Pith. sign in

REVIEW 3 major objections 6 minor 38 references

DeepThink: Aligning Language Models with Domain-Specific User Intents

T0 review · 3 major / 6 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 DeepThink shows that fine-tuning a 7B model on simulated, refined conversations beats a GPT-4-turbo+RAG assistant by 7.92% on real advertising support questions.

desk verdict Clever pipeline for synthetic multi-turn training data, but the headline 7.92% gain is not established because GPT-4-turbo both creates and scores the data; the paper deserves a serious referee but needs human or independent validation. read the letter →

arxiv 2502.05497 v2 pith:PACFHADR submitted 2025-02-08 cs.CL

Yang Li , Mingxuan Luo , Yeyun Gong , Chen Lin , Jian Jiao , Yi Liu , Kaili Huang This is my paper · ORCID
classification cs.CL
keywords DeepThinkinstructionsynthesisdomain-specificQAconversationsimulationretrieval-augmentedSFTLLMalignmentadvertisingdomainGPT-4evaluation
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

DeepThink tests whether a small open model can beat a large commercial assistant on domain-specific support questions by changing the data rather than the model. The paper's central claim is that simulating multi-turn dialogues between a user and an assistant, iteratively refining answers against conversational context and retrieved documents, then fine-tuning Mistral 7B with retrieval-augmented supervised fine-tuning, produces responses that outperform GPT-4-turbo+RAG by an average of 7.92% on 7,801 real advertising-platform questions. The five evaluated dimensions are relevance, completeness, clarity, accuracy, and actionability, with the largest gains in completeness and actionability. A sympathetic reader would care because it offers a recipe for aligning LLMs with hidden user intents in vertical domains without requiring manual annotation or a larger model.

What carries the argument

The load-bearing object is the conversation-simulated instruction synthesis with iterative refinement. DeepThink generates seed questions that copy the linguistic style of sampled real user questions; uses a dual-role prompt (Inquirer and Assistant) to produce multi-turn dialogues capped at three turns, with the Assistant retrieving documents to ground answers and recommending follow-ups; then runs up to three rounds of refinement where GPT-4-turbo scores each answer on relevance, completeness, clarity, accuracy, and actionability, and the refiner revises the answer using conversation context, feedback, and retrieved documents. The best-scoring answer above a threshold becomes the SFT target, and the model is trained with the objective $\mathcal{L}(\Phi) = -\sum_{(x_i,d_i,y_i)} \log p(y_i | x_i, d_i, \Phi)$, conditioning answers on both the question and retrieved documents.

What would settle it

Take a random sample of 500 of the 7,801 real ad-platform questions, present the DeepThink and GPT-4-turbo+RAG responses to human raters in a blinded pairwise comparison, and ask which answer resolves the question; if humans do not prefer DeepThink at rates near the reported 89.69% and 87.58% win rates—or prefer it at or below chance—the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper establishes that instruction data for domain QA should mimic the way real users actually ask, which is often vague, incomplete, and multi-turn. DeepThink's contribution is a three-part pipeline: seed questions stylistically imprinted on real user queries; dual-role conversation synthesis where an Inquirer and an Assistant interact under retrieval-augmented generation; and a conversation-based refiner that iteratively improves answers using a five-dimension assessment. The resulting Mistral 7B, fine-tuned on (question, retrieved document, refined answer) triplets, is claimed to surpass GPT-4-turbo+RAG on real user test sets, with win rates of about 89.7% on the historic subset and 87.6% on the recent subset, and an overall average improvement of 7.92% across the five dimensions.

Load-bearing premise

The entire evaluation rests on GPT-4-turbo's ratings being a trustworthy stand-in for what real users value; if that judge rewards longer, more structured answers regardless of whether they actually solve the user's problem, the reported 7.92% improvement may not hold for human users.

Editorial extensions

If this is right

  • If the central claim holds, a 7B model fine-tuned on DeepThink data can replace or complement a GPT-4-turbo+RAG assistant in the advertising domain, cutting serving cost while improving response quality.
  • The gains concentrate in completeness and actionability, suggesting conversation simulation is what teaches the model to expand vague questions and provide step-by-step guidance.
  • Ablations show that removing conversation-based synthesis, refinement, or both degrades all five metrics, with completeness dropping by 5.12%, 8.34%, and 15.90% respectively, so each component is load-bearing for the reported performance.
  • Incorporating retrieved documents during SFT lowers training loss by about 37% versus non-retrieval SFT, implying the model learns to exploit external knowledge more effectively when documents are part of the input distribution.

