Pith. sign in

REVIEW 4 major objections 2 minor 56 references

Are Today's LLMs Ready to Explain Well-Being Concepts?

T0 review · 4 major / 2 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fine-tuning an open-source LLM with supervised fine-tuning and direct preference optimization makes it explain well-being concepts better than larger models, with a dual-judge automatic evaluation that tracks human qua

desk verdict Plausible and potentially useful benchmark, but the body text I received is unreadable, so the headline empirical claims are currently unadjudicated. read the letter →

arxiv 2508.03990 v1 pith:2FL7XURI submitted 2025-08-06 cs.CL cs.AIcs.HC

Bohan Jiang , Dawei Li , Zhen Tan , Chengshuai Zhao , Huan Liu This is my paper · ORCID
classification cs.CLcs.AIcs.HC
keywords LLM-as-a-judgeexplanationqualitywell-beingconceptsdirectpreferenceoptimizationsupervisedfine-tuningaudienceadaptationopen-sourceLLMs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether large language models can explain well-being concepts—mental, physical, and social—in ways that are accurate and also matched to the reader's level of expertise. To answer it, the authors built a dataset of 43,880 explanations of 2,194 well-being concepts generated by ten different LLMs, and introduced a principle-guided LLM-as-a-judge evaluation framework with dual judges to score quality. They report that these judges align well with human evaluations. The central finding is that fine-tuning an open-source model with supervised fine-tuning (SFT) and direct preference optimization (DPO) significantly improves generated explanations, and that the fine-tuned smaller models outperform their larger counterparts. If the paper is right, teams can make small, open models explain well-being concepts better than much larger systems, using an automatic judge as a reliable yardstick.

What carries the argument

The dual-judge, principle-guided LLM-as-a-judge evaluation framework, combined with supervised fine-tuning (SFT) and direct preference optimization (DPO), is the mechanism that carries the argument. The framework supplies the quality signal; SFT and DPO turn that signal into a better generator.

What would settle it

Pick a set of well-being concepts the judges never saw, have humans rate a random sample of explanations blind, and compare those ratings with the judges' scores; if agreement with humans is at chance, or if the DPO-finetuned model's advantage disappears when humans, not the judges, do the rating, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that explanation quality for well-being concepts is measurable and improvable. Quality is measured by a principle-guided LLM-as-a-judge framework that uses two judges rather than one to assess whether an explanation is accurate, complete, and appropriate for a specified audience level such as expert or general reader. Against human ratings, the proposed judges agree closely enough to stand in for human evaluation at scale. Using those judgments to build preference pairs, the authors fine-tune an open-source LLM with SFT followed by DPO. The result they report is that this small, open model produces explanations that score higher than those of larger, more expensi

Load-bearing premise

The whole result rests on the assumption that the two LLM judges, once aligned with a human-rated sample, produce valid quality scores for all 43,880 explanations across every concept category and audience level; if the judges reward their own stylistic preferences rather than genuine quality, the reported fine-tuning gains may not reflect real improvement.

Editorial extensions

If this is right

  • Small, open-source models fine-tuned this way can be deployed to explain well-being concepts at a fraction of the cost of calling large proprietary APIs.
  • Automatic judges that align with human raters can scale quality evaluation to tens of thousands of explanations without manual review.
  • Preference-based learning (DPO) is effective for specialized explanation tasks, not just general chat alignment.
  • Explanation quality is audience- and concept-dependent, so a single model is unlikely to serve all readers equally; quality scores should be reported by audience and category.

