Pith. sign in

REVIEW 5 major objections 4 minor 52 references

PU-ADKA claims that a fixed $100 consultation budget, spent by selectively asking the right human expert at the right time, improves LLM performance in specialized domains without fine-tuning.

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 ↗

A budget-aware framework (PU-ADKA) selects which domain expert an LLM should query under a fixed $100 budget, improving specialized-domain answers at low cost.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Unreadable text makes this impossible to review; the abstract frames a real problem, but the core claims are unverifiable. the 5 major comments →

arxiv 2508.17202 v2 pith:2B2LM4MK submitted 2025-08-24 cs.CL

Active Domain Knowledge Acquisition with 100-Dollar Budget: Enhancing LLMs via Cost-Efficient, Expert-Involved Interaction in Sensitive Domains

Yang Wu , Raha Moraffah , Rujing Yao , Jinhong Yu , Zhimin Tao , Xiaozhong Liu This is my paper
classification cs.CL
keywords LLM domain specializationactive knowledge acquisitionexpert-in-the-loopbudget-constrained consultationdrug discoveryrare disease researchCKAD benchmarksimulation-based training
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 a general large language model (LLM) can be turned into a specialist in cost-sensitive fields like drug discovery and rare-disease research by spending a small, fixed budget consulting human experts—no fine-tuning required. The authors propose PU-ADKA, a framework that learns which questions to ask, which expert to ask, and what that consultation is worth, balancing availability, expertise, and price. They train the framework in simulations built from PubMed literature and then test it with controlled human-expert interactions and a real drug-development team. If the claim holds, scarce expert time can be spent where it most improves the model, and specialized LLM knowledge can be acquired under strict budget limits.

Core claim

The central claim is that domain expertise can be treated as a purchasable, budgeted resource. PU-ADKA considers a team of experts with different availability, knowledge boundaries, and consultation costs; for each question it decides whether the base LLM can answer alone or whether it should pay a specific expert, and it does this through a policy learned in simulation from PubMed data. The paper reports that the same policy, when run with real human experts and in a live drug-development deployment, raises LLM performance on specialized tasks under a strict $100-style budget. Alongside the method, the paper releases CKAD, a benchmark dataset for measuring cost-effective domain knowledge ac

What carries the argument

PU-ADKA, a cost-aware expert-selection policy. It converts "ask an expert for help" into a sequential budget-allocation problem: for each question it weighs an expert's availability, the boundaries of what that expert knows, and the price of consulting them. Simulation on PubMed data supplies the practice environment in which the policy learns, so that expensive, scarce real experts are only called upon when the learned policy expects the highest marginal knowledge gain.

Load-bearing premise

The load-bearing premise is that the simulated experts built from PubMed behave enough like real experts—same availability, cost, and error patterns—that the query-selection policy trained in simulation still makes good choices with real people.

What would settle it

Re-run the drug-development deployment with a fresh panel of experts and compare PU-ADKA's final answers, under the same $100 budget, against two controls: random expert selection and always consulting the cheapest available expert. If PU-ADKA does not beat both on expert-scored accuracy, the claimed budget-efficient improvement is falsified.

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

If this is right

  • Specializing an LLM no longer requires assembling a large fine-tuning corpus; a targeted expert-consultation budget can substitute.
  • Expert time is allocated by expected knowledge gain rather than by fixed protocol, so scarce specialists are not spent on questions the model already answers well.
  • The policy learned in simulation can be applied in a real expert setting in the same domain, as demonstrated in the drug-development deployment.
  • CKAD gives future work a common, budget-aware benchmark for comparing domain knowledge acquisition methods.

