ranking agents

    Weiwei Sun, Lingyong Y an, Xinyu Ma, Shuaiqiang Wang, Pe ngjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. Is chatgpt good at search? investigat ing large language models as re-ranking agents. arXiv preprint arXiv:2304.09542 , 2023. 17

  15. [23]

    Rearter: Retrieval-augmented reason ing with trustworthy process rewarding

    Zhongxiang Sun, Qipeng Wang, Weijie Y u, Xiaoxue Zang, K ai Zheng, Jun Xu, Xiao Zhang, Song Y ang, and Han Li. Rearter: Retrieval-augmented reason ing with trustworthy process rewarding. arXiv preprint arXiv:2501.07861 , 2025

  16. [24]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, F aisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 , 2023

  17. [25]

    ♪ musique: Multihop questions via single-hop question composition

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot , and Ashish Sabharwal. ♪ musique: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539–554, 2022

  18. [26]

    Plan-rag: Planning-guided retrieval augm ented generation

    Prakhar V erma, Sukruta Prakash Midigeshi, Gaurav Sinh a, Arno Solin, Nagarajan Natarajan, and Amit Sharma. Plan-rag: Planning-guided retrieval augm ented generation. arXiv preprint arXiv:2410.20753, 2024

  19. [27]

    Text embeddings by weakly-supervis ed contrastive pre-training

    Liang Wang, Nan Y ang, Xiaolong Huang, Binxing Jiao, Lin jun Y ang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-supervis ed contrastive pre-training. arXiv preprint arXiv:2212.03533 , 2022

  20. [28]

    Chain-of-thought prompting elicits reas oning in large language models

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

  21. [29]

    Gendec: A robust generative question-decomposition metho d for multi-hop reasoning

    Jian Wu, Linyi Y ang, Y uliang Ji, Wenhao Huang, Börje F Ka rlsson, and Manabu Okumura. Gendec: A robust generative question-decomposition metho d for multi-hop reasoning. arXiv preprint arXiv:2402.11166, 2024

  22. [30]

    Answering complex open- domain questions with multi-hop dense retrieval

    Wenhan Xiong, Xiang Lorraine Li, Srini Iyer, Jingfei Du , Patrick Lewis, William Y ang Wang, Y ashar Mehdad, Wen-tau Yih, Sebastian Riedel, Douwe Kiela,et al. Answering complex open- domain questions with multi-hop dense retrieval. arXiv preprint arXiv:2009.12756 , 2020

  23. [31]

    An Y ang, Baosong Y ang, Beichen Zhang, Binyuan Hui, Bo Zh eng, Bowen Y u, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 techn ical report. arXiv preprint arXiv:2412.15115, 2024

  24. [32]

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

    Zhilin Y ang, Peng Qi, Saizheng Zhang, Y oshua Bengio, Wi lliam W Cohen, Ruslan Salakhut- dinov, and Christopher D Manning. Hotpotqa: A dataset for di verse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600 , 2018

  25. [33]

    Auto-rag: Autono mous retrieval-augmented genera- tion for large language models

    Tian Y u, Shaolei Zhang, and Y ang Feng. Auto-rag: Autono mous retrieval-augmented genera- tion for large language models. arXiv preprint arXiv:2411.19443 , 2024

  26. [34]

    Reasoning over hierarchical question decomposition tree f or explainable question answering

    Jiajie Zhang, Shulin Cao, Tingjia Zhang, Xin Lv, Jiaxin Shi, Qi Tian, Juanzi Li, and Lei Hou. Reasoning over hierarchical question decomposition tree f or explainable question answering. arXiv preprint arXiv:2305.15056 , 2023

  27. [35]

    Processbench: Identify ing process errors in mathematical reasoning

    Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, K eming Lu, Bowen Y u, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identify ing process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559 , 2024

  28. [36]

    #1" or pronouns like

    Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng, Fangkai Y an g, Jia Liu, Shujian Huang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, and Qi Zhang. Efficien trag: Efficient retriever for multi-hop question answering. arXiv preprint arXiv:2408.04259 , 2024. 18 A Supplementary Modules ...

  29. [37]

    What is the company owns the brand that produces iPhones?

  30. [38]

    T o perform the comparison, the populations of Canada and Australia must first be individually achieved

    Who is the CEO of #1? Question: Which country has a larger population, Canada or A ustralia? Reasoning: The question requires a comparative analysis. T o perform the comparison, the populations of Canada and Australia must first be individually achieved. Th en, the comparison c...

  31. [39]

    What is the population of Canada?

  32. [40]

    What is the population of Australia?

  33. [41]

    Australia has a population o f #2

    Canada has a population of #1. Australia has a population o f #2. Which country has a larger population, Canada or Australia? Question: Which river is longer, Nile or Tigris? Reasoning: This question requires determining the lengths of the two rivers to make a comparison. The ...

  34. [42]

    What is the length of Nile?

  35. [43]

    What is the length of Tigris?

  36. [44]

    The length of Tigris is #2

    The length of Nile is #1. The length of Tigris is #2. Which ri ver is longer, Nile or Tigris? Question: What is the capital city of the country where the pa inter of Starry Night was born? Reasoning: The question involves a chain of relationships. First, the painter of Starry ...

  37. [45]

    Who is the painter of Starry Night?

  38. [46]

    What is the country where #1 was born?

  39. [47]

    Once both are identified, their geographical di stance can be obtained

    What is the captital city of #2? Question: What is the distance between the tallest mountain in Japan and the tallest mountain in Nepal? Reasoning: To find the distance, it is first essential to ident ify the tallest mountains in Japan and Nepal in- dividually. Once both are ide...

  40. [48]

    What is the tallest mountain in Japan?

  41. [49]

    What is the tallest mountain in Nepal?

  42. [50]

    First, the country where the Eiffel Tower is loca ted must be found, followed by identifying its capital city

    What is the distance between #1 and #2? Question: What is the distance between the city where the Col osseum is located and the capital city of the country where the Eiffel Tower is located? Reasoning: This question involves two distinct entities an d requires their geographic...

  43. [51]

    What is the country where the Eiffel Tower is located?

  44. [52]

    What is the capital city of #1?

  45. [53]

    What is the city where the Colosseum is located?

  46. [54]

    #number", don’t use the

    What is the distance between #3 and #2? 24 Prompt and T emplate for Question Decomposition Y ou are an expert skilled at analyzing complex questions and decomposing them into distinct and simpler sub-questions. Y our task is to analyze the input question and determine whet her...

  47. [57]

    Instead of asking about where th e company is located, specify that we need the **headquarters** location

    Which city is home to the headquarters of #2? New decomposition instructions 1: The second step is too vague. Instead of asking about where th e company is located, specify that we need the **headquarters** location. Reasoning: The previous decomposition jumps directly from id...

  48. [58]

    Who invented the telephone?

  49. [59]

    Which company was founded by #1?

  50. [60]

    Where are the headquarters of #2?

  51. [61]

    Given a question, an answer, and the supporting passages, ev aluate the answer based on the following criteria: a

    Which city is home to #3? Input fields are: Question: {the input question} Previous Decompositions: {the previous decomposed sub-qu estions and the instructions for improving} Output fields are: Reasoning: {reasoning to the new decomposition} New Decomposition: {the new decompos...

Pith tools

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