Pith. sign in

REVIEW 4 major objections 4 minor 41 references

KaFT: Knowledge-aware Fine-tuning for Boosting LLMs' Domain-specific Question-Answering Performance

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper shows that weighting each training sample by how much it conflicts with the model's existing knowledge improves domain-specific QA fine-tuning, with gains up to 2.40% over standard fine-tuning across four large language models.

desk verdict KaFT is a plausible sample-reweighting method with modest but consistent gains, yet the paper doesn't establish that its score measures knowledge conflict rather than difficulty. read the letter →

arxiv 2505.15480 v2 pith:UOF4TODX submitted 2025-05-21 cs.CL

classification cs.CL
keywords knowledgeconflictsupervisedfine-tuningdomain-specificquestionansweringsamplereweightingconflict-awarerewardsknowledge-awarelargelanguagemodelsmedicalQA
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

Supervised fine-tuning of large language models for domain-specific multiple-choice QA usually treats every training sample equally. This paper argues that equality is the wrong default: samples whose labeled answer conflicts with what the base model already knows can hurt the final model, while samples near the boundary of its knowledge are the most useful. KaFT scores each sample by how often the frozen base model picks the correct answer under shuffled options and repeated sampling, splits the data into four conflict-level groups, and trains with down-weighted losses on the two most conflicted groups. Across four open models and several medical QA benchmarks, it reports consistent gains over standard fine-tuning and over filtering conflict data outright, together with better out-of-domain generalization and less hallucination. The paper's point is that conflict-aware weighting is a cheap, sample-level alternative to data filtering or answer self-alignment.

What carries the argument

The load-bearing object is the conflict score in Eq. (1): for each training question, KaFT reorders the multiple-choice options and samples responses from the frozen base model at temperature 0.7, and the score is the fraction of those sampled responses that match the labeled answer. That score defines a quartile partition of the training set, and the reweighted objective in Eq. (2) multiplies each sample's loss by a reward that is 1 for the two low-conflict quartiles, 0.5 for the 'might-wrong' quartile, and 0.1 for the 'wrong' quartile. The reweighting is what lets the model keep learning from high-conflict examples without letting them dominate, and the option shuffling plus response sampling is what makes the score more stable than a single greedy in-context-learning probe.

What would settle it

Recompute the conflict score twice on the same training set with independent option shuffles and response samples; if the two score lists are near random orderings of each other (e.g., Spearman $\rho < 0.5$) or if KaFT trained on the second ordering no longer beats equal-weight fine-tuning, the central mechanism is not stable.

Watch

Extended reading notes

Core claim

The central claim is that the conflict between a model's internal knowledge and a training sample's labeled answer is a reliable guide to how that sample should be weighted during SFT. Using a query-diversified probe—ten shuffled option orders times ten temperature-0.7 samples from the frozen base model—KaFT computes a score for each question, sorts the training set into right, might-right, might-wrong, and wrong quartiles, and trains with rewards 1, 1, 0.5, and 0.1. Training on the wrong quartile alone produces large accuracy drops, yet mixing some of those samples back into the clean data helps, so the paper concludes that conflict samples should be down-weighted rather than deleted. Across Mistral-7B, Qwen1.5-7B, LLaMA3-3B, and LLaMA3-8B, KaFT reports average gains up to +2.40 percentage points over vanilla SFT and +5.73 points over the base model on a six-benchmark medical QA suite, and similar benefits on out-of-domain QA and hallucination evaluation.

Load-bearing premise

The load-bearing premise is that the probe of the frozen model's answers on shuffled options reliably tells which training samples will help and which will hurt after fine-tuning.

Editorial extensions

If this is right

  • If KaFT's claim is right, SFT pipelines can reduce the harm of knowledge conflict without discarding data: a one-time probing pass over the training set is enough to assign per-sample weights.
  • The finding that 'might-right' samples often outperform 'right' samples implies that the most useful supervised data sits at the boundary of what the base model already knows, not far inside it.
  • Down-weighting conflict data should transfer beyond medical QA, since the paper reports the same recipe helps on history, engineering, and law QA.
  • Because the method lowers overfitting to conflicting labels, it should reduce hallucination on held-out tasks, which the paper supports with HaluEval results.
  • Finer-grained conflict partitions (8 subsets instead of 4) improve results further, so the discrete reward grid is not the limit of the approach.