Where Pith is reading between the lines

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

  • The same budget-allocation machinery could extend to any mix of expensive knowledge sources—paid APIs, specialized models, or human experts—each with its own price and reliability.
  • A natural safety-oriented extension would reserve part of the budget for verification, hiring a second expert to check high-stakes answers instead of spending every dollar acquiring new facts.
  • Because training depends on simulated experts, a direct transfer test—deploying the PubMed-trained policy in a non-biomedical domain with a different expert vocabulary—would measure how much of the gain comes from simulation fidelity versus the selection mechanism.
  • The budget framing invites a head-to-head comparison with retrieval-based augmentation: if the same $100 buys high-quality database or API access rather than human hours, which route yields more domain accuracy?
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

5 major / 4 minor

Summary. This paper proposes PU-ADKA, a framework for active domain knowledge acquisition under a fixed budget. Instead of fine-tuning, it selects which domain expert to consult based on availability, knowledge boundaries, and consultation cost. The authors state that the selection policy is trained on simulations built from PubMed data, validated in controlled expert interactions, and deployed with a real drug-development team; they also introduce a CKAD benchmark. The abstract, however, provides no numeric results, baselines, or protocol details, and the supplied full text is almost entirely mojibake, with a running header from another arXiv paper. Consequently, the methods and experimental evidence cannot be inspected.

Significance. If the claimed results held, the framework would address a practical problem: improving specialized LLMs without expensive fine-tuning while respecting expert availability and cost. The introduction of a benchmark for cost-effective domain knowledge acquisition could also be useful. On the submitted evidence, however, these contributions exist only as statements in the abstract. No machine-checked derivations, reproducible code, or inspectable experimental protocols are available. The significance cannot be assessed until a readable manuscript with baselines and transfer analysis is provided.

major comments (5)
  1. [Full text] The manuscript body is unreadable: the text is mojibake, tables are gibberish, and a header 'arXiv:2508.17199v1 [cs.CV]' appears mid-document. No equation, algorithm, or result can be verified. This presentation defect blocks review; the paper effectively consists only of its abstract.
  2. [Abstract] The abstract states that PU-ADKA is validated through controlled expert interactions and real-world deployment, but it reports no performance metric, baseline, error bar, or statistical comparison. Without these, the claim that the framework 'demonstrat[es] its effectiveness' is unsupported in the submitted artifact. This concern is load-bearing because the entire contribution is empirical.
  3. [Abstract (simulation-to-real)] The policy is 'trained ... using simulations on PubMed data' and then deployed with a drug-development team. The abstract provides no evidence that the simulated experts have realistic availability, cost, and error patterns, nor any analysis of simulator-to-real transfer. If the simulation is faithful only to PubMed text statistics, the learned query policy may not transfer to human experts, and the claimed $100-budget improvement would not follow.
  4. [Abstract (deployment)] The real-world deployment, as described, appears to lack a control arm. To attribute gains to the PU-ADKA selection policy rather than to the act of consulting experts at all, the deployment must compare against a heuristic selector (e.g., random, cheapest-available, or fixed order) under the same budget. No such comparison is visible in the abstract or readable text.
  5. [Abstract (evaluation data)] The abstract does not specify whether the evaluation questions in the controlled interactions and deployment overlap with the PubMed corpus used to train the simulated experts. If the same literature underlies both the simulated training signal and the evaluation, part of the reported gain could be due to rediscovering the training signal rather than to effective expert selection. This leakage risk must be addressed in the experimental design.
minor comments (4)
  1. [Full text] The running header from arXiv:2508.17199 must be removed; the source should be rebuilt with a working font/encoding so the PDF is legible.
  2. [Abstract] The '$100-dollar budget' is never defined in the abstract or readable portion. Please specify the cost model: what counts as a consultation, how expert time is priced, and how the ceiling is applied.
  3. [Abstract (CKAD)] The CKAD benchmark is introduced by name, but no task description, data source, size, or evaluation protocol is given in the readable portion.
  4. [References] The reference list is not accessible due to the encoding corruption, so related-work attribution cannot be checked.

