Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Quantization of LLM weights to 4 bits cuts GPU memory by three-quarters while preserving performance on biomedical tasks, making 70B-parameter models deployable on a 40GB GPU.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Quantizing LLMs to 4 or 8 bits cuts GPU memory by up to 75% with generally small performance changes across eight biomedical NLP benchmarks.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Valuable benchmark of quantized LLMs for biomedical NLP, but the headline 70B-on-40GB claim rests on a memory metric that looks like weight footprint rather than real peak memory requirements. the 4 major comments →

arxiv 2509.04534 v1 pith:2XAHPIR6 submitted 2025-09-04 cs.CL cs.AI

Quantized Large Language Models in Biomedical Natural Language Processing: Evaluation and Recommendation

Zaifu Zhan , Shuang Zhou , Min Zeng , Kai Yu , Meijia Song , Xiaoyi Chen , Jun Wang , Yu Hou
show 1 more author
Rui Zhang
This is my paper
classification cs.CL cs.AI
keywords quantized LLMsbiomedical NLP4-bit quantizationGPU memory reductionlocal LLM deploymentnamed entity recognitionquestion answeringprivacy-preserving AI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 weight quantization is a practical lever for bringing large language models into healthcare settings that must keep data on-site. Across 12 models and eight biomedical benchmarks covering named entity recognition, relation extraction, multi-label classification, and question answering, the authors report that 8-bit and 4-bit weight formats cut GPU memory roughly in half and to about a quarter, respectively, while task scores stay close to full-precision levels. If true, a 70-billion-parameter model becomes runnable on a single 40GB GPU, and domain-specific biomedical models keep their specialized knowledge after compression. The paper's real target is the privacy constraint: local deployment without cloud dependence, which quantization can make affordable.

Core claim

The central claim is that the precision of a model's weights, not its architecture, is the main lever that decides whether a strong biomedical LLM can run locally. Using 8-bit and 4-bit weight quantization, the benchmark shows peak memory falling from roughly 130GB to about 40GB for 70B-class models, with performance on most datasets within a few points of full precision and sometimes slightly better. The paper also reports that biomedical models such as ClinicalCamel-70B retain their domain knowledge, that larger quantized models still scale in performance, and that chain-of-thought and self-consistency prompting remain effective on quantized weights. The authors conclude that quantization

What carries the argument

The load-bearing object is weight quantization at two bit widths: W8A16, where weights are stored in 8-bit and activations in 16-bit, and W4A16, where weights use a 4-bit normal-float format. The weight tensor is compressed before inference, removing most of the parameter memory, and values are reconstructed on the fly during the forward pass so accuracy is largely preserved. This is what makes the reported memory reduction from over 100GB to about 40GB for 70B-parameter models possible, and it explains the observed latency increase as the price of reconstructing weights during generation.

Load-bearing premise

The deployment claim assumes the paper's measured peak memory covers the real runtime allocation, including activations, KV cache, and library overhead, and that 40GB is the target GPU.

What would settle it

Run a 70B model at 4-bit weight precision on a 40GB GPU over the paper's full benchmark, including few-shot and long-output tasks, logging true peak GPU memory; an out-of-memory failure or a required memory above 40GB would refute the headline deployment claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Hospitals and labs with a single 40GB GPU can run 70B-class LLMs locally for named entity recognition, relation extraction, multi-label classification, and question answering without sending patient text to the cloud.
  • 4-bit quantization shrinks memory to about a quarter of full precision, so several large specialized models can be hosted side by side on one server.
  • Domain-tuned biomedical models keep their medical knowledge after quantization, so local deployment does not mean falling back to generic models.
  • Advanced prompting methods such as chain-of-thought and self-consistency still work on quantized models, preserving modern accuracy gains.
  • The practical trade-off is speed: quantized models respond more slowly, so real-time deployments must plan for latency rather than accuracy loss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next test is to check whether 14B-32B models at 4-bit fit the far more common 24GB consumer GPUs, which would widen the privacy-preserving deployment base beyond 40GB cards.
  • The study's pattern of 8-bit sometimes hurting more than 4-bit, as seen with HuatuoGPT-o1-70B and Llama3-Med42-70B, invites task-specific bit-width selection, since the cheapest memory savings are not always the most stable.
  • Quantization's memory savings could be coupled with retrieval-augmented generation, spending the freed GPU memory on long-context or embedding workloads for private biomedical question answering.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper reports a systematic evaluation of 4-bit and 8-bit weight quantization (W4A16, W8A16) for 12 large language models—both general-purpose and biomedical-domain models—across eight biomedical NLP datasets covering multi-label classification, named entity recognition, relation extraction, and question answering. The main empirical claims are that quantization preserves task performance to a negligible degree of loss, reduces GPU memory usage by up to 75%, enables deployment of 70B-parameter models on a 40GB GPU, and is compatible with few-shot learning and advanced prompting strategies. The authors also provide practical recommendations for deploying quantized LLMs in biomedical settings.

