Pith. sign in

REVIEW 5 major objections 6 minor 60 references

RISE: Reasoning Enhancement via Iterative Self-Exploration in Multi-hop Question Answering

T0 review · 5 major / 6 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 The paper claims that an LLM can improve its own multi-hop reasoning by repeatedly decomposing questions, retrieving and reading evidence, critiquing each step, and fine-tuning on those self-generated traces.

desk verdict A plausible self-improvement recipe for multi-hop QA that reports consistent gains, but the missing SFT control leaves the central attribution unproven. read the letter →

arxiv 2505.21940 v1 pith:PGYCWORU submitted 2025-05-28 cs.CL

Bolei He , Xinran He , Mengke Chen , Xianwei Xue , Ying Zhu , Zhenhua Ling This is my paper
classification cs.CL MSC 68T50
keywords multi-hopquestionansweringretrieval-augmentedgenerationself-improvementself-critiquedecompositioniterativefine-tuningself-generatedtrainingdata
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

RISE claims that a language model can get better at multi-hop question answering by exploring its own reasoning paths and then training on the traces it accepts. The loop has three actions: decompose the question into sub-questions, answer each sub-question by retrieving and reading evidence, and critique whether that step actually helps. The accepted steps become training data for a fine-tuning round, and the improved model then generates new, harder questions for the next round. If this works, it matters because current retrieval-augmented pipelines are limited by noisy retrieval and error propagation, and RISE offers a way to improve reasoning without human-annotated reasoning chains or a larger teacher model. On 2WikiMultiHopQA, HotpotQA, and MuSiQue, the paper reports accuracy of 49.40, 40.50, and 21.70, beating Naive RAG and GenGround baselines on the same base model.

What carries the argument

The engine is the self-exploration loop paired with a multi-objective training loss. At each node $t$, the model emits a sub-question $\mathrm{subq}_t = F_d(M,H,q_0)$, answers it from retrieved passages as $\mathrm{suba}_t = F_g(M, \mathrm{subq}_t, r_t)$, and assigns a binary critique $\sigma_t = F_c(M, \mathrm{subq}_t, \mathrm{suba}_t)$. Accepted nodes accumulate in a history $H$ and in three datasets: $D_d$ for decompositions, $D_r$ for retrieve-then-read steps, and $D_c$ for critique judgments. Each round fine-tunes $M$ on $L = \alpha L_d + \beta L_r + \gamma L_c$, where $L_d$ and $L_r$ are autoregressive losses and $L_c$ is a binary classification loss over $\sigma_t$. The updated model then expands the question set through in-context learning, so the loop converts multi-hop QA into a self-supervised curriculum over reasoning steps.

What would settle it

Run RISE with randomized or reversed critique labels while keeping decomposition and retrieve-then-read data unchanged; if 2WikiMultiHopQA accuracy stays near the reported 49.4 percent, the critique signal is not the driver. Alternatively, fine-tune repeatedly on the same seed questions with no question expansion; if the accuracy curve is identical, the self-exploration loop adds nothing beyond in-distribution fine-tuning.

Watch

Extended reading notes

Core claim

The central claim is that a model can bootstrap better multi-hop reasoning from its own exploration traces. Starting from a seed question set, the model iteratively decomposes each question, retrieves and reads evidence for each sub-question, and makes a binary keep-or-discard judgment on the step. The kept steps are recorded as three interlocking datasets: decomposition traces, retrieve-then-read traces, and critique judgments. Each round fine-tunes the model on all three with a multi-objective loss, then uses the updated model to expand the question set. Repeating this loop steadily improves accuracy across iterations and outperforms prompting, RAG, self-reflection, and generate-then-ground baselines on three multi-hop benchmarks, while also transferring to single-hop QA datasets and to a different base model.

Load-bearing premise

The loop assumes that the model's self-generated decompositions and critique labels are accurate enough to teach it something new; if those labels are mostly noise, or if the expanded questions merely replay the benchmark's training distribution, the reported gains may come from in-distribution fine-tuning rather than from the self-exploration mechanism.

