Pith. sign in

REVIEW 4 major objections 4 minor 88 references

Cognivia, a 7B CBT-specialized model, outperforms larger LLMs on structured therapeutic response generation.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 01:47 UTC pith:6ENC6D7E

load-bearing objection Useful CBT dataset and rubric, but the effectiveness claim is circular and the distortion-recognition claim is untested. the 4 major comments →

arxiv 2607.25681 v1 pith:6ENC6D7E submitted 2026-07-28 cs.AI

Cognivia: A Cognitive Behavioral Therapy Copilot for Evidence-Based Mental Healthcare

Qi Chen , Siria Xiyueyao Luo , Jian Wang , Yuan Shi , Haocong Rao , Xuejiao Zhao This is my paper
classification cs.AI
keywords Cognitive Behavioral Therapycognitive distortionrational response generationLLM fine-tuningLoRAdata augmentationAI therapy evaluationmental health chatbot
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.

This paper sets out to show that a small, efficient language model, fine-tuned on roughly 9,400 cognitive-behavioral-therapy (CBT) triplets, can recognize cognitive distortions in user thoughts and produce rational responses that are as good as—or better than—those from far larger proprietary models. The motivation is practical: CBT is effective but therapist-scarce, and ordinary chatbots tend to be sycophantic rather than corrective. To achieve this, the authors build a dataset from expert-curated CBT literature plus augmented mental-health Q&A pairs, then fine-tune a 7B instruction-tuned model with LoRA. They also propose CogEval, a 12-dimension, clinically grounded rubric for judging AI therapy responses. On lexical metrics, LLM-as-judge evaluations under two criteria, and scoring by ten behavioral science experts, Cognivia consistently beats nine baselines including GPT-5 Mini.

Core claim

The central claim is that a 7B-parameter model (Qwen2.5-7B-Instruct) fine-tuned via LoRA on the Augmented CBT Cognitive Triplet Dataset can match or exceed GPT-5 Mini and other strong baselines in both cognitive distortion identification and rational response generation, across automatic, model-based, and human evaluations. The authors attribute this to the structured triplet representation—(thought, cognitive distortion, rational response)—derived from authoritative CBT texts and large-scale QA augmentation, which injects therapeutic reasoning that general conversational models lack. They further claim that CogEval captures the clinical qualities (safety, empathy, boundary integrity, interv

What carries the argument

The key machinery is the CBT Cognitive Triplet Dataset: each sample is (user thought, cognitive distortion, rational response), where the distortion comes from an 11-category taxonomy drawn from Burns's core CBT texts and the response is generated by GPT-5 Mini following a five-component structured prompt (empathy, distortion analysis, reflective questions, CBT exercise, encouragement). The dataset starts from 296 expert-curated seed examples and expands to 9,437 augmented PsyQA entries via two prompt stages (DeepSeek for labeling, GPT-5 Mini for generation) with quality filtering. The fine-tuning uses supervised fine-tuning with LoRA on a 7B instruction-tuned model. CogEval, the evaluation

Load-bearing premise

The load-bearing assumption is that high CogEval scores equal clinically meaningful CBT quality; but since the CogEval rubric is written into the generation prompt and also used to select the dataset and prompt configuration, the evaluation may be partly self-fulfilling, so the effectiveness claim collapses unless the rubric is independently validated and the generation is decoupled from it.

What would settle it

Use a blinded clinical panel: have licensed CBT therapists rate Cognivia and GPT-5 Mini responses generated from prompts that contain no CogEval cues, scoring on a validated CBT competence scale. If Cognivia no longer outperforms GPT-5 Mini, the paper's central claim is weakened. A stronger test: run a randomized trial where users complete a standard cognitive distortion questionnaire before and after interacting with Cognivia versus a supportive chatbot, and see if the CBT model produces larger reductions.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A lightweight 7B model can deliver structured CBT-style responses, making scalable, low-cost mental health support feasible without relying on large proprietary APIs.
  • The CogEval rubric offers a reusable, multi-dimensional standard for comparing AI therapy outputs on clinical quality dimensions rather than just fluency or empathy.
  • The pipeline of expert seed curation, LLM-based augmentation, and LoRA fine-tuning can transfer to other evidence-based psychological interventions beyond CBT.
  • Explicitly identifying cognitive distortions and generating corrective rational responses addresses the over-flattering, advice-only behavior of general conversational models.
  • Near-perfect scores on relational boundary metrics suggest AI therapy systems can be designed to avoid fostering emotional dependence.