Reading between the lines

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

  • Going beyond the paper, the same dual-judge recipe could be adapted to other domains where explanations must match reader expertise, such as health advice or financial literacy; the paper does not test this.
  • The paper's comparison is between a fine-tuned smaller model and its larger base counterparts; an extension would pit the fine-tuned model against similarly fine-tuned large models to see whether the size advantage disappears.
  • A stress test the paper does not report is how the fine-tuned model behaves on well-being concepts outside the 2,194-concept dataset, and how the judges score those out-of-distribution explanations.
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 / 2 minor

Summary. The paper introduces a large-scale dataset of 43,880 LLM-generated explanations for 2,194 well-being concepts and proposes a principle-guided dual LLM-as-a-judge evaluation framework. It reports three empirical claims: (1) the LLM judges align well with human evaluations; (2) explanation quality varies across models, audiences, and concept categories; and (3) SFT- and DPO-fine-tuned open-source models outperform larger base models. The provided full text, however, is unreadable due to severe character-encoding corruption, so only the abstract can be assessed. The abstract states these results without reporting any quantitative evidence such as agreement coefficients, human sample sizes, effect sizes, or significance values.

Significance. If the claims are substantiated, the paper would offer a reusable evaluation framework for explanation quality and evidence that preference-based fine-tuning can make small open models competitive with or superior to much larger models on a specialized explanation task. The scale of the dataset (2,194 concepts, 43,880 explanations from ten LLMs) is a potential community resource. The dual-judge framework with human alignment is a timely and useful contribution. The central caveat is that the entire evaluation pipeline rests on the validity of the LLM judges; the human-validation evidence is the only independent anchor, and that evidence is not visible in the supplied manuscript. The paper also makes a strong practical claim about fine-tuning gains, which must be checked against the same judge-quality concerns.

major comments (4)
  1. [Full text (all sections)] The supplied body of the manuscript is unreadable; nearly every section, table, and figure caption appears as corrupted character encoding. Consequently, no methodology, equation, table, or result can be verified. This is not a minor formatting issue: it prevents the referee from checking the dataset construction, the human-evaluation protocol, the judge rubric, the DPO preference construction, or the statistical analysis. The authors should resubmit a readable, correctly encoded version before the paper can receive substantive review.
  2. [Abstract, claim (1)] The claim that the proposed LLM judges 'align well with human evaluations' is made without reporting any agreement coefficient (e.g., Cohen's kappa, Spearman's rho, or percentage agreement), the number of human annotators, the annotation protocol, or the size of the human-rated sample. Since the human evaluation is the only independent check against judge circularity, the absence of this information leaves the core validity of the evaluation framework unsupported.
  3. [Abstract, claims (1) and (3)] There is a serious circularity risk: if the preference pairs used for DPO were generated by the same family of LLM judges used in the evaluation, then the reported fine-tuning gains may reflect optimization of those judges' stylistic preferences rather than improvement in human-valued explanation quality. The manuscript must state the provenance of the preference pairs and, if they are judge-generated, demonstrate that the preference signal was validated against human judgments. The supplied text does not allow this concern to be resolved, so the strongest result in the paper remains ungrounded.
  4. [Abstract, claim (3)] The claim that SFT/DPO fine-tuned models 'significantly enhance' explanation quality and 'outperform their larger counterparts' is reported without effect sizes, confidence intervals, or significance tests. In a comparison of many models across many categories, such numerical support is essential to distinguish a genuine improvement from judge-specific noise. The abstract alone is insufficient, and the corrupted body prevents further verification.
minor comments (2)
  1. [Abstract] The abstract would be much more informative if it reported the key agreement coefficient and at least one representative effect size for the fine-tuning improvement. As written, the three headline results contain no numbers at all.
  2. [Full text] The corrupted characters appear to result from an encoding mismatch; the authors should ensure that the PDF and the source file compile cleanly. This point is secondary, but it is the immediate cause of the inability to review the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity demonstrable from the provided text; body is undecodable and abstract alone does not reduce claims to their inputs.