Reading between the lines

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

  • Editorial extension: because the conflict score is computed once from the frozen base model, one could recompute it after partial fine-tuning and reweight in a second round, testing whether the benefit comes from capturing the pretraining knowledge boundary.
  • Editorial extension: the hand-set rewards 0.1/0.5/1/1 and the choice of four quartiles are discrete simplifications; a smooth monotone mapping from score to weight, fit on a validation set, would likely remove the need to tune $\alpha$ and $\beta$ by hand.
  • Editorial extension: since the probing step needs only the base model and the training set, KaFT-style scores could be used as a data-selection filter before any SFT, composing naturally with curriculum or deduplication pipelines.
  • Editorial extension: the 'might-right beats right' pattern suggests a general curriculum principle, so the same score might transfer to instruction tuning or other multiple-choice settings, although the paper does not test those.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper addresses knowledge conflict in domain-specific supervised fine-tuning (SFT) of LLMs. It proposes a query diversification strategy that permutes answer options and samples multiple responses from the frozen base model to compute a per-sample conflict score (Eq. 1), splits the training data into four conflict-level subsets, and then fine-tunes with sample-adaptive rewards (Eq. 2), assigning small weights to the most conflicting subsets and full weight to moderate- and low-conflict subsets. The method, KaFT, is evaluated on medical QA benchmarks across four LLMs, with additional experiments on out-of-domain QA, multilingual medical QA, hallucination detection, and other domains. The main claimed results are consistent gains over vanilla SFT and the No-conflict and Self-aligning baselines.

Significance. If the central claim holds, the paper offers a simple and practical way to improve domain-specific QA fine-tuning by reweighting training samples according to a measured conflict level. The strengths of the paper are the breadth of the evaluation (four LLMs, six medical QA benchmarks, OOD generalization, hallucination evaluation, and multilingual evaluation), the use of publicly available benchmarks and the lm-evaluation-harness toolkit, and the honest limitation statement about model scale. The contribution, however, rests on two load-bearing assumptions: that Eq. (1) measures knowledge conflict rather than general difficulty or model uncertainty, and that the reward hyperparameters, which are tuned on the evaluation benchmarks themselves, are not responsible for the reported gains. The paper is therefore potentially useful but not yet fully convincing.