Where Pith is reading between the lines

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

  • Because the CogEval rubric is embedded in the response-generation prompt (Appendix Figure 8) and was used to select the best prompt/dataset configuration (Table 6), the reported CogEval outperformance may partly reflect rubric-following rather than intrinsic clinical quality; an independent evaluation with prompts decoupled from the rubric is needed to confirm the claim.
  • The human evaluation is small (10 experts, 10 sample questions); the reported inter-rater reliability (ICC≈0.95) is promising, but larger and more diverse clinician panels would be needed to generalize the quality finding.
  • A natural next step, which the authors acknowledge, is a longitudinal intervention study: the paper measures response quality, not whether users' cognitive distortions actually decrease over time.
  • The 11-category distortion taxonomy is standard but may not cover culturally specific expressions of distorted thinking; testing on non-English or culturally diverse data would probe the system's generality.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents Cognivia, a CBT-oriented mental-health assistant obtained by fine-tuning Qwen2.5-7B-Instruct on a synthetic 'cognitive triplet' dataset built from PsyQA questions, DeepSeek-assigned cognitive-distortion labels, and GPT-5 Mini-generated rational responses. The authors also propose CogEval, a four-dimension/twelve-sub-dimension rubric for assessing CBT response quality, and evaluate Cognivia with lexical metrics, LLM-as-a-judge under both an existing criterion set and CogEval, and a ten-expert human study. The central claim is that Cognivia consistently outperforms nine baselines in both cognitive-distortion recognition and rational response generation.

Significance. If the claims were valid, the result would be practically significant: a 7B open-weight model fine-tuned on roughly 9.4k LLM-generated triplets could match or exceed a frontier proprietary model on structured CBT responses. The paper has useful components: an explicit 11-category distortion taxonomy, a reproducible pipeline from expert seed curation through augmentation and LoRA fine-tuning, a proposed domain-specific evaluation rubric, and released code. However, the current evidence does not establish the headline claim. The CogEval rubric is embedded in the response-generation prompt and used for data-selection, and the same rubric is then used as the primary evaluation instrument; the final model's cognitive-distortion recognition is not directly measured; automatic metrics lack uncertainty quantification; and the human evaluation is too small and under-reported to carry the load. These are not presentation issues but problems in the load-bearing evaluation design.

major comments (4)
  1. [Appendix B.2 (Figure 8), Table 6, §5.2 (Table 4)] The CogEval evaluation is circular. Figure 8 explicitly inserts the full CogEval dimension list into the rational-response generation prompt as a 'Quality Reference'. Table 6 then selects the prompt version and generator model using those same CogEval scores, and the final training data is produced by the selected configuration. Cognivia is subsequently evaluated with LLM-as-a-judge using the same CogEval rubric (Table 4). At best, this measures adherence to a rubric that was fed to the generator and used for data selection; it does not independently measure clinically meaningful CBT quality. The expert validation in Figure 3 assesses only the reasonableness of the rubric, not whether high CogEval scores correspond to better therapy. The authors should decouple the generation prompt from the evaluation rubric, or validate CogEval scores against an independent clinician-rated gold standar
  2. [Appendix D (Table 8), §5.2] The claim that Cognivia outperforms baselines in 'cognitive distortion recognition' is not directly evaluated for the final model. Table 8 reports accuracy, precision, and recall for prompt versions v1/v2 used with DeepSeek during dataset construction; it does not report distortion-recognition performance of the fine-tuned Cognivia model. No table or experiment measures the final model's ability to identify cognitive distortions on a held-out set with clinician labels. This is a central contribution listed in the abstract and introduction, so the missing evaluation leaves one of the two headline capabilities unsupported.
  3. [§5.2 (Table 2)] The lexical-metric comparisons are reported without error bars, confidence intervals, or significance tests. For example, Cognivia's ROUGE-1 is 0.592 versus GPT-5 Mini's 0.586, ROUGE-L is 0.246 versus 0.232, and BLEU-1 is 0.103 versus 0.099. These differences are small, and the text uses the word 'significant' without any statistical support. With a test set of about 1,435 instances, bootstrap confidence intervals or paired tests should be reported before claiming improvement.
  4. [§5.2 (Figure 4), Appendix E] The human evaluation is too small and insufficiently reported to support the conclusion. Only ten samples and ten experts are used, with no per-dimension scores, no confidence intervals, no description of the statistical test behind 'p < 0.05', and no details on how the reported ICC was computed. Moreover, the human raters apply the same CogEval dimensions that were embedded in the generation prompt and used for dataset selection, so this evaluation does not break the circularity identified above. The authors should provide a fully specified analysis plan, a larger and more diverse sample, and ideally an independent clinical outcome measure.
minor comments (4)
  1. [Section C] The text says the seed data are curated from 'Feeling Good Handbook [13] and Feeling Good: The New Mood Therapy [13]'; the second reference should be to [12], not [13]. Also the two book titles are both attributed to the same reference number.
  2. [§3.2 (Eq. 1)] Equation (1) states 'ChatGPT' but the surrounding text and Table 6 refer to 'GPT-5 Mini'. Please use consistent model names throughout.
  3. [Table 4] Several entries in the table are missing whitespace or contain formatting errors, e.g., '4.884.93' and '4.974.75'. These should be corrected.
  4. [Figure 4 caption] The caption reports 'p<0.05, ICC≈0.95' without specifying the test, the comparison, or the ICC model used. Add this information or remove the claim.