full rationale

The manuscript body as supplied is corrupted/undecodable mojibake; the only intact quotable text is the abstract. The abstract claims (1) a principle-guided LLM-as-a-judge framework that aligns with human evaluations, and (2) that SFT/DPO fine-tuning improves explanation quality as assessed by those judges. Nothing in the abstract states or implies that the DPO preference pairs were generated by the same judges used for evaluation, nor that the judge scores are definitionally identical to the training signals. The 'align well with human evaluations' assertion is an external-validity claim, not a self-referential definition. No equation, fitted parameter renamed as a prediction, or self-citation chain is visible. The reader's concern about judge-fitting is a plausible empirical risk, but without paper text confirming the same-judge preference source, flagging it as circularity would violate the hard rule requiring exhibit of a specific reduction. Accordingly, no circularity step can be identified, and the score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The evaluation loop is only as good as the human validation, which the abstract references but does not specify. No free parameter values are disclosed, and the dataset is the only new artifact; it has no independent external evidence.

free parameters (2)
  • LLM judge rubric and dual-judge aggregation thresholds
    The principle-guided dual-judge protocol requires prompt templates, scoring scales, and an agreement or aggregation rule between the two judges. The abstract discloses none of these; any calibration against the human sample is a fitted component of the evaluation instrument.
  • SFT/DPO training hyperparameters and preference pair selection
    Fine-tuning gains depend on data proportions, learning rates, and how preference pairs were constructed. These choices can determine whether the fine-tuned model 'outperforms' larger counterparts, and none are disclosed at abstract level.
assumptions (3)
  • domain assumption Well-being explanation quality is meaningfully captured by a small set of principles scored by LLM judges (e.g., accuracy, clarity, audience-appropriateness).
    The abstract introduces 'principle-guided LLM-as-a-judge evaluation' without listing the principles or showing they cover well-being explanation quality; if the principles omit key dimensions, every score inherits the omission. Location: Abstract, methodology sentence.
  • domain assumption Human ratings on the validation sample are ground truth, and judge alignment on that sample transfers to all 2,194 concepts and all audience levels.
    Claim (1), that 'judges align well with human evaluations', implies a human sample exists; generalization to the full corpus is assumed. Location: Abstract, results statement (1).
  • ad hoc to paper The DPO preference pairs, if LLM-generated, reflect true human preferences rather than judge idiosyncrasies.
    The abstract reports DPO gains without stating the preference source; if preferences come from the same judge family, the gain is partially circular. Location: Abstract, results statement (3).
invented entities (1)
  • Well-being explanation evaluation dataset (2,194 concepts, 43,880 explanations from ten LLMs)
    purpose: Benchmark for measuring LLM-generated well-being explanations across models, audiences, and concept categories
    The dataset is constructed by the authors for this paper and carries no external evidence of validity beyond the authors' own judge framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Today's LLMs Ready to Explain Well-Being Concepts?." pith.science (2026). https://pith.science/paper/2FL7XURI

@misc{pith2026250803990,
  author       = {Pith},
  title        = {Pith review of: Are Today's LLMs Ready to Explain Well-Being Concepts?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2FL7XURI}},
  note         = {Machine review of arXiv:2508.03990}
}
read the original abstract

Well-being encompasses mental, physical, and social dimensions essential to personal growth and informed life decisions. As individuals increasingly consult Large Language Models (LLMs) to understand well-being, a key challenge emerges: Can LLMs generate explanations that are not only accurate but also tailored to diverse audiences? High-quality explanations require both factual correctness and the ability to meet the expectations of users with varying expertise. In this work, we construct a large-scale dataset comprising 43,880 explanations of 2,194 well-being concepts, generated by ten diverse LLMs. We introduce a principle-guided LLM-as-a-judge evaluation framework, employing dual judges to assess explanation quality. Furthermore, we show that fine-tuning an open-source LLM using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) can significantly enhance the quality of generated explanations. Our results reveal: (1) The proposed LLM judges align well with human evaluations; (2) explanation quality varies significantly across models, audiences, and categories; and (3) DPO- and SFT-finetuned models outperform their larger counterparts, demonstrating the effectiveness of preference-based learning for specialized explanation tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 29 canonical work pages

  1. [1]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Achiam, J.; Held, D.; Tamar, A.; and Abbeel, P. 2017. Constrained policy optimization. In International conference on machine learning, 22--31. PMLR

    work page 2017

  3. [3]

