Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A 7B counseling model is claimed to match a 671B reasoning model on psychology exams.

desk verdict A solid systems paper with a plausible pipeline and big reported gains, but the missing contamination analysis makes the headline 'comparable to DeepSeek-R1' claim conditional rather than proven. read the letter →

arxiv 2508.10848 v3 pith:ARAZLOO5 submitted 2025-08-14 cs.CL

Chongyuan Dai , Jinpeng Hu , Hongchang Shi , Zhuo Li , Dan Guo , Xun Yang , Meng Wang This is my paper
classification cs.CL
keywords psychologicalLLMempathyreasoningGRPOdatasynthesischain-of-thoughtPCEBCPsyExam
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

The paper tries to establish that a language model gives reliable psychological responses only when it combines three capacities at once—empathetic engagement, professional psychological knowledge, and explicit step-by-step reasoning—and that a small model can acquire all three from curated training data. To test this, it builds Psyche-R1, a 7B Chinese psychological LLM, from a data pipeline that produces over 75,000 exam-style questions with detailed rationales and 73,000 empathetic dialogues. Training uses supervised fine-tuning on the easier data plus GRPO reinforcement learning on 19,000 questions that three different LLMs all answer incorrectly, so RL effort is concentrated where the model family is collectively weak. On the PCEB Chinese counselor-examination benchmark the 7B model reaches 74.37% average accuracy, which the paper reports as comparable to the 671B DeepSeek-R1, and it also scores higher than specialized empathy-focused models on counseling empathy metrics. The reason to care: if the claim holds, a modestly sized, domain-specialized model can approach frontier-scale reasoning performance in a high-stakes professional field, which matters when mental-health expertise is scarce.

What carries the argument

The load-bearing mechanism is a two-stage data-and-training loop. First, the data pipeline: LLMs generate psychological questions from textbooks and curricula, rationales are produced by chain-of-thought prompting, and each prompt-rationale pair is iteratively refined over three rounds, keeping only revisions that still reproduce the ground-truth answer. Second, multi-LLM cross-selection: a question is classified as challenging only if three distinct LLMs (Qwen, Llama, Phi) all answer it incorrectly; those 19k questions are trained with GRPO, while the remaining questions plus 73k empathetic dialogues go through SFT. The GRPO reward combines a format reward (+1.25 for proper <think>...</thin

What would settle it

Run a near-duplicate analysis between the training QA pool (the ~90k/75k curated questions, the 8,000 filtered CPsyExam training items, and the Type II public question-bank data) and the PCEB and CPsyExam test items, using normalized exact matching, n-gram overlap, or embedding similarity. If a substantial share of test questions are duplicates or near-duplicates of training questions, the reported accuracy and the claim of parity with a 671B reasoning model are inflated. A cleaner test is to retrain the same pipeline with all overlapping or same-source questions held out and check whether the

Watch

Extended reading notes

Core claim

The central discovery claimed is that jointly integrating empathy, expertise, and reasoning—rather than optimizing any one in isolation—yields more reliable psychological responses, and that a 7B model can achieve this integration. The paper's evidence is Psyche-R1, built on Qwen2.5-7B-Instruct, whose pipeline generates and prunes synthetic psychological QA pairs with iterative prompt-rationale optimization, adds 73k empathetic dialogues, and splits the data into an SFT portion and a GRPO portion selected by multi-LLM cross-selection. In experiments, Psyche-R1 outperforms other psychological LLMs and same-scale general/reasoning models on PCEB, achieves an average standard accuracy of 74.37%

Load-bearing premise

The entire comparison is only trustworthy if the training corpus does not overlap with the test sets: the paper trains on public Chinese psychology question banks and on the CPsyExam training split, then evaluates on PCEB and the CPsyExam test set—all drawn from the same official Chinese psychology examination ecosystem—and it reports no deduplication or overlap analysis.