Circularity Check

0 steps flagged

No specific circularity identifiable; the derivation chain is not inspectable and the abstract alone does not reduce to its inputs.

full rationale

The readable portion of the paper is limited to the abstract; the supplied full text is garbled mojibake and includes a running header from a different arXiv paper (arXiv:2508.17199, cs.CV), so no equations, algorithm definitions, fit targets, or evaluation protocols can be quoted to exhibit a specific reduction. The abstract's claims—training PU-ADKA on PubMed simulations, validating via controlled expert interactions and a real drug-development deployment, and introducing the CKAD benchmark—are transfer/evaluation claims rather than self-definitional derivations. Nothing in the abstract defines the method's target quantity in terms of its output or fits a parameter and then renames that fit as a prediction. The concern that simulation-trained selection may not transfer to real experts, or that CKAD may share source data with the training simulations, is a validity and evaluation-design concern, not a demonstrated circular step; without quotable evidence that CKAD is constructed from the same PubMed data used for training, or that the reported gains are forced by the construction, the rules require a non-finding. No load-bearing self-citations, imported uniqueness theorems, or smuggled ansätze are visible in the readable text. Accordingly, the honest circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The ledger is abstract-only because the full text is corrupted. The visible free parameters are the quantities the framework must model (cost, availability, knowledge boundaries, budget) whose numeric values are not disclosed. The axioms are domain assumptions about LLM ignorance, simulation fidelity, usefulness of consultation, and PubMed as a knowledge proxy. The invented entities are the framework itself and the benchmark, neither of which currently carries independent, inspectable evidence.

free parameters (4)
  • Expert consultation cost model = unknown (abstract only)
    The framework accounts for 'each expert's consultation costs'; the numeric cost schedule used in the simulation and in the real deployment is not stated in the abstract.
  • Expert availability distribution = unknown (abstract only)
    The selection policy must know expert availability; the abstract does not say how availability probabilities were estimated or set in the PubMed-based simulation.
  • Expert knowledge boundary estimates = unknown (abstract only)
    The policy picks the 'most appropriate expert' using predicted knowledge boundaries; how these boundaries were derived from PubMed data is a fitted component not described in the abstract.
  • Budget ceiling ($100) = $100 (framing)
    The '100-dollar budget' is the headline constraint and is effectively a design choice of the authors; whether it was tuned to make the method look favorable or imposed externally is unclear.
axioms (4)
  • domain assumption LLMs lack sufficient expert knowledge in specialized, sensitive domains such as drug discovery and rare disease research
    Motivates the entire framework (abstract, first sentence). If general models already handle such queries well, expert acquisition adds no value. This is not proven in the abstract.
  • domain assumption Simulated experts trained on PubMed data faithfully represent real expert behavior, including knowledge boundaries, availability, and cost
    Bridges training and validation: the learned selection policy can only transfer to real experts if the simulation is a faithful stand-in. This is the weakest_assumption of the report and is unverifiable from the abstract.
  • domain assumption Expert consultation under a strict budget measurably improves LLM outputs in the target domains
    The validation design presupposes that the acquired knowledge is useful and that the chosen evaluation metric captures that utility; neither metric nor magnitude is given in the abstract.
  • domain assumption PubMed text is an adequate proxy for expert knowledge in the target sensitive domains
    The simulation is built on PubMed data; if the literature does not encode tacit expertise such as failed trials or clinical heuristics, the simulated experts will teach the policy the wrong query behavior.
invented entities (2)
  • PU-ADKA framework no independent evidence
    purpose: Manager policy that selects which expert to query under a fixed budget, based on availability, cost, and knowledge boundaries
    The framework is a new artifact introduced by the paper. The abstract provides no external falsifiable handle beyond announced evaluations that cannot be inspected due to full-text corruption.
  • CKAD benchmark dataset no independent evidence
    purpose: Benchmark for cost-effective LLM domain knowledge acquisition to support future research
    Announced in the abstract with no download link, schema, statistics, or evaluation harness. Until released, it offers no independent handle outside this paper.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Domain Knowledge Acquisition with 100-Dollar Budget: Enhancing LLMs via Cost-Efficient, Expert-Involved Interaction in Sensitive Domains." pith.science (2026). https://pith.science/paper/2B2LM4MK

