Pith. sign in

REVIEW 3 major objections 4 minor 35 references

Gaokerena: A Small Persian Medical Language Model Family

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A small Persian medical model reaches 52.98% on translated medical MMLU, beating its 8B baseline by 6.3 points.

desk verdict The datasets and open models are a real contribution to Persian medical NLP, but the headline claim for Gaokerena-R is an evaluation artifact: the RLAIF model gets five CoT samples plus a verifier while baselines get one direct answer. read the letter →

arxiv 2608.00932 v1 pith:WWUAC3TY submitted 2026-08-02 cs.CL

Mehrdad Ghassabi , Hamidreza Baradaran Kashani , Pedram Rostami , Sadra Hakim , Zahra Kazemi , Audrina Ebrahimi This is my paper
classification cs.CL
keywords Persianmedicallanguagemodellow-resourceNLPRLAIFDirectPreferenceOptimizationchain-of-thoughtquestionansweringuncertaintyestimationMMLU
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a small, open, Persian-speaking medical model can be built from a general multilingual 8B checkpoint and can beat its base model on Persian medical multiple-choice benchmarks with domain data and reasoning training. The authors construct the missing raw materials—a 90-million-token Persian medical corpus, 20,000 physician–patient QA pairs, 18,000 machine-translated training MCQs, and a 3,000-question Persian benchmark—and release them. Their first model, Gaokerena-V, lifts the Aya-Expanse-8B baseline from 46.64% to 49.31% on translated medical MMLU and is the first sub-8B Persian model to pass the Iranian Basic Medical Sciences Entrance Exam threshold. Gaokerena-R adds RLAIF reasoning training and reaches 52.98% with chain-of-thought self-consistency plus an external verifier—3.67 points above Gaokerena-V, 6.34 above baseline—although this gain requires the CoT-SC protocol. The authors also introduce uncertainty heads that output confidence from internal hidden states. They present the models as research artifacts, not clinical tools.

What carries the argument

A two-pronged RLAIF preference-generation framework: Expert-Guided Synthesis, where the teacher (DeepSeek-R) generates a correct chain-of-thought conditioned on the ground-truth answer as the preferred trajectory and the student's wrong answer as the rejected one, and Iterative Self-Correction, where the teacher critiques the wrong answer and the student's corrected retry becomes the preferred trajectory. These preference pairs train the student policy by Direct Preference Optimization. At inference, Gaokerena-R uses chain-of-thought prompting with self-consistency: five sampled reasoning paths, majority answer chosen when 3/5 agree, and Aya-Expanse-8B as external verifier for ties.

What would settle it

Take a random sample of about 100 preference pairs from Section 4.2 and have independent physicians judge whether each preferred chain-of-thought is medically correct, not merely consistent with the answer key. If a substantial fraction of teacher CoTs are factually wrong, the DPO gain cannot be attributed to sound reasoning. Also compare Gaokerena-R with and without the Aya verifier and with a single CoT sample: if the 52.98% collapses to baseline when the verifier and five-sample vote are removed, the claim that RLAIF improved reasoning is weakened.

Watch

Extended reading notes

Core claim

Gaokerena is a family of two 8B Persian medical models built from Aya-Expanse-8B with LoRA. Gaokerena-V is domain-tuned on a new 90-million-token Persian medical corpus plus 20,000 physician–patient QA pairs. Gaokerena-R is trained with two RLAIF pipelines—teacher-corrected chain-of-thought and critique-based self-correction—optimized with DPO. Under chain-of-thought self-consistency with five samples and an Aya verifier, Gaokerena-R scores 52.98% on Persian-translated medical MMLU versus 49.31% for Gaokerena-V and 46.64% for the baseline; with direct prompting it drops to 46.28%. The paper's aim is a research-grade, locally deployable foundation for Persian medical NLP, not yet clinical use

Load-bearing premise

The load-bearing premise is that DeepSeek-R, when shown the ground-truth answer, produces medically correct chain-of-thought and accurate critiques, and that the correctness check reliably identifies wrong student answers; if the teacher's reasoning is plausible but wrong, DPO trains the student toward that wrong reasoning and the measured gain is not genuine clinical competence.