Editorial extensions

If this is right

  • Applying RISE to LLaMA-3.1-8B raises accuracy over four rounds, reaching 49.40 on 2WikiMultiHopQA, 40.50 on HotpotQA, and 21.70 on MuSiQue, surpassing the reported retrieval baselines in the same setup.
  • Accuracy improves with each self-exploration round while reasoning length trends downward, suggesting that the model learns to plan more efficient decomposition paths.
  • The same training loop transfers to a different base model (Qwen2.5-7B) and to single-hop QA datasets such as Natural Questions, Web Questions, and TriviaQA, where accuracy also rises across rounds.
  • Ablations show that each of the three objectives contributes: removing decomposition, retrieve-then-read, or self-critique lowers accuracy, and joint training beats training the three capabilities separately.
  • Because the framework only needs the model itself plus a retriever, it offers a path to improving reasoning without manual annotation or a larger teacher model during training.

Reading between the lines

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

  • Editorial inference: if the mechanism works as described, the retriever is the natural next target, since RISE already filters noisy evidence through critique and could likewise fine-tune retrieval with the same self-generated traces instead of treating it as fixed.
  • Editorial inference: the paper's own calibration table shows the model accepts 23.43 wrong steps per 100 on the initial question set, so a decisive control is to train RISE on critique labels supplied by a stronger judge and compare gains; this would isolate whether the model's own critique is the active ingredient.
  • Editorial inference: the decompose-retrieve-critique-train cycle looks applicable beyond question answering, for example to code generation or mathematical proofs where executable checkers can provide cleaner critique labels than the model's self-judgment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes RISE (Reasoning Enhancement via Iterative Self-Exploration), a self-iterative fine-tuning framework for multi-hop question answering (MHQA). Starting from a seed set of 800 questions per dataset sampled from the training splits of 2WikiMultiHopQA, HotpotQA, and MuSiQue, a LLaMA-3.1-8B model performs iterative 'self-exploration' consisting of three actions: question decomposition, retrieve-then-read, and self-critique. The traces are collected as datasets D_d, D_r, D_c and used to fine-tune the model via a weighted multi-objective loss; the model then expands the question set, and the process repeats. The paper reports accuracy improvements over prompting-based and RAG baselines on the three MHQA benchmarks and three single-hop QA datasets, and includes analyses of reasoning length, GPT-4o-based decomposition quality, token consumption, and robustness on Qwen2.5-7B.

Significance. If the reported gains are attributable to the self-exploration mechanism, RISE would be a useful contribution: it combines RAG with a self-improvement loop, reduces reliance on human annotation or larger teacher models, and the paper provides several supporting analyses (Qwen2.5-7B robustness, token-consumption comparison, and detailed prompt templates in the appendix). However, the significance is conditional. The experimental design does not exclude a much simpler explanation for the Table 2 gains, namely ordinary supervised fine-tuning on in-distribution seed questions, and the single-run methodology does not establish that the reported differences are statistically reliable. The paper also contains incomplete appendix sections with missing tables that are referenced as evidence for the central iterative-improvement claim.