    Alexandrova, A. 2017. A philosophy for the science of well-being. Oxford University Press

    work page 2017

  4. [4]

    Askell, A.; Bai, Y.; Chen, A.; Drain, D.; Ganguli, D.; Henighan, T.; Jones, A.; Joseph, N.; Mann, B.; DasSarma, N.; et al. 2021. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861

    arXiv 2021

  5. [5]

    Chae, H.; Kim, M.; Kim, C.; Jeong, W.; Kim, H.; Lee, J.; and Yeo, J. 2023. TUTORING: instruction-grounded conversational agent for language learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 16413--16415

    work page 2023

  6. [6]

    Chen, C.; and Shu, K. 2023. Can llm-generated misinformation be detected? arXiv preprint arXiv:2309.13788

    arXiv 2023

  7. [7]

    Chen, Y.; Zhang, X.; Wang, J.; Xie, X.; Yan, N.; Chen, H.; and Wang, L. 2024. Structured dialogue system for mental health: An llm chatbot leveraging the pm+ guidelines. In International Conference on Social Robotics, 262--271. Springer

    work page 2024

  8. [8]

    Cho, Y.; and Choi, I. 2018. Writing from sources: Does audience matter? Assessing Writing, 37: 25--38

    work page 2018

Show all 56 references
  1. [9]

    Cohen, J. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1): 37--46

    1960

  2. [10]

    Comanici, G.; Bieber, E.; Schaekermann, M.; Pasupat, I.; Sachdeva, N.; Dhillon, I.; Blistein, M.; Ram, O.; Zhang, D.; Rosen, E.; et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv ...

    2025 arXiv

  3. [11]

    Diener, E. 2000. Subjective well-being: The science of happiness and a proposal for a national index. American psychologist, 55(1): 34

    2000

  4. [12]

    Gao, W.; Liu, Q.; Yue, L.; Yao, F.; Lv, R.; Zhang, Z.; Wang, H.; and Huang, Z. 2025. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 23923--23932

    2025

  5. [13]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  6. [14]

    T.-Z.; Williams, J

    Grossman, J.; Lin, Z.; Sheng, H.; Wei, J. T.-Z.; Williams, J. J.; and Goel, S. 2019. MathBot: Transforming online resources for learning math into conversational interactions. AAAI 2019 Story-Enabled Intelligence

    2019

  7. [15]

    Gunjal, A.; Wang, A.; Lau, E.; Nath, V.; Liu, B.; and Hendryx, S. 2025. Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains. arXiv preprint arXiv:2507.17746

    2025 arXiv

  8. [16]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  9. [17]

    Hu, B.; Sheng, Q.; Cao, J.; Shi, Y.; Li, Y.; Wang, D.; and Qi, P. 2024. Bad actor, good advisor: Exploring the role of large language models in fake news detection. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 22105--22113

    2024

  10. [18]

    Jarden, A.; and Roache, A. 2023. What is wellbeing?

    2023

  11. [19]

    R.; and Liu, H

    Jiang, B.; Li, D.; Tan, Z.; Zhou, X.; Rao, A.; Lerman, K.; Bernard, H. R.; and Liu, H. 2024 a . Assessing the impact of conspiracy theories using large language models. arXiv preprint arXiv:2412.07019

    2024 arXiv

  12. [20]

    Jiang, B.; Tan, Z.; Nirmal, A.; and Liu, H. 2024 b . Disinformation detection: An evolving challenge in the age of llms. In Proceedings of the 2024 siam international conference on data mining (sdm), 427--435. SIAM

    2024

  13. [21]

    A.; Jung, H.; and Kim, Y.-H

    Jo, E.; Epstein, D. A.; Jung, H.; and Kim, Y.-H. 2023. Understanding the benefits and challenges of deploying conversational AI leveraging large language models for public health intervention. In Proceedings of the 2023 CHI conference on human factors in computing systems, 1--16

    2023

  14. [22]