Editorial extensions

If this is right

  • If the gains hold, a small on-device Persian model can outperform its own general 8B base and a domain-tuned sibling on translated medical MMLU and IBMSEE, without cloud inference.
  • The two-step RLAIF recipe—teacher-corrected CoT plus critique-based retries—can be ported to other low-resource languages and specialized domains where expert annotation is scarce.
  • The released corpus, MF3QA dataset, translated MCQs, and benchmark allow other teams to build and compare Persian medical models on shared ground truth.
  • The uncertainty heads provide a token-level confidence signal from internal states, enabling applications to flag low-confidence answers instead of silently returning them.

Reading between the lines

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

  • A direct comparison would give Gaokerena-V and Aya-Expanse the same five-sample CoT-SC budget; the paper does not report this, so the marginal contribution of the RLAIF training alone is not fully isolated from the inference-time protocol.
  • Because the teacher's chain-of-thought is conditioned on ground-truth answers, the framework guarantees consistency with the answer key but not medical soundness of the reasoning itself; auditing distilled trajectories for factual fidelity is a natural next step.
  • The same data and benchmark could be used to test whether the recipe transfers to other backbones (e.g., encoder-decoder or MoE architectures); nothing in the training loop is specific to Aya-Expanse.
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 / 4 minor

Summary. The paper introduces Gaokerena-V and Gaokerena-R, 8B-parameter Persian medical language models initialized from Aya-Expanse-8B. Gaokerena-V is domain-adapted on a newly crawled 90M-token Persian medical corpus and instruction-tuned on MF3QA, a curated set of 20,000 physician-patient QA pairs. Gaokerena-R is post-trained with DPO on roughly 11,000 preference pairs generated by DeepSeek-R through two proposed RLAIF strategies: Expert-Guided Synthesis and Iterative Self-Correction. Evaluation is performed on a manually translated 1,000-question medical MMLU subset and a 2,000-question IBMSEE set. The paper reports that Gaokerena-V improves over Aya-Expanse-8B (49.31% vs. 46.64% on MMLU-fa) and that Gaokerena-R reaches 52.98% with CoT self-consistency and an external verifier. Custom uncertainty heads are also described as a contribution.

Significance. If the central comparison were valid, this would be a useful contribution to low-resource Persian medical NLP: the authors release a medical corpus, MF3QA, translated training/evaluation sets, and two open models. The proposed RLAIF framework for reasoning distillation is also interesting. However, the main empirical claim is currently not supported because the models are evaluated under different inference protocols, and the uncertainty heads are never evaluated. The datasets and models appear genuinely useful for the community, but the paper's headline result needs substantial revision.

major comments (3)
  1. [Section 6.2, Tables 5 and 6] The central claim that Gaokerena-R reaches 52.98% and outperforms Gaokerena-V (49.31%) and Aya-Expanse-8B (46.64%) is confounded by the evaluation protocol. In Table 6, Gaokerena-R is evaluated with CoT prompting, five sampled responses, a 3/5 majority threshold, and an Aya-Expanse-8B verifier for ties, while the two comparators are evaluated with a single direct zero-shot completion. Table 5 shows that Gaokerena-R under the same direct protocol scores 46.28%, below both comparators. The reported advantage is therefore introduced by additional test-time compute and the verifier, not demonstrated to come from the RLAIF/DPO training. Please rerun all models under identical protocols—e.g., all three with CoT-SC+verifier—and report matched-protocol numbers, or temper the claim accordingly.
  2. [Section 4.2] The RLAIF preference construction assumes that DeepSeek-R, when prompted with the ground-truth answer, produces medically correct chain-of-thought and accurate critiques, and that the correctness check correctly identifies the initial response as wrong. If the teacher's reasoning is plausible but incorrect, DPO trains the student toward that wrong reasoning. No audit of the 11k preference pairs or of the teacher's reasoning is reported, so the mechanism claimed to produce Gaokerena-R's advantage is unverified. Add a human-annotated audit of a sample of preference pairs (e.g., 100–200) and report agreement with gold labels, or otherwise quantify the quality of the teacher-generated reasoning.
  3. [Section 5 and Section 6] The uncertainty heads are described in detail but never evaluated. There are no quantitative results showing calibration, AUROC, or correlation with answer correctness, nor a comparison with sampling-based uncertainty baselines. Since the abstract and contributions present uncertainty estimation as a key feature, the manuscript should include an evaluation of these heads or explicitly state that this component is not yet validated.
minor comments (4)
  1. [Abstract and Table 5] The abstract says Gaokerena-V improved MMLU-fa from 46.28% to 49.31%, but 46.28% is the straight-prompt score of Gaokerena-R in Table 5; the Aya baseline is 46.64% in Table 4. Please correct this inconsistency.
  2. [Section 3.3] The reference [8] is used for MedSSS, but [8] is the Phi-3 technical report; the intended reference appears to be [6]. Please fix the citation.
  3. [Table 3] The table caption reads 'Table 4.1' but should be 'Table 3' or a proper caption. The internal numbering is inconsistent.
  4. [Section 6.2] The modified self-consistency procedure with the Aya-Expanse-8B verifier is introduced only in the results section. Please provide details in the methods section, including how the verifier is prompted and its own accuracy, and specify the tie-breaking rule precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the RLAIF claims are evaluated on external benchmarks and the cited prior work is not load-bearing.