Significance. If the central claims hold, the paper would be a useful practical reference for biomedical NLP practitioners who need local, privacy-preserving deployment. Its strengths are breadth—12 models, 8 datasets, 4 tasks—and the public release of code and benchmark data, which supports reproducibility. However, the two headline claims—'negligible performance degradation' and '70B models on 40GB consumer GPUs'—rest on measurement choices and statistical evidence that are not adequately documented. The memory metric in particular appears, from the numbers in Table 1, to reflect model weight footprint rather than true peak inference memory; if that is the case, the deployment claim is not supported. The recommendations in Figure 7 are a useful synthesis, but they inherit the validity of the underlying measurements.

major comments (4)
  1. [Section 2.1, Table 1, Section 4.5] The paper's central deployment claim rests on the 'Mem' column of Table 1, but the manuscript never states how this quantity was measured. The reported values are suspiciously close to weight-only footprints: Llama-3.3-70B is 132.37 GB at full precision and 39.38 GB at W4A16, which are approximately the FP16 and 4-bit weight sizes, respectively; the values vary by less than 1 GB across datasets with very different output lengths. If 'Mem' was obtained with model.get_memory_footprint() or a similar weight-summing routine, it omits activations, KV cache, CUDA context, and BitsAndBytes runtime buffers, and a W4 70B model would not fit on a 40GB A100. The authors should specify the exact measurement API (e.g., torch.cuda.max_memory_allocated()) and, if the current metric is not peak runtime memory, re-measure and revise the abstract and discussion claims accordingly. In addition, a 40GB A100
  2. [Section 2.1, Table 1] The blanket statement that quantized models exhibit 'negligible performance degradation' is not supported by the table. Several entries show substantial drops: Llama-3.3-70B on MedQA drops from 0.732 (full precision) to 0.555 at W8A16; Llama-3.3-70B on DDI2013 drops from 0.524 to 0.450 at W8A16; Deepseek-llm-65B on PubMedQA drops from 0.166 to 0.118 at W4A16; Qwen2.5-72B on NCBI-disease drops from 0.413 to 0.357 at W4A16. The paper should quantify the fraction of model/task/bit-width combinations with large degradations and either qualify the 'negligible' claim or identify the conditions under which degradation is non-negligible.
  3. [Section 2.1, Figure 2, Table 1] Table 1 reports single runs with no repeated trials, confidence intervals, or standard deviations, even though the performance differences between quantization levels are often small relative to the expected run-to-run variability of LLM decoding. Figure 2 shows error bars labeled 'standard deviation,' but the number of repetitions and the variance estimation procedure are not described, and the same information is absent from Table 1. Without uncertainty quantification, the conclusion that quantization 'preserves' performance is not statistically grounded. Adding a small number of repeats, or at least reporting the variance from the runs that produced Figure 2, would materially strengthen the paper.
  4. [Section 4.3, Section 4.5] The quantization setup is described at the configuration level, but the use of BitsAndBytes with 'full-precision computation during inference' is not sufficient to explain the reported latency increases and memory figures. The paper acknowledges that latency measurements are noisy due to server load (Section 3), but the memory measurements are treated as exact. The authors should document the hardware, batch size, the exact quantization API call, and whether the reported latency is per-example averaged over the entire dataset or measured on a subset. This is necessary for readers to reproduce the efficiency numbers and to assess the '70B on 40GB' claim.
minor comments (5)
  1. [Abstract, Section 3] The abstract and discussion use 'consumer-grade GPUs' to describe 40GB A100s. This is inaccurate; consumer GPUs typically have 8–24GB. Please revise to 'a single 40GB GPU' or similar.
  2. [Abstract, Section 2.1] The abstract claims memory reduction 'by up to 75%,' but the largest reductions in Table 1 are about 70% (e.g., 132.37->39.38 GB = 70.2%). Please verify the maximum reduction across all experiments and correct the number if needed.
  3. [Figure 2 caption] The caption says 'error bars indicate standard deviation' but does not state the number of repeated runs. Please add this information.
  4. [Section 2.2] The text notes that HuatuoGPT-o1-70B and Llama3-Med42-70B show 'substantial performance degradation' at 8 bits, but no scores are given in the text. Since this is a caveat to the main claim, please include the relevant numbers or refer precisely to Figure 3.
  5. [Section 4.4] The exact prompts used for the instruction, chain-of-thought, and self-consistency conditions are not provided. Please include them in an appendix or repository so that the prompting experiments are reproducible.