major comments (5)
  1. [Section 2.3, Table 2] The central empirical claim that iterative self-exploration causes the accuracy gains is not supported by the experimental design. The seed question set Q0 consists of 800 questions per dataset sampled from the training sets of 2WikiMultiHopQA, HotpotQA, and MuSiQue (Section 2.3, Initialization), and all training data Dd, Dr, Dc are generated from Q0 and its expansions. A standard supervised fine-tuning baseline trained on the same Q0 (e.g., with gold answers, or on the same generated traces without the iterative loop) is absent. Without this control, the differences in Table 2 (e.g., 49.40 vs 43.50 on 2Wiki, 40.50 vs 37.50 on HotpotQA, 21.70 vs 17.80 on MuSiQue) could be explained by the model learning benchmark-specific patterns from a few thousand in-distribution examples rather than by the proposed self-exploration mechanism. This is especially pressing because the self-critique signal is shown to be weak in Table 10.
  2. [Section 4.1, Figure 3] All accuracy numbers in Table 2 and Figure 3(a) are single-run point estimates. No standard deviations, number of random seeds, or statistical significance tests are reported. The claim that RISE 'significantly improves' reasoning accuracy is therefore not established; differences such as 40.50 vs 37.50 on HotpotQA could be within run-to-run variation. The authors should report variance over multiple seeds, or a significance test, and include error bars in Figure 3(a).
  3. [Section 2.2, Table 10] The binary critique labels sigma in Formula 4 are generated by the model itself and then used as training targets in the self-critique loss (Formula 8). Table 10 shows that on Q0 the model's self-judgments are poorly calibrated: 23.43% of wrong answers are judged correct, compared to 28.65% of correct answers judged correct. This means the critique training signal is substantially noisy. The paper treats the self-critique dataset as a key component without addressing the reliability of these labels. The authors should either provide evidence that training on these noisy labels improves downstream answer accuracy beyond the other two losses, or explicitly measure and discuss the impact of label noise. As it stands, the contribution of self-critique to the reported gains is not established.
  4. [Section 4.3, Table 4] The ablation study appears to use only the round-1 training data, but the 'RISE' accuracy in Table 4 (41.13 on 2Wiki) does not match the final multi-iteration model in Table 2 (49.40). The text does not state the iteration round used in Table 4 or the exact training setup for the 'Separate' condition (three LLMs individually trained on one task each). This makes it impossible to attribute the component-wise accuracy drops to the final model's performance. The authors should clarify the iteration setting and, ideally, ablate components at the same iteration as the main results.
  5. [Appendix A.3.3 and A.3.4] The subsections 'The necessity of multiple iterations' and 'Reliability Analysis of Self-Evaluation Mechanism' contain references to 'Table??' and 'Table ??' that are not present in the manuscript. Figure 3(a) claims consistent accuracy improvement across iterations, but the supporting numerical table is missing. Because the multi-iteration improvement is a load-bearing part of the paper's central claim, these omitted tables must be supplied and the text references fixed.
minor comments (6)
  1. [Abstract] The sentence 'MHQA requires integrating evidence from diverse sources while managing intricate logical dependencies, often leads to errors in reasoning' is grammatically incorrect; it should be rephrased, e.g., '...logical dependencies, which often leads to errors in reasoning.'
  2. [Table 2] The column header 'Trival' should be 'TriviaQA'.
  3. [Table 3] The entry '79.3379.33' is missing a space and should be '79.33 79.33'.
  4. [Section 2.3] The reference 'detailed statistics are provided in Appendix 5' should point to the actual appendix table (Table 5 in Appendix A.2.2), not 'Appendix 5'.
  5. [Table 1] The relationship between the weight grid search in Table 1 and the final uniform weighting (alpha=beta=gamma=1) is confusing. The text states that no weight tuning is performed, yet Table 1 shows a sweep on 2Wiki. Please clarify whether Table 1 is a post-hoc analysis or an ablation, and state explicitly that the weights were not selected from this table for the main experiments.
  6. [References] Shi et al. 2024a and Shi et al. 2024b are the same 'Generate-then-ground' paper and should be merged into one reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RISE's accuracy claims are tested on held-out benchmark test sets, and no load-bearing step reduces to its own input by construction.

full rationale

RISE is a self-training framework: its training data Dd, Dr, Dc are generated by the model itself, and the critique target sigma in Eq. 4 is literally the model's own binary judgment reused as the target of Lc in Eq. 8. Self-referential as this loop is, it is not a circular derivation in the sense of this review, because the paper's central claim—reasoning accuracy on MHQA—is evaluated by held-out accuracy, F1, and EM on standard test sets (Table 2, Tables 11-12), not by the training loss or by the model's own judgments. The expansion loop and iterative fine-tuning are similarly data-generation mechanisms whose output is checked against external benchmarks each round (Figure 3, Table 9). Table 10's finding that the model judges 23.43% of wrong answers as correct is a limitation of the self-critique signal, and the absence of an SFT-on-the-same-seed-questions control is a genuine control-comparison gap, but neither makes the test-set gain a restatement of the training data. There are no load-bearing self-citations: the external inspirations cited (Self-Instruct, Self-Ask) are not by the present authors. The only by-construction element is that Lc fits the model's own sigma, which is the proposed self-training mechanism, not a hidden equivalence between premise and conclusion.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the model's self-generated data (decomposition, critique) being useful supervision and on the external retriever supplying relevant evidence. No new entities are introduced; the only hand-set numbers are task weights, stopping thresholds, seed sizes, iteration count, and standard training hyperparameters. The heaviest load-bearing assumptions are the validity of self-generated critique labels and the sufficiency of in-distribution seed questions as a proxy for the task.