Editorial extensions

If this is right

  • If the central claim holds, a carefully curated 7B model can match a 671B general reasoning model on a professional exam benchmark, implying that domain-specific data and training strategy can substitute for a great deal of scale.
  • The ablation results indicate that adding GRPO on hard questions improves accuracy beyond SFT alone, with the largest gains on case-analysis items, so reinforcement learning transfers from math/code to structured psychological reasoning.
  • Reasoning training does not appear to sacrifice empathy: Psyche-R1 scores higher than prior psychological LLMs on emotional empathy, cognitive empathy, and conversation strategy metrics.
  • The data pipeline—synthetic QA generation, iterative rationale refinement, and multi-LLM hard-sample selection—is presented as a reusable recipe for building specialized reliable LLMs in other professional domains.

Reading between the lines

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

  • The multi-LLM cross-selection rule is itself a claim about curriculum: questions that three different model families all miss are the ones where RL pays off. A direct ablation would train GRPO on a random 19,000-question subset and test whether the selection rule, rather than RL alone, drives the gains.
  • The parity claim is benchmark-specific: on the CPsyExam zero-shot average, Psyche-R1's 74.90 trails DeepSeek-R1's 78.28, while on PCEB standard accuracy it leads. A fair reading is 'comparable overall, stronger on PCEB, weaker on CPsyExam,' not uniform parity.
  • A natural next test is whether exam accuracy transfers to live counseling: blind comparisons by licensed counselors or standardized patient interactions would show whether the reasoning gains produce more reliable real-world psychological responses rather than just better multiple-choice performance.
  • The same synthesis-and-selection recipe could be applied to other licensed professional exams with well-defined answers, such as clinical licensing, legal bar, or accounting certification, where the hardest samples can be identified by multi-model disagreement.
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 / 6 minor

Summary. The paper presents Psyche-R1, a 7B Chinese psychological LLM obtained from Qwen2.5-7B-Instruct through a two-stage recipe. The authors first build a large training corpus: about 75k psychological QA pairs with chain-of-thought rationales, generated by an iterative prompt-rationale optimization pipeline; 73k empathetic dialogues; plus existing PsycoLLM and CPsyExam training data. They then apply supervised fine-tuning (SFT) on the non-challenging and dialogue subsets, followed by GRPO on a 19k 'challenging' subset selected by multi-LLM cross-selection. Evaluations on PCEB and CPsyExam report strong accuracy at 7B scale and the abstract claims results comparable to 671B DeepSeek-R1; a GPT-4o-based counseling evaluation on 200 sampled PsyDT items reports improved empathy-related scores. The central claim is the joint integration of empathy, psychological expertise, and reasoning via the proposed data pipeline and hybrid SFT/GRPO training.

Significance. If the reported results withstand scrutiny, the paper is a useful and practical contribution: the data-synthesis pipeline is described in enough detail to follow, the two-stage SFT+GRPO recipe is simple, the ablation in Table 3 shows both stages improve PCEB accuracy, and the authors state they release code. The paper also addresses an underexplored direction, applying reasoning-oriented RL to the psychological domain rather than only empathy fine-tuning. The advance is, however, conditional. The headline comparison with DeepSeek-R1 rests on single-run accuracy numbers from benchmarks that share a source ecosystem with the training data, and the manuscript reports no overlap analysis between training and test sets, no error bars or significance tests, and only LLM-based counseling evaluation. These are not mere presentation issues; they determine whether the 'comparable to DeepSeek-R1' and 'reliable psychological LLM' claims are supported.