full rationale

This paper presents empirical model training and evaluation; there is no formal derivation whose conclusion is presumed. The central comparison (Gaokerena-R vs. Gaokerena-V vs. Aya-Expanse-8B) uses external benchmarks (Persian-translated MMLU subset, IBMSEE) that are disjoint from the training corpora. The RLAIF preference dataset is constructed from training questions with ground-truth labels and teacher-generated chains of thought; it is not reused as an evaluation set, so the DPO-trained model's benchmark accuracy is not an input to its training objective. The only self-citation is PersianMind [10], co-authored by a current author, but it is used merely as a baseline and no conclusion depends on its score. The use of Aya-Expanse-8B as a tie-breaker in Gaokerena-R's CoT-SC protocol is a methodological confound (the baselines are not given the same test-time compute), but it is disclosed and does not make Gaokerena-R's reported accuracy equivalent to Aya's by construction: Gaokerena-R's own sampled answers must fail to reach a 3/5 majority before the verifier is invoked, and the final score is a mixture, not an identity. Concerns about teacher correctness (DeepSeek-R's CoT may be plausible but wrong) are validity risks, not circularity. Therefore no load-bearing circular step is identified.

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

The central claims rest on standard fine-tuning assumptions plus domain-specific data quality assumptions: the base model is suitable, the teacher's reasoning is correct, and the translated benchmarks preserve validity. No new physical entities are introduced. The main free parameters are training and decoding choices (LoRA hyperparameters, DPO beta, self-consistency threshold) plus data filtering decisions.

free parameters (4)
  • LoRA rank, alpha, dropout for Gaokerena-V = stage 1 rank 8 alpha 16 dropout 0.05; stage 2 rank 2 alpha 2 dropout 0.4
    Hand-selected in the hyperparameter table; affects domain adaptation and instruction tuning outcomes.
  • DPO beta (β) = not reported
    Equation 1 defines beta as the DPO strength parameter; no value is given, but the preference optimization result depends on it.
  • Self-consistency majority threshold and verifier choice = 3 of 5 agreement; Aya-Expanse-8B as tie-breaker
    Defined in Section 6.2; changing either would change Gaokerena-R's reported 52.98%.
  • MF3QA minimum answer length filter = 50 tokens; random omission in train/dev
    Section 3.2: answers under 50 tokens randomly omitted in training/dev but not test, shaping the instruction-tuning distribution.
assumptions (4)
  • domain assumption Aya-Expanse-8B is a capable base for Persian medical reasoning and instruction following.
    Chosen as the baseline in Sections 2.2 and 4; no independent evidence that this base is optimal or sufficient for medical reasoning.
  • domain assumption DeepSeek-R produces correct chain-of-thought and accurate critiques when prompted with ground-truth answers.
    Used to label preferred vs rejected trajectories in Section 4.2; errors would be distilled into Gaokerena-R via DPO.
  • domain assumption Translated MMLU and IBMSEE items measure the intended medical knowledge in Persian.
    Evaluation relies on translated items; cross-lingual validity is assumed, not demonstrated with human rater agreement.
  • domain assumption The crawled forums and magazines contain expert-authored, clinically valid content after filtering.
    Claims in Section 3 rely on source curation; no clinician agreement study is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gaokerena: A Small Persian Medical Language Model Family." pith.science (2026). https://pith.science/paper/WWUAC3TY