@misc{pith2026250817202,
  author       = {Pith},
  title        = {Pith review of: Active Domain Knowledge Acquisition with 100-Dollar Budget: Enhancing LLMs via Cost-Efficient, Expert-Involved Interaction in Sensitive Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2B2LM4MK}},
  note         = {Machine review of arXiv:2508.17202}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs) have demonstrated an impressive level of general knowledge. However, they often struggle in highly specialized and cost-sensitive domains such as drug discovery and rare disease research due to the lack of expert knowledge. In this paper, we propose a novel framework (PU-ADKA) designed to efficiently enhance domain-specific LLMs by actively engaging domain experts within a fixed budget. Unlike traditional fine-tuning approaches, PU-ADKA selectively identifies and queries the most appropriate expert from a team, taking into account each expert's availability, knowledge boundaries, and consultation costs. We train PU-ADKA using simulations on PubMed data and validate it through both controlled expert interactions and real-world deployment with a drug development team, demonstrating its effectiveness in enhancing LLM performance in specialized domains under strict budget constraints. In addition to outlining our methodological innovations and experimental results, we introduce a new benchmark dataset, CKAD, for cost-effective LLM domain knowledge acquisition to foster further research in this challenging area.

discussion (0)

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

Reference graph

Works this paper leans on

52 extracted references · 26 canonical work pages · 2 internal anchors

  1. [1]

    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.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

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

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  4. [4]

    Elliot Bolton, Abhinav Venigalla, Michihiro Yasunaga, David Hall, Betty Xiong, Tony Lee, Roxana Daneshjou, Jonathan Frankle, Percy Liang, Michael Carbin, et al. 2024. Biomedlm: A 2.7 b parameter language model trained on biomedical text. arXiv preprint arXiv:2403.18421

    Pith/arXiv arXiv 2024

  5. [5]

    Shayok Chakraborty, Vineeth Balasubramanian, Qian Sun, Sethuraman Panchanathan, and Jieping Ye. 2015. Active batch selection via convex relaxations with guaranteed solution bounds. IEEE transactions on pattern analysis and machine intelligence, 37(10):1945--1958

    work page 2015

  6. [6]

    Gui Citovsky, Giulia DeSalvo, Claudio Gentile, Lazaros Karydas, Anand Rajagopalan, Afshin Rostamizadeh, and Sanjiv Kumar. 2021. Batch active learning at scale. Advances in Neural Information Processing Systems, 34:11933--11944

    work page 2021

  7. [7]

    Clarivate . 2025. https://mjl.clarivate.com/home Master journal list . Accessed: 2025-01-02

    work page 2025

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171--4186

    2019

  9. [9]

    UU Dhar. 2024. Enabling domain expert evaluation of emerging ai technologies in healthcare settings

    work page 2024

  10. [10]

    Marthinus Du Plessis, Gang Niu, and Masashi Sugiyama. 2015. Convex formulation for learning from positive and unlabeled data. In International conference on machine learning, pages 1386--1394. PMLR

    work page 2015

  11. [12]

    Yann Dubois, Bal \'a zs Galambosi, Percy Liang, and Tatsunori B Hashimoto. 2024 b . Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475

    Pith/arXiv arXiv 2024

  12. [13]

    Yarin Gal, Riashat Islam, and Zoubin Ghahramani. 2017. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183--1192. PMLR

    work page 2017

  13. [14]

    Ruijiang Gao and Maytal Saar-Tsechansky. 2020. Cost-accuracy aware adaptive labeling for active learning. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 2569--2576

    work page 2020

  14. [15]

    Preni Golazizian, Alireza S Ziabari, Ali Omrani, and Morteza Dehghani. 2024. Cost-efficient subjective task annotation and modeling through few-shot annotator adaptation. arXiv preprint arXiv:2402.14101

    Pith/arXiv arXiv 2024

  15. [16]

    Suchin Gururangan, Ana Marasovi \'c , Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964

    Pith/arXiv arXiv 2020

  16. [17]

    Guy Hacohen, Avihu Dekel, and Daphna Weinshall. 2022. Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794

