Pith. sign in

REVIEW 5 major objections 5 minor 59 references

ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance

T0 review · 5 major / 5 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 A new framework, ODRA, generates CBT therapy dialogues in which patients show realistic, turn-by-turn resistance, and licensed psychologists rank its sessions first on 12 of 13 clinical metrics.

desk verdict A serious and well-built synthetic CBT data pipeline whose central clinical-robustness claim is undermined by evaluating on patients generated by the same simulator. read the letter →

arxiv 2608.04524 v1 pith:I3CMQA6Z submitted 2026-08-05 cs.CL cs.LG

Javier Rodriguez-Juan , Hiba Arnaout , Jose Garcia-Rodriguez , David Tomás , Iryna Gurevych This is my paper
classification cs.CLcs.LG
keywords syntheticdialoguegenerationcognitivebehavioraltherapychain-of-thoughtreasoningpatientresistancesycophancyLLM-as-a-judgedownstreamfine-tuningmulti-agentsimulation
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 argues that synthetic Cognitive Behavioral Therapy (CBT) sessions can satisfy two competing demands at once: following the staged structure of CBT and showing the resistant, sometimes uncooperative behavior of real patients. The authors introduce ODRA, a multi-agent generation framework in which a therapist model and a patient model alternate turns under the control of a CBT Aligner and a Resistance Orchestrator. Every therapist utterance is preceded by stage-specific chain-of-thought reasoning; every patient utterance is preceded by an update to a continuous resistance value that bounds how open or dismissive the patient is. On automatic measures and in blind ranking by licensed psychologists, ODRA sessions are rated closer to real CBT than script-based and multi-agent baselines, and models fine-tuned on ODRA data perform better against both cooperative and resistant patients. The paper thus claims that explicit resistance modeling in synthetic training data transfers to downstream therapeutic robustness.

What carries the argument

The load-bearing mechanism is the Resistance Orchestrator, a turn-level controller that maintains a continuous patient resistance $r\in[0,1]$ and its therapist-side estimate. The patient update uses the mean-reverting exponential smoothing equation $r_{t+1} = r_t + \Delta_{\mathrm{LLM}} - \alpha(r_t - r_{\mathrm{base}})$ with $\alpha=0.15$ and per-turn shifts clipped to $[-0.10,+0.15]$; the therapist estimate is updated by an exponential smoothing with a six-turn calibration window and the same clipping bounds. Around this core, the Behavioral Profiler converts the resistance value into one of three attitude intervals (Positive, Neutral, Negative) and selects a steering prompt with general rules, calibration rules, should-not rules, and in-context examples. The CBT Aligner is the second pillar: it enforces the five stages and generates stage-specific reasoning traces before each therapist utterance, so the final utterance is a distillation of a clinically checkable plan.

What would settle it

A decisive test: collect or construct a test set of therapy dialogues whose patient resistance is annotated independently of ODRA—for example, transcripts of real CBT sessions labeled for resistance by clinicians—then run the fine-tuned models against those patients. If the resistance-specific gains over text-only baselines disappear or reverse, then the claimed downstream robustness is an artifact of evaluating on the generator's own patient model.

Watch

Extended reading notes

Core claim

On its own terms, ODRA is a five-stage CBT pipeline (Opening, Cognitive Conceptualization, CBT Work, Homework, Closing) whose therapist model emits reasoning traces for each stage, checks for therapy-interfering behaviors, and predicts the patient's resistance before responding. The patient model carries a scalar resistance $r \in [0,1]$ updated each turn by mean-reverting exponential smoothing with asymmetric bounds $[-0.10,+0.15]$, and its utterances are steered by attitude-specific prompts selected from the current resistance interval. The paper reports that this design eliminates the sycophantic patient behavior of earlier methods: behavioral alignment harmonic mean reaches 1.95 versus 1.29 for the best multi-agent baseline, and expert rank is 1.09 out of 3 with agreement $\alpha=0.80$. It further reports that Llama-3 and Qwen-3.5 models fine-tuned on ODRA data exceed text-only baselines by up to 26.20% in counseling skills with cooperative patients and 17.02% with resistant ones. The authors are explicit that when resistance is deliberately disabled (ODRA-NR), that variant scores highest on therapist skills, because resistant patients make sessions harder—which they interpret as a sign of realism rather than a defect.