Circularity Check

2 steps flagged

CogEval rubric is embedded in the rational-response generation prompt (Figure 8) and used to select the training dataset (Table 6); the same rubric is then used as judge (Table 4), so the headline effectiveness claim is partly self-fulfilling rather than independently measured.

specific steps
  1. self definitional [Appendix B.2, Figure 8 (Rational Response Generation prompt)]
    "6. Quality Reference (Do Not Explicitly Output): When generating the response, ensure it aligns with the following evaluation dimensions : (A) Semantic Fidelity • Structural Clarity : ... (B) Robustness and Fault Tolerance • Situational Safety ... (C) Deployment Feasibility and User Adoption • Empathy Validation ... (D) Relational Boundary Integrity • Boundary Framing ... Anthropomorphic Restraint ..."

    The 'Quality Reference' in the generation prompt is the full CogEval rubric from Table 1 (SC, DO, SS, CA, EV, IC, CC, WF, BF, NE, DA, AR). The training responses are generated under this instruction, so the later CogEval-based evaluation measures adherence to criteria that were already fed into the data-generation step. The quality target and the evaluation metric are the same object, so high CogEval scores are not an independent confirmation of clinical quality.

  2. fitted input called prediction [Appendix D, Table 6 and dataset-selection paragraph]
    "To assess the quality of rational responses generated by different LLMs, we further select 5000 test sets randomly from the augmented cognitive triplet dataset and analyze their responses according to the evaluation criteria of Table 1. As shown in Table 6 of the Appendix, GPT-5 Mini consistently outperforms other models in most categories, so we ultimately selected the dataset generated by the GPT-5 Mini model for the subsequent training."

    The prompt version and generator used for the final SFT dataset are selected using the CogEval criteria themselves. The final model is then evaluated on the same CogEval rubric (Table 4, Figure 4). This is selection on the outcome metric: the configuration is tuned to maximize CogEval, and the same metric is later reported as evidence of 'consistently outperforms'. The prediction is statistically coupled to the selection procedure, not an out-of-sample validation of CBT quality.

full rationale

The core loop is concrete: Figure 8 injects the CogEval dimensions into the rational-response generation prompt as a 'Quality Reference'; Appendix D/Table 6 chooses the GPT-5 Mini prompt/dataset configuration by those same CogEval scores; and Table 4 / Figure 4 evaluate Cognivia with the same CogEval rubric, in both LLM-as-judge and human-expert form. Cognivia's near-perfect relational-boundary scores and general CogEval superiority are therefore largely a measure of how well the system reproduces a rubric that was written into its training-data generation and used to filter its data. This makes the paper's central effectiveness claim partly self-fulfilling. Not all evidence is circular: cognitive-distortion identification is checked against human-annotated labels, and Table 2 lexical metrics provide an external surface-level signal (though small, e.g., ROUGE-1 0.592 vs 0.586). The Table 3 existing-criteria evaluation [47] is less contaminated, but it is still administered under the Figure 6 system prompt that contains CogEval cues. Overall, the strongest numerical claim (CogEval outperformance) reduces by construction to the rubric the authors themselves built into the pipeline, so the 'consistently outperforms' conclusion requires decoupling before it can stand as independent evidence.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 1 invented entities

The central claim depends on assumptions about the validity of CBT as implemented, the completeness of the Burns taxonomy, the trustworthiness of LLM-generated labels and responses, and the clinical meaningfulness of the authors' own rubric. The rubric's use as both an optimization target and an evaluation endpoint is the largest unexamined load-bearing assumption.