free parameters (5)
  • task weights alpha, beta, gamma = 1, 1, 1 (uniform)
    Table 1 explores weight ratios and reports accuracy variation between 37.6 and 44.3 on 2Wiki; the paper chooses uniform weights to avoid manual tuning, but this is still a hand-chosen configuration that affects results.
  • maximum exploration nodes Nmax = 20
    Algorithm 1 sets Nmax=20 as the stopping condition for self-exploration; no sensitivity analysis is provided.
  • seed question count per dataset = 800 per dataset (2,400 total)
    Cold-start Q0 size is chosen without justification; it determines the scale of the self-generated training data.
  • number of self-iteration rounds = 4
    Experiments run for four iterations; no convergence criterion or reported curve beyond round 4.
  • training hyperparameters (learning rate, batch size, cutoff) = 1e-4, 64, 8192
    Reported in A.2.1 and not varied; standard hyperparameters but still selected by the authors.
assumptions (6)
  • domain assumption Self-generated critique labels (sigma) are valid training targets for the self-critique loss Lc
    Sections 2.2 and 2.3 (Formula 4, Formula 8): labels come from the model itself; no external ground truth. Table 10 shows the model's judgments are poorly calibrated, so this assumption is questionable.
  • domain assumption Self-generated decompositions and answers provide useful supervision for improving reasoning
    The training data Dd and Dr are produced by the model being trained; there is no external verification of decomposition quality beyond GPT-4o win-rate analysis.
  • ad hoc to paper Question expansion via in-context generation yields diverse, valid, non-duplicate multi-hop questions
    Section 2.3 Question Expansion, Figure 10: no validation of generated question quality; relies on the model's ability to self-instruct.
  • domain assumption GPT-4o judgments are an acceptable reference for decomposition quality and critique consistency
    Section 4.2 uses GPT-4o as a judge without calibration against human judgments.
  • domain assumption The two-stage web retrieval pipeline returns sufficient relevant evidence
    Section 3 Retrieval: retrieval is an external component not optimized or analyzed for recall; noise in retrieval motivates the method.
  • standard math Standard cross-entropy objectives for sub-question generation, reading, and binary critique transfer to the benchmark
    Formulas 6-8 define standard losses; the transfer to held-out QA accuracy is an empirical claim, not a mathematical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RISE: Reasoning Enhancement via Iterative Self-Exploration in Multi-hop Question Answering." pith.science (2026). https://pith.science/paper/PGYCWORU

@misc{pith2026250521940,
  author       = {Pith},
  title        = {Pith review of: RISE: Reasoning Enhancement via Iterative Self-Exploration in Multi-hop Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PGYCWORU}},
  note         = {Machine review of arXiv:2505.21940}
}
read the original abstract

Large Language Models (LLMs) excel in many areas but continue to face challenges with complex reasoning tasks, such as Multi-Hop Question Answering (MHQA). MHQA requires integrating evidence from diverse sources while managing intricate logical dependencies, often leads to errors in reasoning. Retrieval-Augmented Generation (RAG), widely employed in MHQA tasks, faces challenges in effectively filtering noisy data and retrieving all necessary evidence, thereby limiting its effectiveness in addressing MHQA challenges. To address these challenges, we propose RISE:Reasoning Enhancement via Iterative Self-Exploration, a novel framework designed to enhance models' reasoning capability through iterative self-exploration. Specifically, RISE involves three key steps in addressing MHQA tasks: question decomposition, retrieve-then-read, and self-critique. By leveraging continuous self-exploration, RISE identifies accurate reasoning paths, iteratively self-improving the model's capability to integrate evidence, maintain logical consistency, and enhance performance in MHQA tasks. Extensive experiments on multiple MHQA benchmarks demonstrate that RISE significantly improves reasoning accuracy and task performance.

Figures

Figures reproduced from arXiv: 2505.21940 by the authors.

Figure 1
Figure 1. The upper part of the figure (blue) illustrates an Evidence Aggregation Error, where the Blu-ray re￾lease year of Fire Birds (2015) is mistaken for its theatri￾cal release year. The lower part (green and red) shows a Reasoning Decomposition Error. The incorrect path formulates the sub-question as the production year of The Book of Eli (2009) instead of its release year (2010). pieces of evidence and constructing coh… view at source ↗
Figure 2
Figure 2. A complete iteration cycle in RISE. a) Self-Exploration: Model Mi decomposes complex questions q0 into simpler sub-questions, generates sub-answers via retrieve-then-read, and evaluates their validity, leading to a final answer a0. Interactions are stored as historical data D. b) Iterative Optimization: RISE optimizes model Mi using historical data D to create an enhanced model Mi+1, which generates new questions Qi… view at source ↗
Figure 3
Figure 3. Changes in model accuracy (a) and reasoning [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Evaluating the win rates between the current and previous iterations using GPT-4o to assess model’s [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Changes in the model’s retrieve-then-read capability. (a) Results on simpler datasets (NQ, TriviaQA, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Question Decomposition prompt template. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Retrieve-then-Read prompt template. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Self-Critique prompt template. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: GPT-4o decomposition prompt template. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Multi-hop question generation prompt template. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Example of Question Decomposition task input. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Example of Question Decomposition task output. [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Example of Retrieve-then-Read task input. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Example of Retrieve-then-Read task output. [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]
Figure 15
Figure 15. Figure 15: Example of Self-Critique task input. Case Study: Self-Critique Output flag = True [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]
Figure 16
Figure 16. Figure 16: Example of Self-Critique task output. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 17 canonical work pages