major comments (4)
  1. [Data Collection / Data Split / Benchmarks (Tables 1-2)] The evaluation may be contaminated by training/test overlap. The training corpus includes Type II 'psychological question banks from public Chinese educational platforms' and Dcp, the CPsyExam train split, while the test sets are PCEB (from the official National Psychological Counselor Examination) and the CPsyExam test set, all drawn from the same public exam ecosystem. The MinHash-LSH step is described only for clustering the synthesized QA pool, not for removing near-duplicates between training and test. No overlap statistics are reported. Because the central claim of the paper is that a 7B model is comparable to DeepSeek-R1, please report exact and near-duplicate overlap (e.g., edit-distance or n-gram based) between each training component and both test sets, and report results on a guaranteed disjoint subset, or argue convincingly why overlap is impossible.
  2. [Table 1 and Abstract] The 'comparable to DeepSeek-R1' claim is metric-sensitive. Under the authors' standard-accuracy average, Psyche-R1 achieves 74.37 vs. DeepSeek-R1's 72.95, but under the SMCQ-standard/MMCQ-elastic average, DeepSeek-R1 is ahead, 79.18 vs. 77.64. On case-analysis SMCQ, DeepSeek-R1 is 79.25 vs. Psyche-R1's 63.31, and on theory SMCQ, 92.19 vs. 87.70. The paper averages over heterogeneous tasks without confidence intervals or significance tests. Given the small aggregate gap, the headline should be qualified and accompanied by per-task uncertainty estimates (e.g., bootstrap intervals) and, where possible, significance tests.
  3. [Performance on Counseling Tasks / Table 4] The counseling evaluation is based on only 200 randomly sampled items from the PsyDT test set, evaluated by GPT-4o as the sole judge. The paper does not provide the evaluation prompt/rubric, any human validation, inter-rater reliability, or confidence intervals, yet the text states 'significant improvements.' Since empathy is one of the three pillars of Psyche-R1, this evidence is too weak to support the claim of unified empathy gains. Please provide the evaluation protocol, sample-size justification, human or multi-LLM agreement, and per-item score distributions.
  4. [Table 3 / Discussion] The ablation isolates SFT vs. SFT+RL only on PCEB standard accuracy. It does not ablate the multi-LLM cross-selection (e.g., GRPO on a random 19k subset instead of Dpc), does not separate the effect of Dcp from the effect of the synthesized challenging subset, and does not report CPsyExam or counseling results for the ablated models. Consequently, the paper's claim that the challenging-subset selection and hybrid training jointly drive the reported gains is not fully established by the evidence presented.
minor comments (6)
  1. [Figure 1] The axis label 'PCBE' should be 'PCEB.' The legend entries such as '-> Reasoning' and 'Closed' are cryptic; please use full model names or a clearer legend.
  2. [Table 4] The column header 'eognitive empathy' should read 'cognitive empathy.'
  3. [Throughout] The text references Appendix A (prompt templates), Appendix B (model versions), and Appendix C (hyperparameters), but these appendices are not present in the submitted manuscript. They must be included for the training recipe to be reproducible.
  4. [Abstract] There is a LaTeX control sequence '\logopsyche' before the model name; this appears to be a formatting artifact that should be removed.
  5. [Data Collection / Type IV] Type IV data are described as 'harvested' from mental-health support platforms. The manuscript does not discuss anonymization, consent, or data-use restrictions. For sensitive mental-health disclosures, this should be addressed even if it is not central to the technical claims.
  6. [Benchmarks and Evaluation Metrics] The exact number of items in each CPsyExam subcategory (knowledge vs. case, SMCQ vs. MMCQ) and in the PCEB case-analysis set would help readers judge statistical stability; some subsets appear small.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-level circularity; the CPsyExam/PCEB overlap risk is a validity concern, not a by-construction reduction.

full rationale

The paper's central claim is an empirical system result rather than a derived equation. Training on the CPsyExam train split (Dcp) and evaluating on the CPsyExam test set is the standard benchmark protocol; it is not a fitted parameter renamed as a prediction. The PCEB benchmark (Hu et al. 2024) comes from the same research group, but it is an externally sourced, publicly available exam-based benchmark, so the self-citation is not load-bearing for the main conclusion. The only genuinely weak point is that the paper does not report a deduplication or overlap analysis between the Type II public question banks / Dcp and the PCEB and CPsyExam test sets; if overlap existed, the reported accuracies could be inflated. However, this is a data-contamination risk requiring evidence of overlap, not a circular step exhibited by the paper's own definitions, equations, or training/evaluation construction. Per the hard rules, speculation about possible leakage does not constitute demonstrated circularity. Therefore, no circular step is identified; score 1 reflects the minor benchmark-adjacent self-citation and the unaddressed dedup risk rather than a reduction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces a trained model and dataset, not a mathematical derivation. Its central quantitative claims rest on hand-set RL rewards, an arbitrary difficulty threshold, unvalidated synthetic rationales, and an unstated assumption of no train/test overlap.