@misc{pith2026260800932,
  author       = {Pith},
  title        = {Pith review of: Gaokerena: A Small Persian Medical Language Model Family},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWUAC3TY}},
  note         = {Machine review of arXiv:2608.00932}
}
read the original abstract

The integration of artificial intelligence into medical question-answering systems has advanced rapidly; however, research remains predominantly focused on English, leaving low resource languages like Persian significantly underserved. To address this gap, this paper introduces Gaokerena, a novel family of compact Persian medical language models optimized for deployment on consumer grade hardware. As a foundational step toward localized digital healthcare, we first present Gaokerena-V, developed by training a baseline model on a newly curated 90-million-token Persian medical corpus and 20,000 expert-vetted physician Q&A pairs, which improved performance on a translated medical MMLU benchmark from 46.28% to 49.31%. Second, recognizing the critical demands of clinical reasoning, we developed Gaokerena-R by integrating a Chain-of-Thought approach with two novel Reinforcement Learning with AI Feedback (RLAIF) frameworks to optimize preference-based reasoning. Despite utilizing the same baseline architecture and a smaller dataset than Gaokerena-V, Gaokerena-R achieved a superior benchmark score of 52.98%. Furthermore, both models are equipped with custom-developed uncertainty heads that predict the model's confidence in its responses based solely on internal hidden states. While these results demonstrate significant progress in Persian medical language modeling and proactive safety estimation, current performance levels remain insufficient for direct clinical application, highlighting the necessity for further research into robust knowledge acquisition and rigorous safety verification prior to real world deployment.

Figures

Figures reproduced from arXiv: 2608.00932 by the authors.

Figure 1
Figure 1. Medical corpus sources language no. tokens collected by English 1.1B I. Garcia Ferrero et al. Spanish 950M I. Garcia Ferrero et al. French 675M I. Garcia Ferrero et al. Italian 143M I. Garcia Ferrero et al. Persian 90M us [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. MF3QA sources 3.3. Persian Medical Multiple Choice Question Dataset While free-form question-answering (QA) datasets are invaluable, the inclu￾sion of multiple-choice question (MCQ) datasets is equally critical for the devel￾opment of robust medical language models. Various training paradigms, such as those utilized for introducing the MedSSS medical language model [8], specif￾ically require high-quality MCQ dataset… view at source ↗
Figure 3
Figure 3. In the subsequent instruction-tuning stage, we transitioned to specialized tuning aimed at refining the model’s conversational fluency. We utilized the full training split of our newly constructed MF3QA dataset, which comprises high-quality, real-world doctor-patient interactions. To further optimize the model for task-specific performance specifically the interpretation of symptoms and the provision of contextually… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: instruction-tuning loss curve to learn medical reasoning behaviors from the teacher. The first strategy, Expert-Guided Synthesis, contrasts the baseline model’s flawed logic with expert-level reasoning. For a given question x, the student model πS produces an initial r…
Figure 7
Figure 7. Figure 7: During data synthesis, we find that Expert-Guided Synthesis is the dominant contributor to the preference dataset construction; for every 19 preference pairs generated by this strategy, typically only one pair is successfully synthesized through Iterative Self-Correcti…
Figure 5
Figure 5. Figure 5: DPO loss curve their exact token spans, and trains the classification head using ternary labels 1 for hallucinated tokens, 0 for supported ones, and -100 to mask out all non￾claim tokens (e.g., stopwords, punctuation, and auxiliary phrases) so they do not contribute to…
Figure 6
Figure 6. Figure 6: Method 1 Block Diagram construction pipeline that carefully preserves both semantic meaning and struc￾tural consistency across languages. This involves translating and re-validating each atomic claim to ensure it remains a self-contained, verifiable unit in Persian bef…
Figure 7
Figure 7. Figure 7: Method 2 Block Diagram uncertainty signal. This yields a training set that is fully compatible with the gaokerena tokenizer’s vocabulary and seamlessly integrates with the uncertainty head’s learning objective, enabling effective cross-lingual transfer while main￾taini…
Figure 8
Figure 8. Figure 8: Pass@k results on the Persian Translated Medical Portion of MMLU dataset [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Pass@k results on the IBMSEE Sept2023 dataset [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 14 canonical work pages

  1. [1]

    Attention is all you need

    Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017)

    work page 2017

  2. [2]