Reading between the lines

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

  • Beyond the paper: the same recipe—imitate real queries, simulate dialogues, refine with an LLM judge, and condition SFT on retrieved passages—should transfer to other vertical domains with scarce pretraining coverage, such as healthcare, finance, or e-commerce support, though the paper only tests advertising.
  • Beyond the paper: because the judge and the data generator are both GPT-4-turbo, part of the reported 7.92% gain could reflect stylistic preferences of that judge rather than user utility, so a production deployment should re-measure with human raters or online A/B metrics.
  • Beyond the paper: the conversation simulation could be reused as a preference data generator for RLHF or as a way to construct multi-turn dialogue evaluation sets, not just for SFT training data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes DeepThink, a supervised fine-tuning pipeline for domain-specific QA. It synthesizes seed questions by imitating real user queries, simulates multi-turn conversations with a dual-role GPT-4-turbo setup, refines answers iteratively using a GPT-4-turbo assessor and feedback, and fine-tunes Mistral 7B with retrieval-augmented SFT. The central claim is that DeepThink improves average response quality by 7.92% over a GPT-4-turbo+RAG assistant on real advertising-platform user questions, based on GPT-4-turbo scoring across five dimensions (relevance, completeness, clarity, accuracy, actionability). The paper also reports ablations showing the contributions of conversation synthesis, refinement, and RAG-augmented SFT, plus evaluations with additional LLM judges.

Significance. If the reported gains reflect genuine user utility, DeepThink offers a practical and cost-effective method for adapting 7B-parameter models to vertical domains using synthetic conversational data, with a realistic test bed of real user questions. The paper is honest about its limitations, including the lack of human evaluation and A/B testing. However, the significance is currently conditional: the main evidence rests entirely on LLM-as-a-judge scores, and the judge model is the same one used to generate and refine the training data. The appendix's additional judges and the inclusion of real user questions strengthen the empirical basis, but they do not fully resolve the circularity concern.

major comments (3)
  1. [Section 3.3; Figure 20] The central claim that DeepThink outperforms GPT-4-turbo+RAG by 7.92% is evaluated using GPT-4-turbo as the judge, which is the same model used to generate, refine, and select training answers in Section 3.3. Specifically, Section 3.3 selects the answer with the highest GPT-4-turbo score r, and Section 4.1 uses the same model to score final responses on the same five dimensions. This creates a closed loop: the training pipeline is optimized against the exact scoring function used for evaluation. The additional judges in Appendix C (DeepSeek-R1, Llama-3.1-405B) were not used in training-data selection, which mitigates but does not eliminate the concern, since all are LLMs that may share stylistic preferences. The paper's own Limitations section (Section 6) acknowledges that human-in-the-loop evaluation and A/B testing are future work. To support the abstract's claim of beating a production assistant, the paper needs human evaluation (e.g., advertising professionals rating responses) or an online A/B test, or at least a judge model that was not involved in any part of training-data generation or selection.
  2. [Section 4.2; Table 2] The refiner is explicitly prompted to produce 'Step-by-step detailed answer... in 100 words' (Figure 20), and the evaluator prompt (Figure 21/22) instructs the judge not to let length influence evaluation. LLM judges are known to be sensitive to length and structure despite such instructions. Since the training data is refined to match the requested format, the measured gains on completeness and actionability (12.09% and 13.69% in Table 2) may partly reflect format adherence rather than user utility. The paper should analyze the correlation between response length and judge scores, and ideally report results with length-matched or length-controlled baselines.
  3. [Section 4.2; Table 2] The headline 7.92% improvement is the average of per-dimension percentage improvements (3.43, 12.09, 6.69, 3.74, 13.69), not the percentage improvement of the average score (which would be about 7.66%). This aggregation should be stated explicitly in the abstract and Section 4.2, and the numbers should be consistent with the 'Ovr.' column in Table 2. Additionally, no variance, confidence intervals, or significance tests are reported for any of the scores; a single run with no error bars is insufficient to establish a 7.92% improvement, especially given the small differences on some dimensions (e.g., 3.43% for relevance).
minor comments (6)
  1. [Section 4.2] The subsection heading 'Human Preference Evaluation' is misleading because the judgments come from GPT-4-turbo, not from humans. Consider renaming it to 'WinRate against GPT-4-turbo' or similar to avoid implying human judgements.
  2. [Section 4.2; Figure 4] Figure 4 is referenced but not included in the text; the winrate numbers (89.69% and 87.58%) should be reported in a table if the figure is omitted.
  3. [Section 3.3] The selection criterion 'r = rarg max0≤t≤T rt, rt>rθ' is ambiguous; clarify whether the answer with the highest score is chosen only if it exceeds the threshold, and what happens if no refined answer exceeds rθ.
  4. [Appendix B; Table 2] DSF is described as 'performing standard supervised finetuning, without documents in context' but is grouped under 'RAG-augmenting-SFT' in Table 2; clarify the categorization and why it is listed as a RAG-augmented baseline.
  5. [Appendix D; Figure 8] The centroid similarity values are reported as 0.93 for DeepThink and -0.97 for Self Instruct; since the embeddings are from a sentence-transformer model, a cosine similarity of -0.97 is unusual and likely erroneous. Verify the calculation and report the actual similarity metric.
  6. [Appendix C] There are placeholder citations for DeepSeek-R1 and RAFT (shown as '?') in the appendix; these should be completed before publication.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 7.92% improvement is scored by the same GPT-4-turbo model that selected and refined DeepThink's training answers; the 'Human Preference' win rates are also GPT-4-turbo judgments, so the central result is partially circular.