Circularity Check

0 steps flagged

No significant circularity; this is an empirical benchmark with independent measurements.

full rationale

This paper is an empirical evaluation study. It contains no theoretical derivation, no fitted parameters renamed as predictions, no uniqueness theorem, and no ansatz smuggled in via citation. The central claims—quantization reduces GPU memory requirements and preserves task performance—are supported by measured values in Table 1 and Figures 2–6, obtained by running standard BitsAndBytes NF4/int8 quantization on public biomedical datasets. The reported memory numbers (e.g., Llama-3.3-70B at W4A16: 39.38 GB) are empirical outputs, not quantities defined to equal the conclusion that a 70B model can run on a 40GB GPU. Even if one questions whether the 'Mem' column reflects true peak inference memory rather than weight footprint, that is a measurement-validity or reproducibility concern, not circularity. The reuse of datasets and prompts from reference 2 (Chen et al., Nature Communications 2025) is a benchmark-design choice; the evaluation outputs are newly measured and are externally checkable against public datasets, so it is not load-bearing self-citation. The paper's own stated limitations (Section 3: latency affected by server load; Section 4.5: memory measurement methodology not fully described) also raise correctness/reproducibility concerns rather than circular steps. No equation or claim reduces by construction to its own input.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper contains no equations and fits no parameters, so the axiom ledger consists of measurement assumptions: that single-run scores are stable, that the BitsAndBytes quantization configuration represents quantization broadly, and that reported peak memory equals real deployment memory. These assumptions are load-bearing for the headline practical recommendations.

axioms (3)
  • domain assumption Single forward-pass scores on a fixed prompt template are representative of a model's performance on biomedical NLP tasks.
    Table 1 reports point estimates without repeated runs or confidence intervals, so the 'preserved performance' conclusion assumes these single measurements are stable and not noise.
  • domain assumption BitsAndBytes 4-bit NF4 and 8-bit quantization with full-precision compute is a faithful proxy for quantization in general.
    The recommendations generalize from one quantization library and format (Section 4.3) to 'quantization' as a strategy; other schemes such as GPTQ or AWQ may behave differently.
  • domain assumption Peak GPU memory figures fully capture local deployment requirements.
    Deployability on a 40GB GPU is inferred from the Mem columns in Table 1 without separately accounting for activations, KV cache, or CUDA context overhead during inference.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantized Large Language Models in Biomedical Natural Language Processing: Evaluation and Recommendation." pith.science (2026). https://pith.science/paper/2XAHPIR6

@misc{pith2026250904534,
  author       = {Pith},
  title        = {Pith review of: Quantized Large Language Models in Biomedical Natural Language Processing: Evaluation and Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XAHPIR6}},
  note         = {Machine review of arXiv:2509.04534}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models have demonstrated remarkable capabilities in biomedical natural language processing, yet their rapid growth in size and computational requirements present a major barrier to adoption in healthcare settings where data privacy precludes cloud deployment and resources are limited. In this study, we systematically evaluated the impact of quantization on 12 state-of-the-art large language models, including both general-purpose and biomedical-specific models, across eight benchmark datasets covering four key tasks: named entity recognition, relation extraction, multi-label classification, and question answering. We show that quantization substantially reduces GPU memory requirements-by up to 75%-while preserving model performance across diverse tasks, enabling the deployment of 70B-parameter models on 40GB consumer-grade GPUs. In addition, domain-specific knowledge and responsiveness to advanced prompting methods are largely maintained. These findings provide significant practical and guiding value, highlighting quantization as a practical and effective strategy for enabling the secure, local deployment of large yet high-capacity language models in biomedical contexts, bridging the gap between technical advances in AI and real-world clinical translation.