Load-bearing premise

The central assumption is that the patient simulator's resistance dynamics—the mean-reverting exponential smoothing with $\alpha=0.15$ and bounds $[-0.10,+0.15]$—faithfully represent how real patients resist in therapy; the downstream evaluation uses this same simulator to create test patients, so the reported robustness could stem from matching the generator rather than from clinical transfer.

Editorial extensions

If this is right

  • If ODRA is right, synthetic CBT training data can be produced with realistic resistant patient behavior, not only compliant dialogue.
  • Models fine-tuned on ODRA-generated sessions should inherit better counseling skills and improved handling of resistant patients, as shown for Llama-3 and Qwen-3.5.
  • The released 150-session, 9,577-turn dataset with 18,496 reasoning traces gives downstream researchers a clinically grounded chain-of-thought resource for therapist-model training.
  • The TIB classification step matters: ablating it lowers performance, so including therapy-interfering behavior detection is a necessary component for faithful sessions.
  • Because ODRA-NR (resistance off) achieves the highest raw therapist-skill scores, the framework's value is not uniformly higher metric scores but realistic difficulty; downstream training is where that difficulty pays off.

Reading between the lines

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

  • An independent test of the paper's strongest downstream claim would evaluate the fine-tuned models against a patient simulator that ODRA did not train or tune on, or against transcripts of real therapy sessions; the reported resistant-patient gain may otherwise reflect alignment with ODRA's own resistance dynamics.
  • The asymmetric bounds $[-0.10,+0.15]$ operationalize the negativity bias, but the paper does not fit them to real patient data; a direct comparison with resistance annotations from real CBT sessions would show whether this parameter choice transfers.
  • If the exponential-smoothing resistance model is faithful, the same orchestrator could be ported to other structured psychotherapies by swapping the stage definitions and keeping the steering mechanism.
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

5 major / 5 minor

Summary. This paper proposes ODRA, a two-model framework for synthetic CBT session generation. A CBT Aligner enforces a five-stage protocol and emits structured CoT traces, while a Resistance Orchestrator maintains a latent patient resistance variable r_t updated by mean-reverting exponential smoothing with bounded LLM-inferred shifts and uses steering prompts to condition patient utterances on attitude intervals. The authors generate 150 sessions from CACTUS intake forms, evaluate generation quality with GPT-4o-judge CTRS and behavioral metrics plus a three-psychologist ranking study, and fine-tune Llama-3 and Qwen-3.5 on the resulting datasets. The main claims are that ODRA improves over existing CBT synthesis methods and that fine-tuning on ODRA yields better downstream therapists, especially with resistant patients.

Significance. If the claims hold, this is a substantial engineering contribution: a detailed, reproducible recipe for producing structurally coherent CBT dialogues with controllable patient resistance, backed by a released dataset of 150 sessions and 18,496 reasoning traces. The design is thorough, the ablations are informative, and the expert evaluation with licensed psychologists is a genuine strength relative to most prior synthetic-counseling work. However, the significance is constrained by three evidence gaps in the paper's central claims: the headline CTRS advantage comes from the ablation variant ODRA-NR rather than the full ODRA system, the resistance hyperparameters are tuned and evaluated with the same LLM judge, and the resistant-patient downstream test reuses ODRA's own patient simulator. These issues make the current evidence insufficient for the broad conclusions drawn in the abstract.