free parameters (6)
  • GRPO format reward = +1.25 / -1
    Hand-set reward values in Stage 2; changing them changes RL behavior and final accuracy.
  • GRPO accuracy reward = +1 exact, |pred∩gold|/|gold| subset, -1 otherwise
    Partial-credit scheme chosen by hand; no sensitivity analysis.
  • Multi-LLM cross-selection threshold = incorrect from all three Qwen/Llama/Phi models
    Defines which 19k samples count as challenging; arbitrary and likely affects RL gains.
  • Rationale optimization rounds R = 3
    Number of prompt-rationale refinement rounds; paper states R=3 without ablation.
  • Rationale regeneration attempts T = not reported
    Questions failing to yield correct predicted answers after T tries are pruned; T is not specified, affecting dataset size.
  • SFT/GRPO epochs = 2 epochs each
    Training schedule chosen without ablation.
assumptions (5)
  • domain assumption Synthetic rationales that lead to the correct final answer are valid training targets for reasoning.
    Section 'Rationale Generation' accepts any rationale whose predicted answer matches ground truth as high-quality, without human validation of the reasoning steps.
  • domain assumption Public question banks and CPsyExam training split do not overlap the PCEB and CPsyExam test sets.
    Invoked implicitly in Data Collection Type II and Data Split Dcp; no deduplication analysis is reported.
  • domain assumption Multi-LLM cross-selection (all three models incorrect) identifies genuinely challenging samples.
    Section 'Question Selection' assumes disagreement among Qwen, Llama, Phi marks difficulty; no validation that these are the items where RL training helps.
  • domain assumption GPT-4o ratings of empathy, cognitive empathy, conversation strategy, state, and safety reflect counseling quality.
    Section 'Performance on Counseling Tasks' uses 200 GPT-4o-rated samples; no human correlation study is reported.
  • domain assumption SFT on empathetic dialogues increases empathy and SFT on QA increases domain knowledge.
    Ablation supports this only on aggregate exam accuracy, not isolating dialogue vs QA contributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning." pith.science (2026). https://pith.science/paper/ARAZLOO5

@misc{pith2026250810848,
  author       = {Pith},
  title        = {Pith review of: Psyche-R1: Towards Reliable Psychological LLMs through Unified Empathy, Expertise, and Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ARAZLOO5}},
  note         = {Machine review of arXiv:2508.10848}
}
read the original abstract

Amidst a shortage of qualified mental health professionals, the integration of large language models (LLMs) into psychological applications offers a promising way to alleviate the growing burden of mental health disorders. Recent reasoning-augmented LLMs have achieved remarkable performance in mathematics and programming, while research in the psychological domain has predominantly emphasized emotional support and empathetic dialogue, with limited attention to reasoning mechanisms that are beneficial to generating accurate responses. Therefore, in this paper, we propose \logopsyche\textit{Psyche-R1}, the first Chinese psychological LLM that jointly integrates empathy, psychological expertise, and reasoning, built upon a novel data curation pipeline. Specifically, we design a comprehensive data synthesis pipeline that produces over 75k high-quality psychological questions paired with detailed rationales, generated through an iterative prompt-rationale optimization procedure, along with 73k empathetic dialogues. Subsequently, we employ a hybrid training strategy wherein challenging samples are identified through a multi-LLM cross-selection strategy for group relative policy optimization (GRPO) to improve reasoning ability, while the remaining data are used for supervised fine-tuning (SFT) to enhance empathetic response generation and psychological domain knowledge. Extensive experiment results demonstrate the effectiveness of \textit{Psyche-R1} across several psychological benchmarks, where our 7B \textit{Psyche-R1} achieves comparable results to 671B \texttt{DeepSeek-R1}.

Figures

Figures reproduced from arXiv: 2508.10848 by the authors.