Figures

Figures reproduced from arXiv: 2509.04534 by Jun Wang, Kai Yu, Meijia Song, Min Zeng, Rui Zhang, Shuang Zhou, Xiaoyi Chen, Yu Hou, Zaifu Zhan.

Figure 1
Figure 1. Figure 1: Comparison between default LLMs and their quantized counterparts. Additionally, Li et al.31 explored the reasoning capabilities of quantized LLMs in mathematical tasks, while Marchisio et al.32 investigated quantization strategies for multimodal models. Further studies have examined the combination of quantization with low-rank adaptation techniques such as LoRA for fine-tuning, as seen in works by Xu et a… view at source ↗
Figure 2
Figure 2. Figure 2: Quantitative comparison of different quantization schemes and LLMs. Bars indicate mean, error bars indicate standard deviation. 2.2 Domain quantized LLMs on MLC, NER, RE, QA tasks An increasing number of models have been further trained with domain-specific knowledge, and many large models now inherently contain rich domain expertise. Therefore, it is essential to explore the effectiveness of quantization … view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different quantized LLMs from the biomedical domain on DDI2013, HoC, MedQA, and NCBI Disease datasets across MLC, NER, RE, and QA tasks [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of various quantized Qwen models on DDI2013, HoC, MedQA, and NCBI Disease datasets across MLC, NER, RE, and QA tasks. 5/11 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of quantized Qwen2.5-70B model with few-shot setting on DDI2013, HoC, MedQA, and NCBI Disease datasets across MLC, NER, RE, and QA tasks. 6/11 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparison of the quantized Qwen2.5-70B model using instruction prompts, chain-of-thought prompts, and self-consistent prompting methods on the HoC and MedQA datasets. 3 Discussion In this study, we conducted a systematic evaluation of quantization techniques applied to large language models (LLMs). As LLMs continue to scale rapidly, quantization has become an essential strategy for enabling th… view at source ↗
Figure 7
Figure 7. Figure 7: Recommendations for using quantized LLMs in biomedical natural language processing applications. models continue to exhibit strong generalization, instruction-following, and reasoning capabilities. Despite a slight drop in performance in some cases, the trade-off is minimal and acceptable, especially given the drastic improvements in memory efficiency. This makes it possible to deliver nearly the same user… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs

    cs.AI 2026-07 conditional novelty 5.0

    Quantized LLMs diverge from their base models at the decision level even when accuracy is preserved, with query and key attention projections showing the greatest structural distortion under low-bit compression.

Reference graph

Works this paper leans on

56 extracted references · 30 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    Zhou, S. et al. Large language models for disease diagnosis: A scoping review. npj Artif. Intell. 1, 9 (2025)

    work page 2025

  2. [2]

    Chen, Q. et al. Benchmarking large language models for biomedical natural language processing applications and recommendations. Nat. communications 16, 3280 (2025)

    work page 2025

  3. [3]

    Zhan, Z. et al. An evaluation of deepseek models in biomedical natural language processing. arXiv preprint arXiv:2503.00624 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Zhou, S. et al. Automating expert-level medical reasoning evaluation of large language models. arXiv preprint arXiv:2507.07988 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

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

    Pith/arXiv arXiv 2023

  6. [6]

    Team, G. et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Liu, A. et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    Pith/arXiv arXiv 2024

  8. [8]

    Luo, R. et al. Biogpt: generative pre-trained transformer for biomedical text generation and mining.Briefings bioinformatics 23, bbac409 (2022)

    work page 2022

  9. [9]

    Tu, T. et al. Towards generalist biomedical ai. Nejm Ai 1, AIoa2300138 (2024)

    work page 2024

  10. [10]

    Sellergren, A. et al. Medgemma technical report. arXiv preprint arXiv:2507.05201 (2025). 9/11

    Pith/arXiv arXiv 2025

  11. [11]

    & Zhang, R

    Zhan, Z., Zhou, S., Li, M. & Zhang, R. Ramie: retrieval-augmented multi-task information extraction with large language models on dietary supplements. J. Am. Med. Informatics Assoc. 32, 545–554 (2025)

    work page 2025

  12. [12]

    Brown, T. et al. Language models are few-shot learners. Adv. neural information processing systems 33, 1877–1901 (2020)

    1901

  13. [13]

    Zaghir, J. et al. Prompt engineering paradigms for medical applications: scoping review. J. Med. Internet Res. 26, e60501 (2024)

    work page 2024

  14. [14]