major comments (5)
  1. [Abstract and Section 5.1, Table 1] The abstract claims that ODRA significantly outperforms existing methods across therapeutic skills, but Table 1 shows that the full ODRA system receives lower CTRS scores than MIRROR, SQPsych, and ODRA-NR on all six metrics; only the ablation variant ODRA-NR is competitive. Section 5.1 acknowledges this and redefines ODRA-NR as the appropriate benchmark, yet the abstract and conclusions still attribute the superiority to ODRA. Please reconcile the abstract and conclusions with the per-configuration results, or clearly state that the therapeutic-skill advantage holds for the version without the resistance orchestrator while the full system trades therapeutic fluidity for behavioral fidelity.
  2. [Appendix H and Section 5.2] The resistance hyperparameters (clipping bounds [−0.10,+0.15] and homeostatic reversion rate alpha = 0.15) are selected by maximizing GPT-4o judge scores, and the behavioral alignment and realism metrics in Section 5.2 are computed with the same GPT-4o judge family. This creates a tuning-evaluation dependency, so the high Resistance Alignment scores cannot be read as evidence of clinical fidelity. Please either report scores on a judge configuration or metric not used during hyperparameter selection, or add human ratings of patient resistance independent of the LLM judge.
  3. [Section 5.5, Table 4] The 'w/ Resistance' rows evaluate fine-tuned models against 'DeepSeek Patient', which is the same ODRA patient simulator (Section 3.2, Eq. (1), Appendix H) used to generate ODRA's training dialogues. A model fine-tuned on ODRA is therefore tested on the exact resistance distribution it was optimized to match, so its CTRS advantage over baselines can be explained by distribution matching rather than by transferable skill with genuinely resistant patients. The paper's claim that explicit resistance modeling 'directly translates to downstream clinical robustness' is not established by this experiment. Please test on an independent resistant-patient corpus, a separately implemented patient model, or human-validated real or simulated resistant clients.
  4. [Section 5.1 and Appendix F] The reported resistance-prediction accuracy of 0.80, MAE of 0.10, and RMSE of 0.13 measure the therapist model's agreement with the framework's internal latent variable r_t, which is itself generated by the same exponential-smoothing process. This is an internal consistency check, not a validation of the resistance variable against any external clinical ground truth. Please state this limitation in the main text and, if possible, validate a sample of r_t values against independent clinician judgments of patient resistance.
  5. [Section 5.4, Table 3] The expert ranking study compares only CACTUS, MAGneT, and ODRA; MIRROR and SQPsych, which appear in the automated evaluations, are not included. The abstract's '12 of 13 clinical metrics' claim should therefore be attributed to the three-method comparison actually performed, rather than to all existing methods.
minor comments (5)
  1. [Equation (3)] The notation p_{t+1} = p_t + [lambda_LLM(hat{p}_LLM - p_t)]^{delta_max}_{delta_min} is not defined; please state explicitly that the bracket denotes clipping to [delta_min, delta_max].
  2. [Table 1] The Turns column for ODRA-NT reads '68.1263.54', which appears to be a formatting error; insert a space or clarify the intended value.
  3. [Table 4 and throughout] The baseline name is written inconsistently as both 'MAGNET' and 'MAGneT'; unify the spelling.
  4. [Figure 12 prompt] The word 'Aditionally' should read 'Additionally'.
  5. [Section 4.2] The text reports '50 samples evaluated by at least two experts, resulting in 300 session evaluations'; with three experts and 50 samples the expected count is 150, so please clarify how the 300 figure is computed.

Circularity Check

3 steps flagged · score 6.0 of 10

Resistance-alignment scores are self-definitional and the resistant-patient downstream evaluation is generated by ODRA's own patient simulator; the abstract's 'directly translates to downstream clinical robustness' claim is partially circular.

  1. self definitional [Section 3.2 (Behavioral Profiler); Section E.1 (Behavioral Alignment); Appendix L (Negative Attitudinal Steering Prompt)]
    "'Resistance Alignment: Evaluates whether the patient’s resistance r is accurately reflected in the dialogue.' (Sec. E.1) ... 'The main objective is to make the patient reject the therapeutic process or the therapist’s current line to the degree implied by the exact resistance value.' (Appendix L)."