free parameters (2)
  • Response generation prompt version selected by CogEval scores = V5 with GPT-5 Mini (highest scores in Table 6)
    The generation prompt embeds the CogEval dimensions as a 'Quality Reference' (Appendix Figure 8), and the V5/GPT-5 Mini configuration was selected because it scored highest on those same CogEval criteria, which are later used as the primary evaluation.
  • Eleven-category cognitive distortion taxonomy = 11 categories from Burns texts; seed counts per category in Table 7
    The closed taxonomy is chosen by hand and determines what counts as a valid distortion label, shaping both the training data and the evaluation. An incomplete or incorrect taxonomy would propagate through every downstream result.
axioms (6)
  • domain assumption CBT cognitive restructuring—identifying distortions and generating rational responses—is an effective mechanism for improving mental health.
    The system's objective and evaluation assume this clinical mechanism, but no clinical outcome is measured in the paper.
  • domain assumption Burns' 11-category taxonomy and the curated seeds are authoritative and complete for the target user population.
    §3.1 and Appendix A/Table 7 constrain all labels and response generation to this taxonomy; no evidence is given that it covers the range of real user concerns.
  • domain assumption PsyQA forum questions can be meaningfully adapted into English CBT triplets by LLMs without losing clinical or cultural fidelity.
    §3.2 uses PsyQA, a Chinese-language dataset, yet generates English-only outputs; no cross-lingual or cross-cultural validation is reported.
  • domain assumption LLM-generated annotations (DeepSeek labels, GPT-5 Mini rational responses) are acceptable ground truth for training and evaluation.
    Appendix D reports identification accuracy of only 76% with recall 60% on a 100-case sample; the rational responses are generated by a proprietary LLM without clinical validation.
  • ad hoc to paper The CogEval rubric's 12 sub-dimensions capture the clinically relevant quality of CBT responses, and Likert ratings on it are valid outcome measures.
    §3.4 and Table 1 introduce the rubric; it is the main endpoint of the paper and was also used to select the data-generation configuration, yet its clinical validity is not established.
  • domain assumption Supervised fine-tuning on synthetic single-turn triplets transfers to real-world multi-turn users.
    Section H itself notes the single-turn simplification and calls for future multi-turn and intervention studies, so real-user transfer is an unvalidated assumption.
invented entities (1)
  • CogEval quality rubric (4 dimensions, 12 sub-dimensions) no independent evidence
    purpose: Primary evaluation instrument for rational CBT responses
    It is introduced in this paper and immediately used to select the best data-generation prompt and to judge the final model. No external validation against patient outcomes or independent clinical benchmarks is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 20693 in / 12384 out tokens · 137980 ms · 2026-08-01T01:47:35.133590+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Cognivia: A Cognitive Behavioral Therapy Copilot for Evidence-Based Mental Healthcare." pith.science (2026). https://pith.science/paper/6ENC6D7E

@misc{pith2026260725681,
  author       = {Pith},
  title        = {Pith review of: Cognivia: A Cognitive Behavioral Therapy Copilot for Evidence-Based Mental Healthcare},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ENC6D7E}},
  note         = {Machine review of arXiv:2607.25681}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cognitive distortion amplifies negative emotions and contributes to mental health disorders. Cognitive Behavioral Therapy (CBT) is an effective way to address cognitive distortions, but its large-scale application is limited by the shortage of professional therapists. Although large language models (LLMs) have recently been explored for mental health applications, existing methods still suffer from limited domain specificity, overly flattering responses, and the absence of well-defined annotations for cognitive distortions. This paper proposes Cognivia, an evidence-based artificial intelligence therapist that integrates automatic cognitive distortion identification and rational response generation. Our framework is built on authoritative CBT texts widely regarded as core paradigms and standard references. It is further augmented with mental health question-answer (Q and A) data, and employs multi-stage prompting and structured generation strategies under the supervision of behavioral science experts. Then we fine-tune a lightweight LLM on this augmented CBT dataset to obtain Cognivia. In addition, we propose the first hierarchical quality evaluation framework for assessing LLM-generated rational responses, developed through collaboration between AI researchers and behavioral science experts. Cognivia is evaluated using lexical metrics, LLM-based Judges with two complementary criteria, and human evaluation by 10 behavioral science experts. It consistently outperforms the baseline methods in cognitive distortion recognition and rational response generation, demonstrating its effectiveness. Our code is available at https://github.com/SNOWTEAM2023/Cognivia.

Figures

Figures reproduced from arXiv: 2607.25681 by Haocong Rao, Jian Wang, Qi Chen, Siria Xiyueyao Luo, Xuejiao Zhao, Yuan Shi.

Figure 1
Figure 1. Figure 1: An illustrative comparison between Cognivia and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed CBT cognitive triplet construction and model training framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Expert evaluation of the CogEval criteria. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Result of human evaluation between Cognivia and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the Five-Stage CBT Response Frame [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CBT-Structured system prompt used for response [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prompt template for cognitive distortion identification [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt template for generating CBT rational responses [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Questionnaire for Human Expert Evaluation of CBT [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of human evaluation participants’ [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The interactive interface (UI) of Cognivia, which [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗

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

88 extracted references · 7 linked inside Pith

  1. [1]

    Mahyar Abbasian, Elahe Khatibi, Iman Azimi, David Oniani, Zahra Shakeri Hos- sein Abad, Alexander Thieme, Ram Sriram, Zhongqi Yang, Yanshan Wang, Bryant Lin, et al. 2024. Foundation metrics for evaluating effectiveness of healthcare conversations powered by generative AI.Npj digital medicine7, 1 (2024), 82