major comments (4)
  1. [§2.2, Eq. (1)] The load-bearing conflict score is the fraction of correct answers over No×Nr samples generated from the frozen base model. This measure conflates knowledge conflict with question difficulty and with pretraining unfamiliarity: under the definitions cited in §1 (Ren et al. 2024; Xu et al. 2024), a model that simply does not know the answer is not in conflict with the training context. Eq. (1) therefore cannot distinguish 'the model holds a specific competing answer' from 'the model is guessing' or 'the question is hard.' Because the reward assignment in Eq. (2) and all downstream gains in Table 2 depend on this score, the paper should provide evidence that low-Score samples exhibit a consistent alternative answer rather than random or uncertain behavior, and should compare KaFT against a difficulty-based reweighting baseline (e.g., down-weighting samples with low base-model accuracy without option permutation). Without such evidence, the 'knowledge-aware' mechanism is not established.
  2. [§4.2, Fig. 4, Table 2] The default rewards α=0.1 and β=0.5 are selected from the grid in Fig. 4 and Table 13 as the configuration that maximizes average performance on the same six medical QA benchmarks used for the main results in Table 2. This makes the comparison against Vanilla SFT and the other baselines partly self-referential, because the baselines do not receive equivalent hyperparameter tuning on the evaluation benchmarks. The authors should either tune α, β, and the number of subsets on a held-out validation split, or report a sensitivity analysis showing that the relative ordering over baselines is stable across a range of settings and random seeds. In addition, the main gains in Table 2 (e.g., +2.40 for Mistral-7B, +1.06 for Qwen1.5-7B) are reported without error bars or significance tests; for the smaller gains, the 'consistent and significant' claim is currently unsupported.
  3. [§2.3, Table 9] The text states that 'might-right performs better among all LLMs,' but Table 9 shows the opposite for LLaMA3-3B: the right subset averages 46.98 while might-right averages 46.05. Since the reward design in Eq. (2) is motivated by the claim that might-right is the most valuable subset, this contradiction needs to be resolved. If the pattern is not consistent across backbones, the motivation for assigning weight 1 to might-right and the broader claim of a universal trend are weakened.
  4. [Fig. 2(c), Table 10] The row labels in Table 10 appear to be swapped for Mistral-7B and LLaMA3-3B: the 0% column matches the No-conflict averages in Table 2 (Mistral-7B: 48.22; LLaMA3-3B: 46.98), so what is labeled 'Mistral-7B' is actually LLaMA3-3B, and vice versa. After correcting the labels, introducing 25% wrong data improves the average by only +0.17 (Mistral-7B) and +0.57 (LLaMA3-3B), which is much weaker than the text suggests and also within the range of sampling noise given that Figure 2(c) uses only three seeds. The authors should correct the table and temper the claim that 'appropriately applying conflict data is more beneficial than filtering.'
minor comments (4)
  1. [Table 6, Table 7] There are typographical and formatting issues: 'Sumarization' in Table 6 should be 'Summarization', and Table 7 contains 'inbold' as a formatting artifact.
  2. [Eq. (2)] The objective is written as E_{(q,o,a,R)∼D}[R log M(a|q,o)], but R is a deterministic function of the reward assignment rather than part of the data distribution; writing the reward outside the expectation or as R_i · log M(a_i|q_i,o_i) would be clearer and would avoid confusion about what is being sampled.
  3. [Appendix A.3] The probing prompt uses three randomly selected few-shot examples, but §2.2 does not state how these examples are chosen or whether the reported Score_i averages over different few-shot selections; since a stated motivation is to reduce sensitivity to few-shot examples, this detail should be specified.
  4. [§4.1] The paper says evaluation uses temperature 0 for reproducibility, while conflict detection uses temperature 0.7; this is fine, but the two decoding settings should be stated together in the setup so readers do not confuse the detection and evaluation protocols.

Circularity Check

1 steps flagged · score 3.0 of 10

Headline gains partly selected: α/β and bin count are tuned on the same six medical-QA benchmarks later reported as KaFT's improvements; the core empirical chain is otherwise self-contained.

  1. fitted input called prediction [Section 3 (Eq. 2), Section 4.3 'Parameter Analysis' and Table 5]
    "α and β are rewards between 0 and 1, where α is generally smaller than β. Empirically, we set α and β as 0.1 and 0.5, respectively. ... More specifically, the case of α = 0.1 and β = 0.5 performs best, thus leaving as our default experimental settings. ... We find that splitting the dataset into more subsets generally results in better performance ... we split the dataset into 4 subsets for a better trade-off."

    The default rewards in Eq. (2) and the number of conflict bins are chosen by maximizing average accuracy on the same six medical-QA benchmarks (MedQA, MedMCQA, MMLU*, CMB, CMExam, CMMLU*) that later appear as the headline claim ('KaFT brings consistent and significant performance gains'). Section 4.3 sets α=0.1, β=0.5 because that grid point 'performs best' on that suite, and Table 5 justifies the 4-subsets split by 'better performance' on the same suite after 'manually tuning the rewards'. The reported gains versus vanilla SFT (+2.40/+1.06/+1.34/+1.08) are thus evaluated at the configuration selected to maximize the very metric being reported — the headline magnitude is selected, not freely predicted.