  1. fitted input called prediction [Section 3.3 (Instruction Update and Filtering); Section 4.1 (Evaluation); Section 4.2 (Comparison of Human Preference)]
    "We select the best answer with the highest score in the pool that exceeds a predefined quality threshold, i.e., r = rarg max0≤t≤T rt,rt>rθ. ... we leverage GPT-4-turbo to evaluate the quality of model-generated responses. Specifically, we input the question, the most relevant documents, and the model's response into GPT-4-turbo, prompting it to score the model's answer based on relevance, completeness, clarity, accuracy, and actionability. ... To reflect the degree of human preference, we use the judgments from GPT-4-turbo."

    Training and evaluation share the same fitness function. Section 3.3 uses GPT-4-turbo's five-dimension scores to select which refined answer becomes the SFT target (r = rargmax rt), and Section 4.1 computes the headline quality scores with the same GPT-4-turbo five-dimension rubric. Section 4.2's 'Human Preference' win rate is also 'the judgments from GPT-4-turbo.' The reported 7.92% improvement and 87.58-89.69% win rates therefore measure how well DeepThink reproduces the exact GPT-4-turbo preferences that the refiner was optimized against, not independent user satisfaction. This is partial rather than total circularity: the model must still generalize to unseen test questions, and Appendix C adds two alternative LLM judges, which reduces (but does not eliminate) judge-specific bias.

full rationale

The central derivation chain is: GPT-4-turbo generates seed QAs from documents and 15 real user questions; GPT-4-turbo simulates conversations; GPT-4-turbo refines and scores answers, with the best-scoring answer chosen as SFT data; Mistral 7B is fine-tuned; GPT-4-turbo grades the final outputs on the same five dimensions. The key circularity is that the training-data selection objective and the evaluation metric are the same GPT-4-turbo scoring rubric, so the headline improvement is partially aligned to the judge by construction. The paper's own Limitations (Section 6) states that 'the assessment protocol relied primarily on GPT-4-turbo, DeepSeek-R1, and Llama-3.1-405B for automated evaluation, potentially introducing model-specific biases' and lists human-in-the-loop evaluation and A/B testing as future work, confirming the gap. Appendix C does provide some independence by re-evaluating with DeepSeek-R1 and Llama-3.1-405B, and those judges were not used in training selection, though all are LLMs and may share stylistic preferences. No load-bearing self-citation chain is present: the only co-author self-citation (Ma et al., 2023, with co-author Yeyun Gong) is a related-work reference on query rewriting and does not support the central claim. I also flag completeness issues: Appendix B and Appendix C contain unresolved '(?)' citation placeholders (e.g., for DeepSeek-R1 and the CoT critique), and Appendix E references a figure without a number; these are not circularity but indicate unfinished manuscript passages. Overall, the paper's method is not equivalent to its inputs by definition, because the trained model must generalize to held-out real questions and the evaluation is an empirical measurement; however, because the same judge both selected the training answers and scored the final responses, the central 7.92% result is partially circular and should be scored 6.