    2024

  2. [2]

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

    Pith/arXiv arXiv 2023

  3. [3]

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report.arXiv preprint arXiv:2305.10403(2023)

    Pith/arXiv arXiv 2023

  4. [4]

    Elham Asgari, Nina Montaña-Brown, Magda Dubois, Saleh Khalil, Jasmine Bal- loch, Joshua Au Yeung, and Dominic Pimenta. 2025. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. NPJ digital medicine8, 1 (2025), 274

    2025

  5. [5]

    2024.Cognitive therapy of depression

    Aaron T Beck, A John Rush, Brian F Shaw, Gary Emery, Robert J DeRubeis, and Steven D Hollon. 2024.Cognitive therapy of depression. Guilford Publications

    2024

  6. [7]

    2020.Cognitive behavior therapy: Basics and beyond

    Judith S Beck. 2020.Cognitive behavior therapy: Basics and beyond. Guilford Publications

    2020

  7. [8]

    Marthinus J Bekker, Kathleen M Griffiths, and Paula M Barrett. 2017. Improv- ing accessibility of cognitive behavioural therapy for children and adolescents: Review of evidence and future directions.Clinical Psychologist21, 3 (2017), 157–164

    2017

  8. [9]

    Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big?. InProceedings of the 2021 ACM conference on fairness, accountability, and transparency. 610–623

    2021

  9. [10]

    Timothy W Bickmore and Rosalind W Picard. 2005. Establishing and maintain- ing long-term human-computer relationships.ACM Transactions on Computer- Human Interaction (TOCHI)12, 2 (2005), 293–327

    2005

  10. [11]

    Praveen Kumar Bodigutla, Aditya Tiwari, Spyros Matsoukas, Josep Valls-Vargas, and Lazaros Polymenakos. 2020. Joint turn and dialogue level user satisfaction estimation on multi-domain conversations. InFindings of the Association for Computational Linguistics: EMNLP 2020. 3897–3909

    2020

  11. [12]

    1981.Feeling good

    David D Burns. 1981.Feeling good. Vol. 4. Signet Book New York, NY, USA

    1981

  12. [13]

    1989.The feeling good handbook: Using the new mood therapy in everyday life.William Morrow & Co

    David D Burns. 1989.The feeling good handbook: Using the new mood therapy in everyday life.William Morrow & Co

    1989

  13. [14]

    Yuyan Chen, Zhihao Wen, Ge Fan, Zhengyu Chen, Wei Wu, Dayiheng Liu, Zhixu Li, Bang Liu, and Yanghua Xiao. 2023. Mapo: Boosting large language model per- formance with model-adaptive prompt optimization. InFindings of the Association for Computational Linguistics: EMNLP 2023. 3279–3304

    2023

  14. [15]

    Jean Cottraux, Ivan Note, Eliane Albuisson, Saï Nan Yao, Brigitte Note, Evelyne Mollard, Françoise Bonasse, Isabelle Jalenques, Janine Guérin, and André Julien Coudert. 2000. Cognitive behavior therapy versus supportive therapy in social phobia: a randomized controlled trial.Psychotherapy and Psychosomatics69, 3 (2000), 137–146

    2000

  15. [16]

    Emma Croxford, Yanjun Gao, Elliot First, Nicholas Pellegrino, Miranda Schnier, John Caskey, Madeline Oguss, Graham Wills, Guanhua Chen, Dmitriy Dligach, et al. 2025. Evaluating clinical AI summaries with large language models as judges.npj Digital Medicine8, 1 (2025), 640

    2025

  16. [17]

    Amy JC Cuddy, Susan T Fiske, and Peter Glick. 2008. Warmth and competence as universal dimensions of social perception: The stereotype content model and the BIAS map.Advances in experimental social psychology40 (2008), 61–149

    2008

  17. [18]

    Yi Feng, Mingyang Song, Jiaqi Wang, Zhuang Chen, Guanqun Bi, Minlie Huang, Liping Jing, and Jian Yu. 2025. SS-GEN: a social story generation framework with large language models. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 1300–1308

    2025

  18. [19]

    Hadar Fisher, Nigel M Jaffe, Kristina Pidvirny, Anna O Tierney, Mia S Vaidean, Poorvesh Dongre, and Christian A Webb. 2025. Language-Based Detection of Depression with Machine Learning: Systematic Review and Meta-Analysis. (2025)

    2025

  19. [20]

    Nikolaos Flemotomos, Victor R Martinez, Zhuohao Chen, Torrey A Creed, David C Atkins, and Shrikanth Narayanan. 2021. Automated quality assess- ment of cognitive behavioral therapy sessions through highly contextualized language representations.PloS one16, 10 (2021), e0258639