full rationale

Derivation-chain walk. Eq. (1) computes Score_i from the frozen base model's correct-answer rate over No×Nr=100 permuted-option samples at temperature 0.7, probing only the training set. Eq. (2) maps quartile bins of Score_i to rewards (0.1/0.5/1/1) and reweights the SFT loss; the final models are evaluated on held-out test sets never used to compute Score_i. Each link is empirically tested rather than assumed: the differential harm of the 'wrong' subset is measured by training on each subset alone (Fig. 2b, Table 9); the binned reward beats both constant reward (= vanilla SFT) and score-proportional reward (Fig. 3, Table 12); and gains persist on unfitted benchmarks (OOD Table 3, MMedBench Table 14, HaluEval Table 6). Nothing in Eq. (1)'s definition guarantees that down-weighting low-Score samples raises held-out accuracy; that outcome is observed, so the central claim does not reduce to its inputs by construction. Self-citation is negligible and non-load-bearing: the only self-reference (Zhong et al. 2024, Related Works) merely lists an adaptive-learning example; no uniqueness theorem, prior ansatz, or defining equation is imported from the authors' own work. The one genuine circular element is parameter selection on the evaluation suite: α, β and the bin count are chosen because they 'perform best' / give 'better performance' on the six medical-QA benchmarks that constitute the headline result (Section 4.3, Table 5), making the reported gain magnitudes partly a selection artifact. This is a mild fitted-input-called-prediction loop rather than a forced reduction, because the improvement direction is robust across the whole α/β grid and on unfitted benchmarks. Two correctness risks (not circularity) are flagged: (a) Section 2.2 equates 'always fails to output the correct answer' with 'does not learn the knowledge', conflating difficulty or absent knowledge with the cited definition of conflict (Xu et al. 2024), a construct-validity threat to the 'knowledge-aware' interpretation; (b) Appendix Table 10's Mistral-7B and LLaMA3-3B rows appear swapped (their 0%/100% averages equal the other backbone's No-conflict/Vanilla-SFT numbers in Table 2), which after correction reduces the '25% wrong data is more beneficial' evidence (finding ❷) to +0.17/+0.57 — still positive, but weaker than Figure 2(c) suggests. The paper's own Limitations (up to 8B, MCQ-only) are honestly stated and do not bear on circularity.

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

KaFT introduces no new theoretical entities, but its effectiveness rests on two tuned hyperparameters (alpha, beta) and on the domain assumption that the diversified ICL probe reliably measures knowledge conflict that predicts SFT behavior. The probe is heuristic and its validity is only indirectly supported by downstream performance.

free parameters (4)
  • alpha (reward for wrong subset) = 0.1
    Tuned on medical QA test benchmarks in Section 4.3; chosen because alpha=0.1, beta=0.5 gave the best average score.
  • beta (reward for might-wrong subset) = 0.5
    Tuned on medical QA test benchmarks in Section 4.3; chosen because alpha=0.1, beta=0.5 gave the best average score.
  • Number of conflict-level subsets = 4
    Manual choice in Section 3; Section 4.3 shows 8 subsets give slightly better results but require more manual tuning.
  • No, Nr, temperature = 10, 10, 0.7
    Set for conflict probing in Section 2.2; no sensitivity analysis reported.
assumptions (5)
  • domain assumption The query diversification conflict score (Eq. 1) accurately reflects the degree of knowledge conflict for each training sample.
    Used to split data into reward groups; no ground-truth labels or calibration against independent conflict measures.
  • domain assumption The model's knowledge at probing time predicts the effect of the sample during SFT.
    The model changes during training, yet the conflict score is computed once on the frozen base model (Section 2.2).
  • domain assumption Moderate-conflict samples (might-right) are more useful than already-known samples (right).
    Empirical claim in Section 2.3, but contradicted for LLaMA3-3B in Table 9.
  • domain assumption Linearly down-weighting the loss for conflict data is an effective way to reduce its negative impact.
    Eq. 2 applies rewards as multipliers on log-likelihood; no theoretical or empirical justification for linear weighting.
  • domain assumption ICL with three random few-shot examples provides a stable probe of internal knowledge.
    Random selection of few-shot examples in Appendix A.3; diversification mitigates but does not eliminate instability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KaFT: Knowledge-aware Fine-tuning for Boosting LLMs' Domain-specific Question-Answering Performance." pith.science (2026). https://pith.science/paper/UOF4TODX

@misc{pith2026250515480,
  author       = {Pith},
  title        = {Pith review of: KaFT: Knowledge-aware Fine-tuning for Boosting LLMs' Domain-specific Question-Answering Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOF4TODX}},
  note         = {Machine review of arXiv:2505.15480}
}
read the original abstract