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

The central claim depends on several unverified assumptions about the evaluation judge, data representativeness, and knowledge-base completeness. No free parameters are fitted to a held-out validation set; hyperparameters are chosen ad hoc and not shown to be robust.

free parameters (5)
  • max conversation turns = 3
    Set in Appendix A; no sensitivity analysis; larger values might improve or degrade quality.
  • refinement rounds T = 3
    Set in Appendix A; iterations of the refiner; threshold r_theta not specified.
  • quality threshold r_theta = not reported
    Used to select refined answers; value not given, making reproduction ambiguous.
  • retrieval top-k = 3
    Number of document chunks used in RAG; affects grounding and training.
  • LoRA rank r = 64
    LoRA hyperparameter with alpha 16 and dropout 0.05; no sensitivity analysis.
assumptions (4)
  • domain assumption GPT-4-turbo's LLM-based judgments are a valid proxy for real user satisfaction on the five evaluation dimensions.
    Invoked in Section 3.3 (assessor) and Section 4.1 (evaluation); no human validation.
  • domain assumption The 15 randomly sampled real user questions are representative of the general user population and do not overlap with the test set.
    Section 4.1 'Imitation seed data'; overlap would leak style information into training.
  • domain assumption The retrieved document chunks from the 7.5k text segments contain sufficient and accurate information to answer all test questions.
    RAG at synthesis and inference; if the knowledge base is incomplete, answers may be incomplete.
  • domain assumption Supervised fine-tuning on (question, document, answer) triplets generalizes to real user questions.
    Standard SFT assumption, Section 3.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepThink: Aligning Language Models with Domain-Specific User Intents." pith.science (2026). https://pith.science/paper/PACFHADR

@misc{pith2026250205497,
  author       = {Pith},
  title        = {Pith review of: DeepThink: Aligning Language Models with Domain-Specific User Intents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PACFHADR}},
  note         = {Machine review of arXiv:2502.05497}
}
read the original abstract

Supervised fine-tuning with synthesized instructions has been a common practice for adapting LLMs to domain-specific QA tasks. However, the synthesized instructions deviate from real user questions and expected answers. This study proposes a novel framework called DeepThink to generate high-quality instructions. DeepThink first generates a few seed questions to mimic actual user questions, simulates conversations to uncover the hidden user needs, and refines the answer by conversational contexts and the retrieved documents for more comprehensive answers. Experiments demonstrate that DeepThink achieves an average performance improvement of 7.92% compared to a GPT-4-turbo+RAG-based assistant on the real user test set in the advertising domain across dimensions such as relevance, completeness, clarity, accuracy, and actionability.

Figures

Figures reproduced from arXiv: 2502.05497 by the authors.