    2021

  20. [21]

    Milena Gandy, Eyal Karin, Vincent J Fogliati, Susanne Meares, Olav Nielssen, Nickolai Titov, and Blake F Dear. 2018. Emotional and cognitive difficulties, help-seeking, and barriers to treatment in neurological disorders.Rehabilitation Psychology63, 4 (2018), 563

    2018

  21. [22]

    Brandon A Gaudiano. 2008. Cognitive-behavioural therapies: achievements and challenges.Evidence Based Mental Health11, 1 (2008)

    2008

  22. [23]

    Elizabeth A Gosch, Ellen Flannery-Schroeder, Christian F Mauro, and Scott N Compton. 2006. Principles of cognitive-behavioral therapy for anxiety disorders in children.Journal of Cognitive Psychotherapy20, 3 (2006), 247

    2006

  23. [24]

    Jennifer L Greenberg, Katharine A Phillips, Susanne S Hoeppner, Nicholas C Jacobson, Angela Fang, and Sabine Wilhelm. 2023. Mechanisms of cognitive behavioral therapy vs. supportive psychotherapy in body dysmorphic disorder: An exploratory mediation analysis.Behaviour research and therapy161 (2023), 104251

    2023

  24. [25]

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al . 2024. A survey on llm-as-a-judge.The Innovation(2024)

    2024

  25. [26]

    Zhijun Guo, Alvina Lai, Johan H Thygesen, Joseph Farrington, Thomas Keen, and Kezhi Li. 2024. Large language models for mental health applications: systematic review.JMIR mental health11, 1 (2024), e57400

    2024

  26. [27]

    Himanshu Gupta, Kevin Scaria, Ujjwala Anantheswaran, Shreyas Verma, Mi- hir Parmar, Saurabh Arjun Sawant, Chitta Baral, and Swaroop Mishra. 2023. Targen: Targeted data generation with large language models.arXiv preprint arXiv:2310.17876(2023)

    Pith/arXiv arXiv 2023

  27. [28]

    2010.Experience design: Technology for all the right reasons

    Marc Hassenzahl. 2010.Experience design: Technology for all the right reasons. Vol. 8. Morgan & Claypool Publishers

    2010

  28. [29]

    Yuhao He, Li Yang, Chunlian Qian, Tong Li, Zhengyuan Su, Qiang Zhang, and Xiangqing Hou. 2023. Conversational agent interventions for mental health problems: systematic review and meta-analysis of randomized controlled trials. Journal of medical Internet research25 (2023), e43862

    2023

  29. [30]

    Megan Healy, Alison Richard, and Khameer Kidia. 2022. How to reduce stigma and bias in clinical communication: a narrative review.Journal of General Internal Medicine37, 10 (2022), 2533–2540

    2022

  30. [31]

    Nathan Hodson and Simon Williamson. 2024. Can large language models replace therapists? Evaluating performance at simple cognitive behavioral therapy tasks. JMIR AI3, 1 (2024), e52500

    2024

  31. [32]

    Stefan G Hofmann, Anu Asnaani, Imke JJ Vonk, Alice T Sawyer, and Angela Fang

  32. [33]

    Yining Hua, Hongbin Na, Zehan Li, Fenglin Liu, Xiao Fang, David Clifton, and John Torous. 2025. A scoping review of large language models for generative tasks in mental health care.npj Digital Medicine8, 1 (2025), 230

    2025

  33. [34]

    Yue Huang, Siyuan Wu, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Jianfeng Gao, Chaowei Xiao, Lichao Sun, et al . 2024. Datagen: Unified synthetic dataset generation via large language models.arXiv preprint arXiv:2406.18966(2024)

    arXiv 2024

  34. [35]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. 2024. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186(2024)

    Pith/arXiv arXiv 2024

  35. [36]

    Martin Josifoski, Marija Sakota, Maxime Peyrard, and Robert West. 2023. Ex- ploiting asymmetry for synthetic training data generation: SynthIE and the case of information extraction. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 1555–1574

    2023

  36. [37]

    Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa El- houshi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, and Carole-Jean Wu. 2025. Demystifying synthetic data in llm pre-training: A systematic study of scaling laws, benefits, and pitfalls. InProceedings of the 2025 Conference on Empirical Methods in Natural Langu...

    2025

  37. [38]

    Yejin Kim, Chi-Hyun Choi, Selin Cho, Jy-yong Sohn, and Byung-Hoon Kim. 2025. Aligning large language models for cognitive behavioral therapy: a proof-of- concept study.Frontiers in Psychiatry16 (2025), 1583739

    2025

  38. [39]

    Pengxiang Lan, Haoyu Xu, Enneng Yang, Yuliang Liang, Guibing Guo, Jianzhe Zhao, and Xingwei Wang. 2025. Efficient and effective prompt tuning via prompt decomposition and compressed outer product. InProceedings of the 2025 Confer- ence of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol...