    Keil, F. C. 2006. Explanation and understanding. Annu. Rev. Psychol., 57(1): 227--254

    2006

  15. [23]

    M.; and Seo, M

    Kim, S.; Bae, S.; Shin, J.; Kang, S.; Kwak, D.; Yoo, K. M.; and Seo, M. 2023. Aligning large language models through synthetic feedback. arXiv preprint arXiv:2305.13735

    2023 arXiv

  16. [24]

    Y.; Longpre, S.; Kim, C.; Yoon, D.; Son, G.; Cho, Y.; Shafayat, S.; Baek, J.; et al

    Kim, S.; Suk, J.; Cho, J. Y.; Longpre, S.; Kim, C.; Yoon, D.; Son, G.; Cho, Y.; Shafayat, S.; Baek, J.; et al. 2025. The BiGGen Bench: A Principled Benchmark for Fine-grained Evaluation of Language Models with Language Models. In Proceedings of the 2025 Conference of the Natio...

    2025

  17. [25]

    Li, D.; Jiang, B.; Huang, L.; Beigi, A.; Zhao, C.; Tan, Z.; Bhattacharjee, A.; Jiang, Y.; Chen, C.; Wu, T.; Shu, K.; Cheng, L.; and Liu, H. 2024. From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge. arXiv preprint arXiv: 2411.16594

    2024

  18. [26]

    Li, D.; Sun, R.; Huang, Y.; Zhong, M.; Jiang, B.; Han, J.; Zhang, X.; Wang, W.; and Liu, H. 2025. Preference leakage: A contamination problem in llm-as-a-judge. arXiv preprint arXiv:2502.01534

    2025

  19. [27]

    Lin, C.-Y. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 74--81

    2004

  20. [28]

    I.; Siddique, F

    Lin, Z.; Xu, P.; Winata, G. I.; Siddique, F. B.; Liu, Z.; Shin, J.; and Fung, P. 2020. Caire: An end-to-end empathetic chatbot. In Proceedings of the AAAI conference on artificial intelligence, volume 34, 13622--13623

    2020

  21. [29]

    Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  22. [30]

    Liu, R.; and Shah, N. B. 2023. Reviewergpt? an exploratory study on using large language models for paper reviewing. arXiv preprint arXiv:2306.00622

    2023 arXiv

  23. [31]

    M.; Uribe, J

    McAleese, N.; Pokorny, R. M.; Uribe, J. F. C.; Nitishinskaya, E.; Trebacz, M.; and Leike, J. 2024. Llm critics help catch llm bugs. arXiv preprint arXiv:2407.00215

    2024 arXiv

  24. [32]

    Meguellati, E.; Zeghina, A.; Sadiq, S.; and Demartini, G. 2025. LLM-Based Semantic Augmentation for Harmful Content Detection. In Proceedings of the International AAAI Conference on Web and Social Media, volume 19, 1190--1209

    2025

  25. [33]

    Nie, J.; Shao, H.; Fan, Y.; Shao, Q.; You, H.; Preindl, M.; and Jiang, X. 2024. LLM-based conversational AI therapist for daily functioning screening and psychotherapeutic intervention via everyday smart devices. arXiv preprint arXiv:2403.10779

    2024 arXiv

  26. [34]

    OpenAI. 2025. Openai o3 and o4-mini system card. Technical report, OpenAI

    2025

  27. [35]

    Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, 311--318

    2002

  28. [36]

    Post, M. 2018. A call for clarity in reporting BLEU scores. arXiv preprint arXiv:1804.08771

    2018 arXiv

  29. [37]

    V.; and Das, S

    Prakash, A. V.; and Das, S. 2020. Intelligent conversational agents in mental healthcare services: a thematic analysis of user perceptions. Pacific Asia Journal of the Association for Information Systems, 12(2): 1

    2020

  30. [38]

    D.; Ermon, S.; and Finn, C

    Rafailov, R.; Sharma, A.; Mitchell, E.; Manning, C. D.; Ermon, S.; and Finn, C. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 53728--53741