Figure 1
Figure 1. Three phenomena on real-world advertising [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison of DeepThink and GPT-4-turbo across five evaluation dimensions over dif￾ferent time spans ( "Historic," and "Recent."). Deep￾Think performs better than GPT-4-turbo in relevance, completeness, clarity, accuracy, and actionability. 2 Related Works 2.1 Instruction Data Synthesis To address the issue of limited training samples in specific domains, various works have proposed using additional data… view at source ↗
Figure 3
Figure 3. The framework of DeepThink generated. This feedback is then utilized as input for the subsequent iteration, guiding the refiner to improve the response further. Initialization. Refinement focuses on enriching the current answer by incorporating relevant infor￾mation from the conversation’s other turns. Specifi￾cally, DeepThink feeds synthesized questions, cor￾responding answers, and the related conversation context … view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Human Preference Evaluation (WinRate mod [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Score distribution of the instructions ponent in our model. (2) The removal of CDR has the most signif￾icant impact on the model’s performance. This demonstrates the critical role of CDR in leveraging conversational context to enhance response qual￾ity. Specifically, C…
Figure 6
Figure 6. Figure 6: Training loss trend of DeepThink with and without RAG-augmenting-SFT on Recent Rel. Comp. Clar. Acc. Act. 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 Score SFT RAG-augmenting-SFT [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Performance of SFT and RAG-augmenting￾SFT on Recent We have shown that DeepThink achieves a no￾table improvement regarding all performance met￾rics compared with DeepThink-S in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Top 10 Most Common Root Verbs (Inner) and [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Top 10 Most Common Root Verbs (Inner) and Their Top 3 Direct Noun Objects (Outer) in Evol Instruct [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 8
Figure 8. Figure 8: Similarity between different synthetic data [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 12
Figure 12. Figure 12: Performance between Full parameters and QLoRA finetuning F Case Study F.1 Synthesized Conversations and Refining Responses We present a conversation synthesized by Deep￾Think, as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: The case of conversation between the user and the assistant synthesized by [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: The case of the response after Conversation-based Data Refinement. [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Online platform answer vs. DeepThink(Case 1) [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Online platform answer vs. DeepThink(Case 2) [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 17
Figure 17. Figure 17: The prompt of extracting Seed QAs from documents [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]
Figure 18
Figure 18. Figure 18: The prompt of Inquirer in Conversation-based Data Synthesis [PITH_FULL_IMAGE:figures/full_fig_p017_18.png]
Figure 19
Figure 19. Figure 19: The prompt of Assistant in Conversation-based Data Synthesis [PITH_FULL_IMAGE:figures/full_fig_p018_19.png]
Figure 20
Figure 20. Figure 20: The prompt of Conversation-based Data Refinement [PITH_FULL_IMAGE:figures/full_fig_p019_20.png]
Figure 21
Figure 21. Figure 21: The prompt of the evaluation prompt based on the relevant document [PITH_FULL_IMAGE:figures/full_fig_p020_21.png]
Figure 22
Figure 22. Figure 22: The prompt of the winrate evaluation prompt [PITH_FULL_IMAGE:figures/full_fig_p020_22.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 6 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Yelaman Abdullin, Diego Molla-Aliod, Bahadorreza Ofoghi, John Yearwood, and Qingyang Li. 2024. Synthetic dialogue dataset generation using llm agents. arXiv preprint arXiv:2401.17461

    arXiv 2024

  4. [4]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://api.semanticscholar.org/CorpusID:264288947 Self-rag: Learning to retrieve, generate, and critique through self-reflection . ArXiv, abs/2310.11511

    arXiv 2023

  5. [5]

    XTuner Contributors. 2023. Xtuner: A toolkit for efficiently fine-tuning llm. https://github.com/InternLM/xtuner

    work page 2023

  6. [6]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36

    2024

  7. [7]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://api.semanticscholar.org/CorpusID:52967399 Bert: Pre-training of deep bidirectional transformers for language understanding . In North American Chapter of the Association for Computational Linguistics

    2019

  8. [8]

    Masoomali Fatehkia, Ji Kim Lucas, and Sanjay Chawla. 2024. T-rag: lessons from the llm trenches. arXiv preprint arXiv:2402.07483

    arXiv 2024

Show all 38 references
  1. [9]

    Sagar Goyal, Eti Rastogi, Sree Prasanna Rajagopal, Dong Yuan, Fen Zhao, Jai Chintagunta, Gautam Naik, and Jeff Ward. 2024. Healai: A healthcare llm for effective medical documentation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, pages...

    2024

  2. [10]

    Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. https://doi.org/10.18653/v1/2023.acl-long.806 Unnatural instructions: Tuning language models with (almost) no human labor . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics...

    2023 doi

  3. [11]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. https://api.semanticscholar.org/CorpusID:249097975 Unsupervised dense information retrieval with contrastive learning . Trans. Mach. Learn. Res., 2022

    2021

  4. [12]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  5. [13]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating S...

    2023

  6. [14]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdel rahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2019. https://api.semanticscholar.org/CorpusID:204960716 Bart: Denoising sequence-to-sequence pre-training for natural language generation, trans...