    2025

  39. [40]

    Xiaochong Lan, Zhiguang Han, Yiming Cheng, Li Sheng, Jie Feng, Chen Gao, and Yong Li. 2025. Depression detection on social media with large language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2155–2171

    2025

  40. [41]

    Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 15339–15353

    2024

  41. [42]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  42. [43]

    Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. 2024. Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579(2024)

    Pith/arXiv arXiv 2024

  43. [44]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. InText summarization branches out. 74–81

    2004

  44. [45]

    Mingxuan Liu, Yilin Ning, Salinelat Teixayavong, Xiaoxuan Liu, Mayli Mertens, Yuqing Shang, Xin Li, Di Miao, Jingchi Liao, Jie Xu, et al. 2025. A scoping review and evidence gap analysis of clinical AI fairness.npj Digital Medicine8, 1 (2025), 360. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Qi Chen et al

    2025

  45. [46]

    Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. InProceedings of the 58th annual meeting of the association for computational linguistics. 1906–1919

    2020

  46. [47]

    Hongbin Na. 2024. CBT-LLM: A Chinese large language model for cognitive behavioral therapy-based mental health question answering.arXiv preprint arXiv:2403.16008(2024)

    Pith/arXiv arXiv 2024

  47. [48]

    Clifford Nass and Youngme Moon. 2000. Machines and mindlessness: Social responses to computers.Journal of social issues56, 1 (2000), 81–103

    2000

  48. [49]

    Jekaterina Novikova, Ondřej Dušek, and Verena Rieser. 2018. RankME: Reli- able human ratings for natural language generation. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers). 72–78

    2018

  49. [50]

    Kyo-Joong Oh, Dongkun Lee, Byungsoo Ko, and Ho-Jin Choi. 2017. A chatbot for psychiatric counseling in mental healthcare service based on emotional dialogue analysis and sentence generation. In2017 18th IEEE international conference on mobile data management (MDM). IEEE, 371–375

    2017

  50. [51]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics. 311–318

    2002

  51. [52]

    Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2024. Towards building multilingual language model for medicine.Nature Communications15, 1 (2024), 8384

    2024

  52. [53]

    Byron Reeves and Clifford Nass. 1996. The media equation: How people treat computers, television, and new media like real people.Cambridge, UK10, 10 (1996), 19–36

    1996

  53. [54]

    2012.Client centered therapy (new ed)

    Carl Rogers. 2012.Client centered therapy (new ed). Hachette UK

    2012

  54. [55]

    Kyra L Rosen, Margaret Sui, Kimia Heydari, Elizabeth J Enichen, and Joseph C Kvedar. 2025. The perils of politeness: how large language models may amplify medical misinformation.NPJ Digital Medicine8, 1 (2025), 644

    2025

  55. [56]

    Matan Rubin, Joanna Z Li, Federico Zimmerman, Desmond C Ong, Amit Gold- enberg, and Anat Perry. 2025. Comparing the value of perceived human versus AI-generated empathy.Nature Human Behaviour(2025), 1–15

    2025

  56. [57]

    Ananya B Sai, Akash Kumar Mohankumar, and Mitesh M Khapra. 2022. A survey of evaluation metrics used for nlg systems.ACM Computing Surveys (CSUR)55, 2 (2022), 1–39

    2022

  57. [58]

    Aadesh Salecha, Molly E Ireland, Shashanka Subrahmanya, João Sedoc, Lyle H Ungar, and Johannes C Eichstaedt. 2024. Large language models display human- like social desirability biases in Big Five personality surveys.PNAS nexus3, 12 (2024), pgae533

    2024

  58. [59]

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R Johnston, et al. 2023. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548(2023)

    Pith/arXiv arXiv 2023

  59. [60]

    Hao Shen, Zihan Li, Minqiang Yang, Minghui Ni, Yongfeng Tao, Zhengyang Yu, Weihao Zheng, Chen Xu, and Bin Hu. 2024. Are Large Language Models Possible to Conduct Cognitive Behavioral Therapy?. In2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 3695–3700

    2024

  60. [61]

    2022.Human-centered AI

    Ben Shneiderman. 2022.Human-centered AI. Oxford University Press

    2022

  61. [62]

    Hao Sun, Zhenru Lin, Chujie Zheng, Siyang Liu, and Minlie Huang. 2021. Psyqa: A chinese dataset for generating long counseling text for mental health support. InFindings of the association for computational linguistics: ACL-IJCNLP 2021. 1489–1503

    2021

  62. [63]