    Pith/arXiv arXiv 2022

  17. [18]

    o llenbrok, and Beg \

    Julia Henkel, Genc Hoxha, Gencer Sumbul, Lars M \"o llenbrok, and Beg \"u m Demir. 2023. Annotation cost efficient active learning for content based image retrieval. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 4994--4997. IEEE

    work page 2023

  18. [19]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    Pith/arXiv arXiv 2021

  19. [20]

    Sheng-Jun Huang, Jia-Lve Chen, Xin Mu, and Zhi-Hua Zhou. 2017. Cost-effective active learning from diverse labelers. In IJCAI, pages 1879--1885

    work page 2017

  20. [21]

    Jia Ji, Yongshuai Hou, Xinyu Chen, Youcheng Pan, and Yang Xiang. 2024. Vision-language model for generating textual descriptions from clinical images: model development and validation study. JMIR Formative Research, 8:e32690

    work page 2024

  21. [22]

    Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke H \"u llermeier. 2023. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925

    arXiv 2023

  22. [23]

    Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang, and Il-Chul Moon. 2021. Lada: Look-ahead data acquisition via augmentation for deep active learning. Advances in Neural Information Processing Systems, 34:22919--22930

    work page 2021

  23. [24]

    Ryuichi Kiryo, Gang Niu, Marthinus C Du Plessis, and Masashi Sugiyama. 2017. Positive-unlabeled learning with non-negative risk estimator. Advances in neural information processing systems, 30

    work page 2017

  24. [25]

    Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2023 a . From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning. arXiv preprint arXiv:2308.12032

    Pith/arXiv arXiv 2023

  25. [26]

    Shibo Li, Jeff M Phillips, Xin Yu, Robert Kirby, and Shandian Zhe. 2022. Batch multi-fidelity active learning with budget constraints. Advances in Neural Information Processing Systems, 35:995--1007

    work page 2022

  26. [27]

    Yunshui Li, Binyuan Hui, Xiaobo Xia, Jiaxi Yang, Min Yang, Lei Zhang, Shuzheng Si, Junhao Liu, Tongliang Liu, Fei Huang, et al. 2023 b . One shot learning as instruction data prospector for large language models. arXiv preprint arXiv:2312.10302

    Pith/arXiv arXiv 2023

  27. [28]

    Wei Liu, Weihao Zeng, Keqing He, Yong Jiang, and Junxian He. 2023. What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning. arXiv preprint arXiv:2312.15685

    Pith/arXiv arXiv 2023

  28. [29]

    Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. 2022. Biogpt: generative pre-trained transformer for biomedical text generation and mining. Briefings in bioinformatics, 23(6):bbac409

    work page 2022

  29. [30]

    Chaitanya Malaviya, Subin Lee, Sihao Chen, Elizabeth Sieber, Mark Yatskar, and Dan Roth. 2023. Expertqa: Expert-curated questions and attributed answers. arXiv preprint arXiv:2309.07852

    Pith/arXiv arXiv 2023

  30. [31]

    Sanket Mangrulkar, Kaustubh Somasundaram, and Akhilesh Shrivastava. 2022. Peft: Parameter-efficient fine-tuning. Hugging Face. https://github.com/huggingface/peft

    work page 2022

  31. [32]

    Mary L McHugh. 2012. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276--282

    2012

  32. [33]

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. 2023. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435

    Pith/arXiv arXiv 2023

  33. [34]

    OpenAI . 2024. https://openai.com/index/hello-gpt-4o/ Gpt-4o model card

    work page 2024

  34. [35]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730--27744

    2022

  35. [36]

    Soumen Pal, Manojit Bhattacharya, Sang-Soo Lee, and Chiranjib Chakraborty. 2024. A domain-specific next-generation large language model (llm) or chatgpt is required for biomedical engineering and research. Annals of biomedical engineering, 52(3):451--454

    work page 2024

  36. [37]

    Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholamreza Reza Haffari, Anton Van Den Hengel, and Javen Qinfeng Shi. 2022. Active learning by feature mixing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12237--12246

    work page 2022

  37. [38]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2017. Automatic differentiation in pytorch. In Advances in Neural Information Processing Systems

    work page 2017

  38. [39]