    2019

  7. [15]

    Zihan Liu, Wei Ping, Rajarshi Roy, Peng Xu, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Chatqa: Building gpt-4 level conversational qa models. arXiv preprint arXiv:2401.10225

    2024 arXiv

  8. [16]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting for retrieval-augmented large language models. arXiv preprint arXiv:2305.14283

    2023 arXiv

  9. [17]

    Rui Mao, Guanyi Chen, Xulang Zhang, Frank Guerin, and Erik Cambria. 2023. Gpteval: A survey on assessments of chatgpt and gpt-4. arXiv preprint arXiv:2308.12488

    2023 arXiv

  10. [18]

    Dheeraj Mekala, Tu Vu, Timo Schick, and Jingbo Shang. 2022. https://api.semanticscholar.org/CorpusID:253107930 Leveraging qa datasets to improve generative data augmentation . In Conference on Empirical Methods in Natural Language Processing

    2022

  11. [19]

    OpenAI et al. 2023. https://arxiv.org/abs/2303.08774 GPT-4 Technical Report

    2023 arXiv

  12. [20]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 3...

    2022

  13. [21]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. https://api.semanticscholar.org/CorpusID:256459451 In-context retrieval-augmented language models . Transactions of the Association for Computational Linguistics, 11:1...

    2023

  14. [22]

    Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. 2022. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations

    2022

  15. [23]

    Makesh Narsimhan Sreedhar, Traian Rebedea, Shaona Ghosh, and Christopher Parisien. 2024. Canttalkaboutthis: Aligning language models to stay on topic in dialogues. arXiv preprint arXiv:2404.03820

    2024 arXiv

  16. [24]

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Stanford alpaca: An instruction-following llama model

    2023

  17. [25]

    Jian Wang, Chak Tou Leong, Jiashuo Wang, Dongding Lin, Wenjie Li, and Xiao-Yong Wei. 2024. Instruct once, chat consistently in multiple rounds: An efficient tuning framework for dialogue. arXiv preprint arXiv:2402.06967

    2024 arXiv

  18. [26]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. https://api.semanticscholar.org/CorpusID:254877310 Self-instruct: Aligning language models with self-generated instructions . In Annual Meeting of the Associa...

    2022

  19. [27]

    Zirui Wang, Adams Wei Yu, Orhan Firat, and Yuan Cao. 2021. https://api.semanticscholar.org/CorpusID:237572306 Towards zero-label language learning . ArXiv, abs/2109.09193

    2021 arXiv

  20. [28]

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023. Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564

    2023 arXiv

  21. [29]

    Bicknell

    Austin Xu, Will Monroe, and K. Bicknell. 2024 a . https://api.semanticscholar.org/CorpusID:266875606 Large language model augmented exercise retrieval for personalized language learning . Proceedings of the 14th Learning Analytics and Knowledge Conference

    2024

  22. [30]

    Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. https://api.semanticscholar.org/CorpusID:258298159 Wizardlm: Empowering large language models to follow complex instructions . ArXiv, abs/2304.12244

    2023 arXiv

  23. [31]

    Xuhai Xu, Bingsheng Yao, Yuanzhe Dong, Saadia Gabriel, Hong Yu, James Hendler, Marzyeh Ghassemi, Anind K Dey, and Dakuo Wang. 2024 b . Mental-llm: Leveraging large language models for mental health prediction via online text data. Proceedings of the ACM on Interactive, Mobile,...

    2024

  24. [32]

    Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. 2024 c . Magpie: Alignment data synthesis from scratch by prompting aligned llms with nothing. arXiv preprint arXiv:2406.08464

    2024 arXiv

  25. [33]

    Tianjun Zhang, Shishir G Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E Gonzalez. 2024. Raft: Adapting language model to domain specific rag. arXiv preprint arXiv:2403.10131

    2024 arXiv

  26. [34]

    Xuanyu Zhang and Qing Yang. 2023. Self-qa: Unsupervised knowledge guided language model alignment. arXiv preprint arXiv:2305.11952

    2023 arXiv

  27. [35]

    Wenting Zhao, Xiang Ren, John Frederick Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. 2024. https://api.semanticscholar.org/CorpusID:269390491 Wildchat: 1m chatgpt interaction logs in the wild . ArXiv, abs/2405.01470

    2024 arXiv

  28. [36]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric Xing, et al. 2023. Lmsys-chat-1m: A large-scale real-world llm conversation dataset. arXiv preprint arXiv:2309.11998

    2023 arXiv

  29. [37]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36

    2024

  30. [38]

    Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. arXiv preprint arXiv:2310.17631

    2023 arXiv

Pith tools

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