    Wang, J. et al. Data-efficient biomedical in-context learning: A diversity-enhanced submodular perspective. arXiv preprint arXiv:2508.08140 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Lewis, P.et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Adv. neural information processing systems 33, 9459–9474 (2020)

    work page 2020

  16. [16]

    Zhan, Z. et al. Retrieval-augmented in-context learning for multimodal large language models in disease classification. arXiv preprint arXiv:2505.02087 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    & Zhang, R

    Zhan, Z., Wang, J., Zhou, S., Deng, J. & Zhang, R. Mmrag: multi-mode retrieval-augmented generation with large language models for biomedical in-context learning. J. Am. Med. Informatics Assoc. ocaf128, DOI: 10.1093/jamia/ocaf128 (2025). https://academic.oup.com/jamia/advance-article-pdf/doi/10.1093/jamia/ocaf128/63948576/ocaf128.pdf

    work page doi:10.1093/jamia/ocaf128 2025

  18. [18]

    Kaplan, J. et al. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    Pith/arXiv arXiv 2001

  19. [19]

    & Zhang, R

    Zhan, Z. & Zhang, R. Towards better multi-task learning: A framework for optimizing dataset combinations in large language models. In Findings of the Association for Computational Linguistics: NAACL 2025, 5373–5386 (2025)

    work page 2025

  20. [20]

    EPEE: Towards Efficient and Effective Foundation Models in Biomedicine

    Zhan, Z., Zhou, S., Zhou, H., Liu, Z. & Zhang, R. Epee: Towards efficient and effective foundation models in biomedicine. arXiv preprint arXiv:2503.02053 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Act, A. et al. Health insurance portability and accountability act of 1996. Public law 104, 191 (1996)

    work page 1996

  22. [22]

    Price, W. N. & Cohen, I. G. Privacy in the age of medical big data. Nat. medicine 25, 37–43 (2019)

    work page 2019

  23. [23]

    Mehrtak, M. et al. Security challenges and solutions using healthcare cloud computing. J. medicine life 14, 448 (2021)

    work page 2021

  24. [24]

    Gholami, A. et al. A survey of quantization methods for efficient neural network inference. In Low-power computer vision, 291–326 (Chapman and Hall/CRC, 2022)

    work page 2022

  25. [25]

    & Xie, Y

    Deng, L., Li, G., Han, S., Shi, L. & Xie, Y . Model compression and hardware acceleration for neural networks: A comprehensive survey. Proc. IEEE 108, 485–532 (2020)

    work page 2020

  26. [26]

    Li, S. et al. Evaluating quantized large language models. arXiv preprint arXiv:2402.18158 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    Jin, R. et al. A comprehensive evaluation of quantization strategies for large language models. In Ku, L.-W., Martins, A. & Srikumar, V . (eds.) Findings of the Association for Computational Linguistics: ACL 2024 , 12186–12215, DOI: 10.18653/v1/2024.findings-acl.726 (Association for Computational Linguistics, Bangkok, Thailand, 2024)

    doi:10.18653/v1/2024.findings-acl.726 2024

  28. [28]

    Husom, E. J. et al. Sustainable llm inference for edge ai: Evaluating quantized llms for energy efficiency, output accuracy, and inference latency. arXiv preprint arXiv:2504.03360 (2025)

    Pith/arXiv arXiv 2025

  29. [29]

    Girija, S. S. et al. Optimizing llms for resource-constrained environments: A survey of model compression techniques. arXiv preprint arXiv:2505.02309 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Evaluating the generalization ability of quantized llms: Benchmark, analysis, and toolbox

    Liu, Y .et al. Evaluating the generalization ability of quantized llms: Benchmark, analysis, and toolbox. arXiv preprint arXiv:2406.12928 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Li, Z. et al. Quantization meets reasoning: Exploring llm low-bit quantization degradation for mathematical reasoning. arXiv preprint arXiv:2501.03035 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Marchisio, K. et al. How does quantization affect multilingual llms? arXiv preprint arXiv:2407.03211 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Qa-lora: Quantization-aware low-rank adaptation of large language models

    Xu, Y .et al. Qa-lora: Quantization-aware low-rank adaptation of large language models. arXiv preprint arXiv:2309.14717 (2023)

    Pith/arXiv arXiv 2023

  34. [34]