    Medreason: Eliciting factual medical reasoning steps in llms via knowledge graphs. ArXiv, abs/2504.00993, 2025

    Wu, Juncheng, et al. "Medreason: Eliciting factual medical reasoning steps in llms via knowledge graphs. ArXiv, abs/2504.00993, 2025."

    arXiv 2025

  3. [3]

    SINA-BERT: A pre-trained Language Model for Analysis of Medical Texts in Persian

    Taghizadeh, Nasrin, et al. "SINA-BERT: a pre-trained language model for analysis of medical texts in Persian." arXiv preprint arXiv:2104.07613 (2021)

    work page Pith review arXiv 2021

  4. [4]

    Small language models learn enhanced reasoning skills from medical textbooks

    Kim, Hyunjae, et al. "Small language models learn enhanced reasoning skills from medical textbooks." arXiv preprint arXiv:2404.00376 (2024). 26

    arXiv 2024

  5. [5]

    Chain-of-thought prompting elicits reasoning in large language models

    Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in neural information processing systems 35 (2022): 24824-24837

    work page 2022

  6. [6]

    MedS3: Towards Medical Slow Thinking with Self- Evolved Soft Dual-sided Process Supervision

    Jiang, Shuyang, et al. "MedS3: Towards Medical Slow Thinking with Self- Evolved Soft Dual-sided Process Supervision." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 37. 2026

    work page 2026

  7. [7]

    MedMobile: A mobile-sized language model with clinical capabilities

    Vishwanath, Krithik, et al. "MedMobile: A mobile-sized language model with clinical capabilities." arXiv preprint arXiv:2410.09019 (2024)

    work page arXiv 2024

  8. [8]

    Phi-3 Technical Report: A Highly Capa- ble Language Model Locally on Your Phone

    Abdin, Marah, et al. "Phi-3 Technical Report: A Highly Capa- ble Language Model Locally on Your Phone." arXiv, 30 Aug. 2024, arxiv.org/abs/2404.14219

    arXiv 2024

Show all 35 references
  1. [9]

    Aya expanse: Combining research breakthroughs for a new multilingual frontier

    Dang, John, et al. "Aya expanse: Combining research breakthroughs for a new multilingual frontier." arXiv preprint arXiv:2412.04261 (2024)

    2024 arXiv

  2. [10]

    Persian- mind: Across-lingualpersian-englishlargelanguagemodel

    Rostami, Pedram, Ali Salemi, and Mohammad Javad Dousti. "Persian- mind: Across-lingualpersian-englishlargelanguagemodel."arXivpreprint arXiv:2401.06466 (2024)

    2024 arXiv

  3. [11]

    Llama 2: Open foundation and fine-tuned chat models

    Touvron, Hugo, et al. "Llama 2: Open foundation and fine-tuned chat models." arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  4. [12]

    Medical mT5: an open-source multilingual text-to-textLLMforthemedicaldomain

    García-Ferrero, Iker, et al. "Medical mT5: an open-source multilingual text-to-textLLMforthemedicaldomain."arXivpreprintarXiv:2404.07613 (2024)

    2024 arXiv

  5. [13]

    Datasets for large language models: A comprehensive survey

    Liu, Yang, et al. "Datasets for large language models: A comprehensive survey." arXiv preprint arXiv:2402.18041 (2024)

    2024 arXiv

  6. [14]

    Chatdoctor: A medical chat model fine-tuned on a large language model meta-ai (llama) using medical domain knowledge

    Li, Yunxiang, et al. "Chatdoctor: A medical chat model fine-tuned on a large language model meta-ai (llama) using medical domain knowledge." Cureus 15.6 (2023). 27

    2023

  7. [15]

    Zhongjing: Enhancing the chinese medical capa- bilities of large language model through expert feedback and real-world multi-turn dialogue

    Yang, Songhua, et al. "Zhongjing: Enhancing the chinese medical capa- bilities of large language model through expert feedback and real-world multi-turn dialogue." Proceedings of the AAAI conference on artificial in- telligence. Vol. 38. No. 17. 2024

    2024

  8. [16]

    Disc-medllm: Bridging general large language mod- els and real-world medical consultation

    Bao, Zhijie, et al. "Disc-medllm: Bridging general large language mod- els and real-world medical consultation." arXiv preprint arXiv:2308.14346 (2023)