Figure 1
Figure 1. Comparison of different LLMs on the PCEB, plot [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed pipeline for constructing the dataset and Psyche-R1. Our pipeline involves generating [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. A Qualitative example from the CPsyExam test set comparing [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 31 canonical work pages

  1. [1]

    Al Asad, N.; Pranto, M. A. M.; Afreen, S.; and Islam, M. M. 2019. Depression detection by analyzing social media posts of user. In 2019 IEEE international conference on signal processing, information, communication & systems (SPICSCON), 13--17. IEEE

    work page 2019

  2. [2]

    Chen, J.; Cai, Z.; Ji, K.; Wang, X.; Liu, W.; Wang, R.; Hou, J.; and Wang, B. 2024. Huatuogpt-o1, towards medical complex reasoning with llms. arXiv preprint arXiv:2412.18925

    arXiv 2024

  3. [3]

    Chen, Y.; Xing, X.; Lin, J.; Zheng, H.; Wang, Z.; Liu, Q.; and Xu, X. 2023. SoulChat: Improving LLMs’ Empathy, Listening, and Comfort Abilities through Fine-tuning with Multi-turn Empathy Conversations. In Findings of the Association for Computational Linguistics: EMNLP 2023, 1170--1183

    work page 2023

  4. [4]

    Chen, Z.; Lu, Y.; and Wang, W. 2023. Empowering Psychotherapy with Large Language Models: Cognitive Distortion Detection through Diagnosis of Thought Prompting. In Findings of the Association for Computational Linguistics: EMNLP 2023, 4295--4304

    work page 2023

  5. [5]

    D.; Lee, Y.; and Lim, D

    Cho, Y.; Kim, M.; Kim, S.; Kwon, O.; Kwon, R. D.; Lee, Y.; and Lim, D. 2023. Evaluating the efficacy of interactive language therapy based on LLM for high-functioning autistic adolescent psychological counseling. arXiv preprint arXiv:2311.09243

    arXiv 2023

  6. [6]

    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 preprint arXiv:2507.06261

    arXiv 2025

  7. [7]

    S.; Bryan, C

    Demszky, D.; Yang, D.; Yeager, D. S.; Bryan, C. J.; Clapper, M.; Chandhok, S.; Eichstaedt, J. C.; Hecht, C.; Jamieson, J.; Johnson, M.; et al. 2023. Using large language models in psychology. Nature Reviews Psychology, 2(11): 688--701

    work page 2023

  8. [8]

    Gao, L.; Madaan, A.; Zhou, S.; Alon, U.; Liu, P.; Yang, Y.; Callan, J.; and Neubig, G. 2023. Pal: Program-aided language models. In International Conference on Machine Learning, 10764--10799. PMLR

    work page 2023

Show all 50 references
  1. [9]

    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

  2. [10]

    J.; Krishna, R.; Lee, C.-Y.; and Pfister, T

    Hsieh, C.-Y.; Li, C.-L.; YEH, C.-K.; Nakhost, H.; Fujii, Y.; Ratner, A. J.; Krishna, R.; Lee, C.-Y.; and Pfister, T. 2023. Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. In The 61st Annual Meeting Of The Associati...

    2023

  3. [11]

    Hu, J.; Dong, T.; Gang, L.; Ma, H.; Zou, P.; Sun, X.; Guo, D.; Yang, X.; and Wang, M. 2024. Psycollm: Enhancing llm for psychological understanding and evaluation. IEEE Transactions on Computational Social Systems

    2024

  4. [12]

    Huang, Z.; Epps, J.; Joachim, D.; and Sethu, V. 2019. Natural language processing methods for acoustic and landmark event-based features in speech-based depression detection. IEEE Journal of selected topics in Signal Processing, 14(2): 435--448

    2019

  5. [13]

    P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al

    Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  6. [14]

    Jaech, A.; Kalai, A.; Lerer, A.; Richardson, A.; El-Kishky, A.; Low, A.; Helyar, A.; Madry, A.; Beutel, A.; Carney, A.; et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  7. [15]

    Kuribayashi, T.; Oseki, Y.; and Baldwin, T. 2024. Psychometric Predictive Power of Large Language Models. In Findings of the Association for Computational Linguistics: NAACL 2024, 1983--2005

    2024

  8. [16]

    Lai, T.; Shi, Y.; Du, Z.; Wu, J.; Fu, K.; Dou, Y.; and Wang, Z. 2023. Supporting the demand on mental health services with AI-based conversational large language models (LLMs). BioMedInformatics, 4(1): 8--33

    2023

  9. [17]

    Lee, D.; Park, S.; Kang, J.; Choi, D.; and Han, J. 2020. Cross-lingual suicidal-oriented word embedding toward suicide prevention. In Findings of the Association for Computational Linguistics: EMNLP 2020, 2208--2217

    2020

  10. [18]

    Lee, S.; Mac Kim, S.; Kim, M.; Kang, D.; Yang, D.; Kim, H.; Kang, M.; Jung, D.; Kim, M.; Lee, S.; et al. 2024. Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory. In Findings of the Association for Computational Linguistics: EMNLP 2024, 14...