Supervised fine-tuning (SFT) is a common approach to improve the domain-specific question-answering (QA) performance of large language models (LLMs). However, recent literature reveals that due to the conflicts between LLMs' internal knowledge and the context knowledge of training data, vanilla SFT using the full QA training set is usually suboptimal. In this paper, we first design a query diversification strategy for robust conflict detection and then conduct a series of experiments to analyze the impact of knowledge conflict. We find that 1) training samples with varied conflicts contribute differently, where SFT on the data with large conflicts leads to catastrophic performance drops; 2) compared to directly filtering out the conflict data, appropriately applying the conflict data would be more beneficial. Motivated by this, we propose a simple-yet-effective Knowledge-aware Fine-tuning (namely KaFT) approach to effectively boost LLMs' performance. The core of KaFT is to adapt the training weight by assigning different rewards for different training samples according to conflict level. Extensive experiments show that KaFT brings consistent and significant improvements across four LLMs. More analyses prove that KaFT effectively improves the model generalization and alleviates the hallucination.

Figures

Figures reproduced from arXiv: 2505.15480 by the authors.

Figure 1
Figure 1. Comparison between (a) vanilla SFT and (b) our KaFT. Different from vanilla SFT treating all train￾ing data equally, KaFT uses sample-adaptive rewards to facilitate more effective learning of LLMs. desired behavioral norms, is usually required (Zhou et al., 2024; Zhang et al., 2024). However, recent literature (Ren et al., 2024; Gekhman et al., 2024) reveals that domain-specific SFT usually suffers from a crucial pr… view at source ↗
Figure 2
Figure 2. (a) Illustration of distributions of Scorei on MedQA across different LLMs. We use the kernel density estimate for visualizing, where the larger density refers to more training samples. (b) Performance comparison (%) of different subsets. Note that all subsets hold the same number of training samples. (c) Analysis of different proportions of wrong data. Specifically, we randomly select varied samples from wrong and … view at source ↗
Figure 3
Figure 3. Effect of reward strategies in KaFT. The y-axis denotes the average performance of medical QA. and “-w. auto-adapt” outperform the “-w. constant” by a clear margin, proving the effectiveness of knowledge-aware SFT. Moreover, “-w. auto-adapt” usually performs worse than ours. One possible reason is that it assigns a relatively small reward for the more important might-right subset, thus hindering the activation of LL… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Parameter analyses of KaFT. The y-axis and x-axis denote the varied α and β, respectively. We report the average results on medical QA benchmarks. Specifically, we split the training data into 2, 4 and 8 subsets based on the conflict scores, respec￾tively. After manual…
Figure 5
Figure 5. Figure 5: Performance comparison (%) on multilin￾gual medical QA. LLaMA3-8B is used as base model. average performance gains, especially +6.25% gains in Japanese and +7.81% gains in Russian. Hallucination Alleviation. As stated in the prior work (Gekhman et al., 2024), fine-tuni…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 25 canonical work pages

  1. [1]

    David Paul Ausubel, Joseph Donald Novak, Helen Hanesian, et al. 1978. Educational psychology: A cognitive view

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. In Advances in neural information processing systems

  4. [4]

    Zeming Chen, Alejandro Hern \'a ndez Cano, Angelika Romanou, Antoine Bonnet, Kyle Matoba, Francesco Salvi, Matteo Pagliardini, Simin Fan, Andreas K \"o pf, Amirkeivan Mohtashami, et al. 2023. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079

  5. [5]

    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 preprint arXiv:2407.21783

  6. [6]

    Zorik Gekhman, Gal Yona, Roee Aharoni, Matan Eyal, Amir Feder, Roi Reichart, and Jonathan Herzig. 2024. Does fine-tuning llms on new knowledge encourage hallucinations? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

  7. [7]

    Amirata Ghorbani and James Zou. 2019. Data shapley: Equitable valuation of data for machine learning. In International conference on machine learning. PMLR

  8. [8]

    Zhouhong Gu, Xiaoxuan Zhu, Haoning Ye, Lin Zhang, Jianchen Wang, Yixin Zhu, Sihang Jiang, Zhuozhi Xiong, Zihan Li, Weijie Wu, et al. 2024. Xiezhi: An ever-updating benchmark for holistic domain knowledge evaluation. In Proceedings of the AAAI Conference on Artificial Intelligence

Show all 41 references
  1. [9]

    Kai He, Rui Mao, Qika Lin, Yucheng Ruan, Xiang Lan, Mengling Feng, and Erik Cambria. 2025. A survey of large language models for healthcare: from data, technology, and applications to accountability and ethics. Information Fusion