    2023 arXiv

  9. [17]

    Huatuogpt, towards taming language model to be a doctor

    Zhang, Hongbo, et al. "Huatuogpt, towards taming language model to be a doctor." arXiv preprint arXiv:2305.15075 (2023)

    2023 arXiv

  10. [18]

    Huatuo-26M, a Large-scale Chinese Medical QA Dataset

    Wang, Xidong, et al. "Huatuo-26M, a Large-scale Chinese Medical QA Dataset." Findings of the Association for Computational Linguistics: NAACL 2025. 2025

    2025

  11. [19]

    MedDialog: Large-scale medical dialogue datasets

    Zeng, Guangtao, et al. "MedDialog: Large-scale medical dialogue datasets." Proceedings of the 2020 conference on empirical methods in nat- ural language processing (EMNLP). 2020

    2020

  12. [20]

    MedAlpaca–an open-source collection of medical con- versational AI models and training data

    Han, Tianyu, et al. "MedAlpaca–an open-source collection of medical con- versational AI models and training data." arXiv preprint arXiv:2304.08247 (2023)

    2023 arXiv

  13. [21]

    K-qa: A real-world medical q&a benchmark

    Manes, Itay, et al. "K-qa: A real-world medical q&a benchmark." arXiv preprint arXiv:2401.14493 (2024)

    2024 arXiv

  14. [22]

    Measuring massive multitask language understand- ing

    Hendrycks, Dan, et al. "Measuring massive multitask language understand- ing." arXiv preprint arXiv:2009.03300 (2020)

    2009 arXiv

  15. [23]

    Deepseek llm: Scaling open-source language models with longtermism

    Bi, Xiao, et al. "Deepseek llm: Scaling open-source language models with longtermism." arXiv preprint arXiv:2401.02954 (2024)

    2024 arXiv

  16. [24]

    Medmcqa: A large-scale multi-subject multi-choice dataset for medical 28 domain question answering

    Pal, Ankit, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. "Medmcqa: A large-scale multi-subject multi-choice dataset for medical 28 domain question answering." Conference on health, inference, and learn- ing. PMLR, 2022

    2022

  17. [25]

    Grok 3 Beta — The Age of Reasoning Agents

    xAI. "Grok 3 Beta — The Age of Reasoning Agents." xAI Blog, 17 Feb. 2025, x.ai/news/grok-3

    2025

  18. [26]

    Gpt-4 technical report

    Achiam, Josh, et al. "Gpt-4 technical report." arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  19. [27]

    Lora: Low-rank adaptation of large language mod- els

    Hu, Edward J., et al. "Lora: Low-rank adaptation of large language mod- els." ICLR 1.2 (2022): 3

    2022

  20. [28]

    Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback

    Lee, Harrison, et al. "Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback." arXiv preprint arXiv:2309.00267 (2023)

    2023 arXiv

  21. [29]

    Direct preference optimization: Your language model is secretly a reward model

    Rafailov, Rafael, et al. "Direct preference optimization: Your language model is secretly a reward model." Advances in neural information pro- cessing systems 36 (2023): 53728-53741

    2023

  22. [30]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Guo, Daya, et al. "Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning." arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  23. [31]

    Uncertainty quantification for large language models

    Shelmanov, Artem, et al. "Uncertainty quantification for large language models." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 5: Tutorial Abstracts). 2025

    2025

  24. [32]

    Mistral7b

    Q. jiang, Albert et al. "Mistral7b" arXiv preprint arXiv:2310.06825 (2023)

    2023 arXiv

  25. [33]

    Qwen2 Technical Report

    Yang, An, et al. “Qwen2 Technical Report.” arXiv Preprint arXiv:2407.10671, 2024

    2024 arXiv

  26. [34]

    Self-consistency improves chain of thought reasoning in language models

    Wang, Xuezhi, et al. "Self-consistency improves chain of thought reasoning in language models." arXiv preprint arXiv:2203.11171 (2022)

    2022 arXiv

  27. [35]

    Largelanguagemonkeys: Scalinginferencecompute with repeated sampling

    Brown, Bradley, etal."Largelanguagemonkeys: Scalinginferencecompute with repeated sampling." arXiv preprint arXiv:2407.21787 (2024). 29

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.