    2024

  11. [19]

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

    2004

  12. [20]

    Liu, C.; Wang, H.; Pan, J.; Wan, Z.; Dai, Y.; Lin, F.; Bai, W.; Rueckert, D.; and Arcucci, R. 2025. Beyond distillation: Pushing the limits of medical llm reasoning with minimalist rule-based rl. arXiv preprint arXiv:2505.17952

    2025 arXiv

  13. [21]

    U.; Qiu, S.; Saqib, M.; Anwar, S.; Usman, M.; Akhtar, N.; Barnes, N.; and Mian, A

    Naveed, H.; Khan, A. U.; Qiu, S.; Saqib, M.; Anwar, S.; Usman, M.; Akhtar, N.; Barnes, N.; and Mian, A. 2023. A comprehensive overview of large language models. ACM Transactions on Intelligent Systems and Technology

    2023

  14. [22]

    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

  15. [23]

    Qiu, H.; He, H.; Zhang, S.; Li, A.; and Lan, Z. 2024. SMILE: Single-turn to Multi-turn Inclusive Language Expansion via ChatGPT for Mental Health Support. In Findings of the Association for Computational Linguistics: EMNLP 2024, 615--636

    2024

  16. [24]

    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

  17. [25]

    Shen, H.; Li, Z.; Yang, M.; Ni, M.; Tao, Y.; Yu, Z.; Zheng, W.; Xu, C.; and Hu, B. 2024. Are Large Language Models Possible to Conduct Cognitive Behavioral Therapy? In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 3695--3700. IEEE

    2024

  18. [26]

    Sheng, G.; Zhang, C.; Ye, Z.; Wu, X.; Zhang, W.; Zhang, R.; Peng, Y.; Lin, H.; and Wu, C. 2024. HybridFlow: A Flexible and Efficient RLHF Framework. arXiv preprint arXiv: 2409.19256

    2024 arXiv

  19. [27]

    Sorin, V.; Brin, D.; Barash, Y.; Konen, E.; Charney, A.; Nadkarni, G.; and Klang, E. 2024. Large language models and empathy: systematic review. Journal of medical Internet research, 26: e52597

    2024

  20. [28]

    J.; Soma, C

    Tanana, M. J.; Soma, C. S.; Kuo, P. B.; Bertagnolli, N. M.; Dembe, A.; Pace, B. T.; Srikumar, V.; Atkins, D. C.; and Imel, Z. E. 2021. How do you feel? Using natural language processing to automatically rate emotion in psychotherapy. Behavior research methods, 53(5): 2069--2082

    2021

  21. [29]

    Team, E. 2024 a . EmoLLM: Reinventing Mental Health Support with Large Language Models. https://github.com/SmartFlowAI/EmoLLM

    2024

  22. [30]

    Team, M.; Xiao, C.; Li, Y.; Han, X.; Bai, Y.; Cai, J.; Chen, H.; Chen, W.; Cong, X.; Cui, G.; et al. 2025. MiniCPM4: Ultra-Efficient LLMs on End Devices. arXiv preprint arXiv:2506.07900

    2025 arXiv

  23. [31]

    Team, Q. 2024 b . Qwen2.5: A Party of Foundation Models

    2024

  24. [32]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  25. [33]

    A.; and Millard, M

    Wolohan, J.; Hiraga, M.; Mukherjee, A.; Sayyed, Z. A.; and Millard, M. 2018. Detecting linguistic traces of depression in topic-restricted text: Attending to self-stigmatized depression with NLP. In Proceedings of the first international workshop on language cognition and comp...