    Thomas Yu Chow Tam, Sonish Sivarajkumar, Sumit Kapoor, Alisa V Stolyar, Katelyn Polanska, Karleigh R McCarthy, Hunter Osterhoudt, Xizhi Wu, Shyam Visweswaran, Sunyang Fu, et al. 2024. A framework for human evaluation of large language models in healthcare derived from literature review.NPJ digital medicine7, 1 (2024), 258

    2024

  63. [64]

    Adela C Timmons, Jacqueline B Duong, Sierra N Walters, Kayla E Carta, Grace A Jumonville, Alyssa S Carrasco, Daniela N Romero, Matthew W Ahle, Jonathan S Comer, Ishita P Khurd, et al. 2025. Bridging fair-aware artificial intelligence and co-creation for equitable mental healthcare.Nature Reviews Psychology4, 12 (2025), 793–807

    2025

  64. [65]

    2024.The cognitive behavioral therapy workbook: evidence- based cbt skills to help you manage stress, anxiety, depression, and more

    Michael A Tompkins. 2024.The cognitive behavioral therapy workbook: evidence- based cbt skills to help you manage stress, anxiety, depression, and more. New Harbinger Publications

    2024

  65. [66]

    Lara Traeger. 2020. Cognitive restructuring. InEncyclopedia of behavioral medicine. Springer, 498–498

    2020

  66. [67]

    Pengkun Wang, Zhe Zhao, HaiBin Wen, Fanfu Wang, Binwu Wang, Qingfu Zhang, and Yang Wang. 2024. Llm-autoda: Large language model-driven automatic data augmentation for long-tailed problems.Advances in Neural Information Processing Systems37 (2024), 64915–64941

    2024

  67. [68]

    Philip S Wang, Matthias Angermeyer, Guilherme Borges, Ronny Bruffaerts, Wai Tat Chiu, Giovanni De Girolamo, John Fayyad, Oye Gureje, Josep Maria Haro, Yueqin Huang, et al. 2007. Delay and failure in treatment seeking after first onset of mental disorders in the World Health Organization’s World Mental Health Survey Initiative.World psychiatry6, 3 (2007), 177

    2007

  68. [69]

    Zaitian Wang, Jinghan Zhang, Xinhao Zhang, Kunpeng Liu, Pengfei Wang, and Yuanchun Zhou. 2025. Diversity-oriented data augmentation with large lan- guage models. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 22265–22283

    2025

  69. [70]

    Adam Waytz, John Cacioppo, and Nicholas Epley. 2010. Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on psychological science5, 3 (2010), 219–232

    2010

  70. [71]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

    2022

  71. [72]

    Jia Xu, Tianyi Wei, Bojian Hou, Patryk Orzechowski, Shu Yang, Ruochen Jin, Rachael Paulbeck, Joost Wagenaar, George Demiris, and Li Shen. 2025. Men- talchat16k: A benchmark dataset for conversational mental health assistance. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5367–5378

    2025

  72. [73]

    Shicheng Xu, Liang Pang, Mo Yu, Fandong Meng, Huawei Shen, Xueqi Cheng, and Jie Zhou. 2024. Unsupervised information refinement training of large language models for retrieval-augmented generation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 133–145

    2024

  73. [74]

    Xuhai Xu, Bingsheng Yao, Yuanzhe Dong, Saadia Gabriel, Hong Yu, James Hendler, Marzyeh Ghassemi, Anind K Dey, and Dakuo Wang. 2024. Mental-llm: Lever- aging large language models for mental health prediction via online text data. Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies 8, 1 (2024), 1–32

    2024

  74. [75]

    Wei Zhai, Nan Bai, Qing Zhao, Jianqiang Li, Fan Wang, Hongzhi Qi, Meng Jiang, Xiaoqin Wang, Bing Xiang Yang, and Guanghui Fu. 2025. MentalGLM series: Explainable large language models for mental health analysis on chinese social media. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 13599–13614

    2025

  75. [76]

    Xiang Zhang, Juntai Cao, Chenyu You, and Dujian Ding. 2025. Why prompt design matters and works: A complexity analysis of prompt search space in llms. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 32525–32555

    2025

  76. [77]

    Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chunyan Miao. 2025. Medrag: Enhanc- ing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot. InProceedings of the ACM on Web Conference 2025. 4442–4457

    2025

  77. [78]

    distortions

    Ming Zhong, Yang Liu, Da Yin, Yuning Mao, Yizhu Jiao, Pengfei Liu, Chenguang Zhu, Heng Ji, and Jiawei Han. 2022. Towards a unified multi-dimensional evalua- tor for text generation. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2023–2038. Cognivia: A Cognitive Behavioral Therapy Copilot for Evidence-Based Mental...

    2022

  78. [80]

    All-or-nothing thinking

  79. [81]

    Discounting the positives

  80. [82]

    Magnification or minimization

Showing first 80 references.