  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. In The Twelfth International Conference on Learning Representations

    2023

  2. [2]

    Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1533--1544

    2013

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  4. [4]

    Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, J \'e r \'e my Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. 2023. Open problems and fundamental limitations of reinforcement learning from human feedback. Transactions on Machine Learning Research

    work page 2023

  5. [5]

    Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. 2024. Self-play fine-tuning converts weak language models to strong language models. In Forty-first International Conference on Machine Learning

    work page 2024

  6. [6]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1--113

    2023

  7. [7]

    Jingfei Du, Edouard Grave, Beliz Gunel, Vishrav Chaudhary, Onur Celebi, Michael Auli, Veselin Stoyanov, and Alexis Conneau. 2021. https://doi.org/10.18653/v1/2021.naacl-main.426 Self-training improves pre-training for natural language understanding . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational L...

    doi:10.18653/v1/2021.naacl-main.426 2021

  8. [8]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv--2407

    2024

Show all 60 references
  1. [9]

    Zhangyin Feng, Xiaocheng Feng, Dezhi Zhao, Maojin Yang, and Bing Qin. 2024. https://doi.org/10.1109/ICASSP48485.2024.10448015 Retrieval-generation synergy augmented large language models . In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Proc...

    2024

  2. [10]

    C aglar G \"u l c ehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, et al. 2023. Reinforced self-training (rest) for language modeling. CoRR

    2023

  3. [11]

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

    2020

  4. [12]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. https://doi.org/10.18653/v1/2020.coling-main.580 Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps . In Proceedings of the 28th International Conference on Computational Li...

    2020 doi

  5. [13]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299

    2022 arXiv

  6. [14]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1--38

    2023

  7. [15]

    Shuyang Jiang, Yuhao Wang, and Yu Wang. 2023. https://arxiv.org/abs/2306.02907 Selfevolve: A code evolution framework via large language models . Preprint, arXiv:2306.02907

    2023 arXiv

  8. [16]

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551

    2017 arXiv

  9. [17]

    Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christopher Potts, and Matei Zaharia. 2023. https://arxiv.org/abs/2212.14024 Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp . Preprint, arXiv:2212.14024

    2023 arXiv

  10. [18]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

    2019

  11. [19]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  12. [20]

    Ruosen Li and Xinya Du. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.452 Leveraging structured information for explainable multi-hop question answering and reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6779--6789, Singapor...

    2023 doi

  13. [21]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let's verify step by step. In The Twelfth International Conference on Learning Representations

    2023

  14. [22]

    Xiao Liu, Hanyu Lai, Hao Yu, Yifan Xu, Aohan Zeng, Zhengxiao Du, Peng Zhang, Yuxiao Dong, and Jie Tang. 2023. Webglm: Towards an efficient web-enhanced question answering system with human preferences. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and...

    2023

  15. [23]

    Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. https://arxiv.org/abs/2308.09583 Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct . Pre...

    2023 arXiv

  16. [24]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.322 Query rewriting in retrieval-augmented large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 53...