    Zhou, S. et al. Uncertainty-aware large language models for explainable disease diagnosis.arXiv preprint arXiv:2505.03467 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Baker, S. et al. Automatic semantic classification of scientific literature according to the hallmarks of cancer.Bioinformatics 32, 432–440 (2016)

    work page 2016

  36. [36]

    Chen, Q. et al. Multi-label classification for biomedical literature: an overview of the biocreative vii litcovid track for covid-19 literature topic annotations. Database 2022, baac069 (2022). 10/11

    work page 2022

  37. [37]

    Li, J. et al. Biocreative v cdr task corpus: a resource for chemical disease relation extraction. Database 2016 (2016)

    work page 2016

  38. [38]

    I., Leaman, R

    Do˘gan, R. I., Leaman, R. & Lu, Z. Ncbi disease corpus: a resource for disease name recognition and concept normalization. J. biomedical informatics 47, 1–10 (2014)

    work page 2014

  39. [39]

    Krallinger, M. et al. Overview of the biocreative vi chemical-protein interaction track. In Proceedings of the sixth BioCreative challenge evaluation workshop, vol. 1, 141–146 (2017)

    work page 2017

  40. [40]

    & Herrero-Zazo, M

    Segura-Bedmar, I., Martínez, P. & Herrero-Zazo, M. Semeval-2013 task 9: Extraction of drug-drug interactions from biomedical texts (ddiextraction 2013). In Second Joint Conference on Lexical and Computational Semantics (* SEM), Volume 2: Proceedings of the Seventh International Workshop on Semantic Evaluation (SemEval 2013), 341–350 (2013)

    work page 2013

  41. [41]

    Jin, D. et al. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Appl. Sci. 11, 6421 (2021)

    work page 2021

  42. [42]

    Jin, Q., Dhingra, B., Liu, Z., Cohen, W. & Lu, X. PubMedQA: A dataset for biomedical research question answering. In Inui, K., Jiang, J., Ng, V . & Wan, X. (eds.)Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2567–2577, DOI: 10...

    doi:10.18653/v1/d19-1259 2019

  43. [43]

    Dubey, A. et al. The llama 3 herd of models. arXiv e-prints arXiv–2407 (2024)

    work page 2024

  44. [44]

    Qwen2.5 technical report (2025)

    Qwen et al. Qwen2.5 technical report (2025). 2412.15115

    Pith/arXiv arXiv 2025

  45. [45]

    Abdin, M. et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905 (2024)

    Pith/arXiv arXiv 2024

  46. [46]

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

    Pith/arXiv arXiv 2024

  47. [47]

    Toma, A. et al. Clinical camel: An open expert-level medical language model with dialogue-based knowledge encoding. arXiv preprint arXiv:2305.12031 (2023)

    Pith/arXiv arXiv 2023

  48. [48]

    K., Raha, T., Khan, S

    Christophe, C., Kanithi, P. K., Raha, T., Khan, S. & Pimentel, M. A. Med42-v2: A suite of clinical llms. arXiv preprint arXiv:2408.06142 (2024)

    Pith/arXiv arXiv 2024

  49. [49]

    Chen, J. et al. Huatuogpt-o1, towards medical complex reasoning with llms. arXiv preprint arXiv:2412.18925 (2024)

    Pith/arXiv arXiv 2024

  50. [50]

    Chen, Z. et al. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079 (2023)

    Pith/arXiv arXiv 2023

  51. [51]

    Huggingface’s transformers: State-of-the-art natural language processing

    Wolf, T.et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771 (2019)

    Pith/arXiv arXiv 1910

  52. [52]

    & Zettlemoyer, L

    Dettmers, T., Lewis, M., Shleifer, S. & Zettlemoyer, L. 8-bit optimizers via block-wise quantization. arXiv preprint arXiv:2110.02861 (2021)

    Pith/arXiv arXiv 2021

  53. [53]

    & Zettlemoyer, L

    Dettmers, T., Lewis, M., Belkada, Y . & Zettlemoyer, L. Llm. int8 () 8-bit matrix multiplication for transformers at scale. In Proceedings of the 36th International Conference on Neural Information Processing Systems, 30318–30332 (2022)

    work page 2022

  54. [54]

    Wei, J. et al. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021)

    Pith/arXiv arXiv 2021

  55. [55]

    Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Adv. neural information processing systems 35, 24824–24837 (2022)

    work page 2022

  56. [56]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171 (2022)

    Wang, X.et al. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171 (2022). 11/11

    Pith/arXiv arXiv 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.