  2. [10]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. In International Conference on Learning Representations

  3. [11]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2021. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations

  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

  5. [13]

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences

  6. [14]

    Minki Kang, Moonsu Han, and Sung Ju Hwang. 2020. Neural mask generator: Learning to generate adaptive word maskings for language model adaptation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

  7. [15]

    Yanis Labrak, Adrien Bazoge, Emmanuel Morin, Pierre-Antoine Gourraud, Mickael Rouvier, and Richard Dufour. 2024. Biomistral: A collection of open-source pretrained large language models for medical domains. In Findings of the Association for Computational Linguistics: ACL 2024

  8. [16]

    Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. 2020. Boosting few-shot learning with adaptive margin loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

  9. [17]

    Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2024. Cmmlu: Measuring massive multitask language understanding in chinese. In Findings of the Association for Computational Linguistics: ACL 2024

  10. [18]

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023 a . Halueval: A large-scale hallucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

  11. [19]

    Yunxiang Li, Zihan Li, Kai Zhang, Ruilong Dan, Steve Jiang, and You Zhang. 2023 b . Chatdoctor: A medical chat model fine-tuned on a large language model meta-ai (llama) using medical domain knowledge. Cureus

  12. [20]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024 a . Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

  13. [21]

    Junling Liu, Peilin Zhou, Yining Hua, Dading Chong, Zhongyu Tian, Andrew Liu, Helin Wang, Chenyu You, Zhenhua Guo, Lei Zhu, et al. 2024 b . Benchmarking large language models on cmexam-a comprehensive chinese medical exam dataset. In Advances in Neural Information Processing Systems

  14. [22]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

  15. [23]

    OpenAI. 2023. Gpt-4 technical report

  16. [24]

    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. In Advances in neural information processing systems

  17. [25]

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

  18. [26]

    Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2024. Towards building multilingual language model for medicine. Nature Communications

  19. [27]

    Mengjie Ren, Boxi Cao, Hongyu Lin, Cao Liu, Xianpei Han, Ke Zeng, Guanglu Wan, Xunliang Cai, and Le Sun. 2024. Learning or self-aligning? rethinking instruction fine-tuning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

  20. [28]

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. 2023. Large language models encode clinical knowledge. Nature

  21. [29]

    Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, et al. 2025. Toward expert-level medical question answering with large language models. Nature Medicine

  22. [30]

    Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. 2024 a . Openchat: Advancing open-source language models with mixed-quality data. In The Twelfth International Conference on Learning Representations

  23. [31]

    Xidong Wang, Guiming Chen, Song Dingjie, Zhang Zhiyi, Zhihong Chen, Qingying Xiao, Junying Chen, Feng Jiang, Jianquan Li, Xiang Wan, et al. 2024 b . Cmb: A comprehensive medical benchmark in chinese. In Proceedings of the 2024 Conference of the North American Chapter of the As...

  24. [32]

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

  25. [33]

    Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. 2024. Knowledge conflicts for llms: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

  26. [34]

    Junjie Ye, Yuming Yang, Qi Zhang, Tao Gui, Xuanjing Huang, Peng Wang, Zhongchao Shi, and Jianping Fan. 2024. Empirical insights on fine-tuning large language models for question-answering. arXiv preprint arXiv:2409.15825

  27. [35]

    Jinsung Yoon, Sercan Arik, and Tomas Pfister. 2020. Data valuation using reinforcement learning. In International Conference on Machine Learning. PMLR

  28. [36]

    Liang Zhang, Katherine Jijo, Spurthi Setty, Eden Chung, Fatima Javid, Natan Vidra, and Tommy Clifford. 2024. Enhancing large language model performance to answer questions and extract information more accurately. arXiv preprint arXiv:2402.01722

  29. [37]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223

  30. [38]

    Qihuang Zhong, Liang Ding, Li Shen, Juhua Liu, Bo Du, and Dacheng Tao. 2024. Revisiting knowledge distillation for autoregressive language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

  31. [39]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2024. Lima: Less is more for alignment. In Advances in Neural Information Processing Systems

  32. [40]

    URL: " 'urlintro :=

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

  33. [41]

    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.