    2023 doi

  17. [25]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36

    2024

  18. [26]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. 2021. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332

    2021 arXiv

  19. [27]

    Ansong Ni, Miltiadis Allamanis, Arman Cohan, Yinlin Deng, Kensen Shi, Charles Sutton, and Pengcheng Yin. 2024. Next: Teaching large language models to reason about code execution. In Forty-first International Conference on Machine Learning

    2024

  20. [28]

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5687--5711

    2023

  21. [29]

    Shuofei Qiao, Ningyu Zhang, Runnan Fang, Yujie Luo, Wangchunshu Zhou, Yuchen Eleanor Jiang, Huajun Chen, et al. 2024. Autoact: Automatic agent learning from scratch for qa via self-planning. In ICLR 2024 Workshop on Large Language Model (LLM) Agents

    2024

  22. [30]

    Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922

    2023 arXiv

  23. [31]

    Zhengliang Shi, Shuo Zhang, Weiwei Sun, Shen Gao, Pengjie Ren, Zhumin Chen, and Zhaochun Ren. 2024 a . Generate-then-ground in retrieval-augmented generation for multi-hop question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lingui...

    2024

  24. [32]

    Zhengliang Shi, Shuo Zhang, Weiwei Sun, Shen Gao, Pengjie Ren, Zhumin Chen, and Zhaochun Ren. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.397 Generate-then-ground in retrieval-augmented generation for multi-hop question answering . In Proceedings of the 62nd Annual Meet...

    2024 doi

  25. [33]

    Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. In Findings of the Association for Computational Linguistics: ACL 2023, pages 7059--7073

    2023

  26. [34]

    Avi Singh, John D Co-Reyes, and Rishabh Agarwal. 2024. Beyond human data: Scaling self-training for problem-solving with language models. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models

    2024

  27. [35]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  28. [36]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. https://doi.org/10.1162/tacl_a_00475 ♫ M u S i Q ue: Multihop questions via single-hop question composition . Transactions of the Association for Computational Linguistics, 10:539--554

    2022 doi

  29. [37]

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. https://arxiv.org/abs/2211.14275 Solving math word problems with process- and outcome-based feedback . Preprint, arXiv:2211.14275

    2022 arXiv

  30. [38]

    Tianduo Wang, Shichen Li, and Wei Lu. 2024. Self-training with direct preference optimization improves chain-of-thought reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11917--11928

    2024

  31. [39]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023 a . Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations

    2023

  32. [40]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023 b . https://openreview.net/forum?id=1PL1NIMMrw Self-consistency improves chain of thought reasoning in language models . In The Eleventh International Confer...

    2023

  33. [41]

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023 c . Self-instruct: Aligning language models with self-generated instructions. In The 61st Annual Meeting Of The Association For Computational Linguistics

    2023

  34. [42]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022 a . https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf Chain-of-thought prompting elicits reason...

    2022

  35. [43]

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

    2022

  36. [44]

    Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-Seng Chua. 2024. https://doi.org/10.1145/3589334.3645363 Search-in-the-chain: Interactively enhancing large language models with search for knowledge-intensive tasks . In Proceedings of the ACM Web Conference 2024, WWW...

    2024

  37. [45]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/v1/D18-1259 H otpot QA : A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

    2018 doi

  38. [46]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2024. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36

    2024

  39. [47]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations

    2023

  40. [48]

    Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. 2023. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794

    2023 arXiv

  41. [49]

    Xunjian Yin, Baizhou Huang, and Xiaojun Wan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.87 ALCUNA : Large language models meet new knowledge . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1397--1414, Singapore. Associa...

    2023 doi

  42. [50]

    Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, S Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators. In International Conference on Learning Representations

    2023

  43. [51]

    Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  44. [52]

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason E Weston. 2024. Self-rewarding language models. In Forty-first International Conference on Machine Learning

    2024

  45. [53]

    Eric Zelikman, Eliana Lorch, Lester Mackey, and Adam Tauman Kalai. 2024. Self-taught optimizer (stop): Recursively self-improving code generation. In OPT 2023: Optimization for Machine Learning

    2024

  46. [54]

    Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model. In The Eleventh International Conference on Learning Representations

    2022

  47. [55]

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Liu Yong, and Shen Huang. 2024. https://doi.org/10.18653/v1/2024.naacl-long.96 End-to-end beam retrieval for multi-hop question answering . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Co...

    2024 doi

  48. [56]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  49. [57]

    Denny Zhou, Nathanael Sch \"a rli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, and Ed H. Chi. 2023. https://openreview.net/forum?id=WZH7099tgfM Least-to-most prompting enables complex reasoning in large language mode...

    2023

  50. [58]

    Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zheng Liu, Zhicheng Dou, and Ji-Rong Wen. 2024. https://arxiv.org/abs/2308.07107 Large language models for information retrieval: A survey . Preprint, arXiv:2308.07107

    2024

  51. [59]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  52. [60]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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