    Sawan Patel, Fred Zhangzhi Peng, Keith Fraser, Adam D Friedman, Pranam Chatterjee, and Sherwood Yao. 2025. Evoflow-rna: Generating and representing non-coding rna with a language model. bioRxiv, pages 2025--02

    work page 2025

  39. [40]

    PubMed . 2024. PubMed Data: Download and Use . https://pubmed.ncbi.nlm.nih.gov/download. Last updated Dec 14, 2024

    work page 2024

  40. [41]

    Annalisa Szymanski, Noah Ziems, Heather A Eicher-Miller, Toby Jia-Jun Li, Meng Jiang, and Ronald A Metoyer. 2025. Limitations of the llm-as-a-judge approach for evaluating llm outputs in expert knowledge tasks. In Proceedings of the 30th International Conference on Intelligent User Interfaces, pages 952--966

    2025

  41. [42]

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

    Pith/arXiv arXiv 2023

  42. [43]

    Hongru Wang, Rui Wang, Fei Mi, Zezhong Wang, Ruifeng Xu, and Kam-Fai Wong. 2023. Chain-of-thought prompting for responding to in-depth dialogue questions with llm. arXiv preprint arXiv:2305.11792

    Pith/arXiv arXiv 2023

  43. [44]

    Jian Wang, Zhe Zhang, and Guosheng Zhao. 2024. Tag-based self-learning task recommendation for mobile crowdsensing via collaborative multi-expert system. Computer Communications, 214:260--269

    work page 2024

  44. [45]

    Jianhao Wang, Zhizhou Ren, Terry Liu, Yang Yu, and Chongjie Zhang. 2020. Qplex: Duplex dueling multi-agent q-learning. arXiv preprint arXiv:2008.01062

    Pith/arXiv arXiv 2020

  45. [46]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. https://doi.org/10.18653/v1...

    doi:10.18653/v1/2020.emnlp-demos.6 2020

  46. [47]

    Yang Wu, Xurui Li, Xuhong Zhang, Yangyang Kang, Changlong Sun, and Xiaozhong Liu. 2023. Community-based hierarchical positive-unlabeled (pu) model fusion for chronic disease prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 2747--2756

    work page 2023

  47. [48]

    Yang Wu, Chenghao Wang, Ece Gumusel, and Xiaozhong Liu. 2024 a . Knowledge-infused legal wisdom: Navigating llm consultation through the lens of diagnostics and positive-unlabeled reinforcement learning. arXiv preprint arXiv:2406.03600

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [49]

    Yang Wu, Huayi Zhang, Yizheng Jiao, Lin Ma, Xiaozhong Liu, Jinhong Yu, Dongyu Zhang, Dezhi Yu, and Wei Xu. 2024 b . Rose: A reward-oriented data selection framework for llm task-specific instruction tuning. arXiv preprint arXiv:2412.00631

    Pith/arXiv arXiv 2024

  49. [50]

    Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen. 2024. Less: Selecting influential data for targeted instruction tuning. arXiv preprint arXiv:2402.04333

    Pith/arXiv arXiv 2024

  50. [51]

    Rujing Yao, Yang Wu, Chenghao Wang, Jingwei Xiong, Fang Wang, and Xiaozhong Liu. 2025 a . Elevating legal llm responses: Harnessing trainable logical structures and semantic knowledge with legal reasoning. arXiv preprint arXiv:2502.07912

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [52]

    Rujing Yao, Yiquan Wu, Tong Zhang, Xuhui Zhang, Yuting Huang, Yang Wu, Jiayin Yang, Changlong Sun, Fang Wang, and Xiaozhong Liu. 2025 b . Intelligent legal assistant: An interactive clarification system for legal question answering. In Companion Proceedings of the ACM on Web Conference 2025, pages 2935--2938

    work page 2025

  52. [53]

    Ruoyu Zhang, Yanzeng Li, Yongliang Ma, Ming Zhou, and Lei Zou. 2023. Llmaaa: Making large language models as active annotators. arXiv preprint arXiv:2310.19596

    Pith/arXiv arXiv 2023

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