    2023

  31. [39]

    M.; Smiley, A.; and Finkelstein, J

    Reategui-Rivera, C. M.; Smiley, A.; and Finkelstein, J. 2025. LLM-Based Chatbot to Reduce Mental Illness Stigma in Healthcare Providers. In 2025 IEEE 15th Annual Computing and Communication Workshop and Conference (CCWC), 00001--00007. IEEE

    2025

  32. [40]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  33. [41]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  34. [42]

    Shen, X.; Wu, Y.; Qu, Y.; Backes, M.; Zannettou, S.; and Zhang, Y. 2025. HateBench: Benchmarking Hate Speech Detectors on LLM-Generated Content and Hate Campaigns. arXiv preprint arXiv:2501.16750

    2025 arXiv

  35. [43]

    W.; stergaard, S

    Topp, C. W.; stergaard, S. D.; S ndergaard, S.; and Bech, P. 2015. The WHO-5 Well-Being Index: a systematic review of the literature. Psychotherapy and psychosomatics, 84(3): 167--176

    2015

  36. [44]

    TOV, W. 2018. Well-being concepts and components. Handbook of subjective well-being, 1--15

    2018

  37. [45]

    Viswanathan, V.; Sun, Y.; Ma, S.; Kong, X.; Cao, M.; Neubig, G.; and Wu, T. 2025. Checklists Are Better Than Reward Models For Aligning Language Models. arXiv preprint arXiv:2507.18624

    2025

  38. [46]

    Wang, J.; Liang, Y.; Meng, F.; Sun, Z.; Shi, H.; Li, Z.; Xu, J.; Qu, J.; and Zhou, J. 2023. Is ChatGPT a Good NLG Evaluator? A Preliminary Study. In Proceedings of the 4th New Frontiers in Summarization Workshop, 1--11

    2023

  39. [47]

    I Like Sunnie More Than I Expected!

    Wu, S.; Cachia, J. Y.; Han, F.; Yao, B.; Xie, T.; Zhao, X.; and Wang, D. 2024. " I Like Sunnie More Than I Expected!": Exploring User Expectation and Perception of an Anthropomorphic LLM-based Conversational Agent for Well-Being Support. arXiv preprint arXiv:2405.13803

    2024 arXiv

  40. [48]

    Xiong, H.; Bian, J.; Li, Y.; Li, X.; Du, M.; Wang, S.; Yin, D.; and Helal, S. 2024. When search engine services meet large language models: visions and challenges. IEEE Transactions on Services Computing

    2024

  41. [49]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  42. [50]

    Ye, S.; Kim, D.; Kim, S.; Hwang, H.; Kim, S.; Jo, Y.; Thorne, J.; Kim, J.; and Seo, M. 2023. Flask: Fine-grained language model evaluation based on alignment skill sets. arXiv preprint arXiv:2307.10928

    2023 arXiv

  43. [51]

    Yuan, W.; Neubig, G.; and Liu, P. 2021. BARTScore: Evaluating Generated Text as Text Generation. In Ranzato, M.; Beygelzimer, A.; Dauphin, Y. N.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information ...

    2021

  44. [52]

    Zhang, C.; Feng, Z.; Zhang, Z.; Qiang, J.; Xu, G.; and Li, Y. 2025. Is LLMs Hallucination Usable? LLM-based Negative Reasoning for Fake News Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 1031--1039

    2025

  45. [53]

    Q.; and Artzi, Y

    Zhang, T.; Kishore, V.; Wu, F.; Weinberger, K. Q.; and Artzi, Y. 2020. BERTScore: Evaluating Text Generation with BERT . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net

    2020

  46. [54]

    P.; Zhang, H.; Gonzalez, J

    Zheng, L.; Chiang, W.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E. P.; Zhang, H.; Gonzalez, J. E.; and Stoica, I. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems 36: Annual Conf...

    2023

  47. [55]

    , " * write output.state after.block = add.period write newline

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

  48. [56]

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