    2018

  26. [34]

    Wu, S.; Huang, X.; and Lu, D. 2025. Psychological health knowledge-enhanced LLM-based social network crisis intervention text transfer recognition method. In Proceedings of the 2025 International Conference on Health Big Data, 156--161

    2025

  27. [35]

    Xiao, M.; Xie, Q.; Kuang, Z.; Liu, Z.; Yang, K.; Peng, M.; Han, W.; and Huang, J. 2024. HealMe: Harnessing Cognitive Reframing in Large Language Models for Psychotherapy. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long...

    2024

  28. [36]

    Xie, H.; Chen, Y.; Xing, X.; Lin, J.; and Xu, X. 2025. P sy DT : Using LLM s to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psychological Counseling. In Che, W.; Nabende, J.; Shutova, E.; and Pilehvar, M. T., eds., Proceedings o...

    2025

  29. [37]

    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

  30. [38]

    Yang, Q.; Wang, Z.; Chen, H.; Wang, S.; Pu, Y.; Gao, X.; Huang, W.; Song, S.; and Huang, G. 2024. PsychoGAT: A Novel Psychological Measurement Paradigm through Interactive Fiction Games with LLM Agents. In Proceedings of the 62nd Annual Meeting of the Association for Computati...

    2024

  31. [39]

    Yao, S.; Yu, D.; Zhao, J.; Shafran, I.; Griffiths, T.; Cao, Y.; and Narasimhan, K. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36: 11809--11822

    2023

  32. [40]

    Ye, J.; Xiang, L.; Zhang, Y.; and Zong, C. 2025. SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent. In Proceedings of the 31st International Conference on Computational Linguistics, 4646--4669

    2025

  33. [41]

    Yu, Q.; Zhang, Z.; Zhu, R.; Yuan, Y.; Zuo, X.; Yue, Y.; Dai, W.; Fan, T.; Liu, G.; Liu, L.; et al. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476

    2025 arXiv

  34. [42]

    Zhang, C.; Li, R.; Tan, M.; Yang, M.; Zhu, J.; Yang, D.; Zhao, J.; Ye, G.; Li, C.; and Hu, X. 2024. CPsyCoun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological Counseling. In Findings of the 62nd Annual Meeting of the Associat...

    2024

  35. [43]

    Zhang, H.; Chen, J.; Jiang, F.; Yu, F.; Chen, Z.; Chen, G.; Li, J.; Wu, X.; Zhiyi, Z.; Xiao, Q.; et al. 2023. HuatuoGPT, Towards Taming Language Model to Be a Doctor. In Findings of the Association for Computational Linguistics: EMNLP 2023, 10859--10885

    2023

  36. [44]

    Zhang, J.; He, H.; Ma, L.; Song, N.; He, S.; Zhang, S.; Qiu, H.; Zhou, Z.; Li, A.; Dai, Y.; et al. 2025. ConceptPsy: A comprehensive benchmark suite for hierarchical psychological concept understanding in LLMs. Neurocomputing, 637: 130070

    2025

  37. [45]

    Zhao, J.; Zhu, J.; Tan, M.; Yang, M.; Li, R.; Di, Y.; Zhang, C.; Ye, G.; Li, C.; Hu, X.; et al. 2025. CPsyExam: A Chinese Benchmark for Evaluating Psychology using Examinations. In Proceedings of the 31st International Conference on Computational Linguistics, 11248--11260

    2025

  38. [46]

    Zheng, C.; Sabour, S.; Wen, J.; Zhang, Z.; and Huang, M. 2023. AugESC: Dialogue Augmentation with Large Language Models for Emotional Support Conversation. In Findings of ACL

    2023

  39. [47]

    Zheng, Y.; Zhang, R.; Zhang, J.; Ye, Y.; and Luo, Z. 2024. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. In Cao, Y.; Feng, Y.; and Xiong, D., eds., Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: Syst...

    2024

  40. [48]

    Zhu, J.; Chen, Q.; Dou, H.; Li, J.; Guo, L.; Chen, F.; and Zhang, C. 2025. Dianjin-r1: Evaluating and enhancing financial reasoning in large language models. arXiv preprint arXiv:2504.15716

    2025 arXiv

  41. [49]

    , " * 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...

  42. [50]

    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 5, 2026 · model on record in the stance chip above.