    The ODRA patient is explicitly prompted to express the orchestrator-assigned resistance scalar ('Always align the utterance with a negative therapeutic attitude calibrated to {scaled_resistance}'), and the Behavioral Profiler selects the steering prompt from the active resistance interval. The Resistance Alignment metric is then defined as whether 'the patient’s resistance r is accurately reflected in the dialogue.' The high ODRA score (1.96/2.00 in Table 2) therefore measures how faithfully the patient model obeys its own steering instruction, not an independent property of real resistant patients.

  2. fitted input called prediction [Section 4 (Models); Section 5.5 (Downstream Fine-tuning); Table 4]
    "'For session synthesis, we conduct an ablation study and select DeepSeek-V3.2 for its superior performance' (Sec. 4); Table 4 uses 'w/ Resistance (DeepSeek Patient)'; abstract: 'models fine-tuned on our dataset demonstrate superior therapeutic performance against both cooperative and resistant patients, validating that explicit resistance modeling in synthetic training data directly translates to downstream clinical robustness.'"

    The resistant test patients in Table 4 are generated by DeepSeek-V3.2, the same patient model (with Eq. 1 resistance dynamics, Appendix H hyperparameters, and Appendix L steering prompts) that produced the ODRA training dialogues. A therapist fine-tuned on ODRA sessions is evaluated on the exact resistance distribution it was optimized to match, so the CTRS advantage in the w/ Resistance rows can be explained by distribution matching to the ODRA generator rather than by transferable skill with genuinely resistant clients.

1 more flagged steps
  1. self definitional [Section 3.2 (Therapist Resistance Update); Section 5.1; Appendix F]
    "'While the patient resistance represents the ground-truth value, the therapist estimated counterpart is a prediction of the most likely patient resistance.' (Sec. 3.2) ... Appendix F: 'In each turn, the therapist estimates the ground-truth patient resistance ... yielding scores of 0.80, 0.10, and 0.13, respectively.'"

    The 'ground-truth' resistance is not measured from real patients; it is computed by Eq. (1) from the patient model's own inferred shift (Delta_LLM) and the framework's hyperparameters. The patient is then prompted to let that exact resistance value shape its tone and attitude, and the therapist's estimate is produced from the same utterance using the same attitude-interval rubric. The reported accuracy of 0.80 therefore quantifies how well the therapist re-reads a latent state that the framework itself wrote into the dialogue, i.e., internal consistency of the simulation.

full rationale

ODRA is not wholly circular: it builds on Beck's CBT manual and CACTUS intake profiles, and the expert preference study (Sec. 5.4) provides independent human comparison against CACTUS and MAGneT on clinical dimensions. The GPT Patient rows in Table 4 also give an external non-resistant condition. However, two load-bearing evaluation claims are closed loops. First, the Resistance Alignment score (Table 2) is generated by prompting the patient to express the exact resistance scalar the judge is then asked to detect, so it is a prompt-obedience check rather than evidence about real resistant patients. Second, the headline downstream claim about 'resistant patients' is tested only with DeepSeek Patient, which is ODRA's own patient simulator; fine-tuned models are thus evaluated on the same resistance distribution used for training, so the reported advantage can be explained by distribution matching rather than clinical transfer. The therapist resistance-prediction accuracy (0.80) similarly measures internal consistency of a framework-generated latent variable. These issues do not invalidate the expert-relative comparisons, but they undermine the abstract's stronger claim that explicit resistance modeling 'directly translates to downstream clinical robustness.'

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claims rely on a self-contained simulation loop: the resistance dynamics, the therapist's prediction target, and the downstream evaluation patients all come from the same ODRA generator. External anchors (CACTUS profiles, expert rankings) provide partial grounding.

free parameters (3)
  • homeostatic reversion rate alpha = 0.15
    Tuned on GPT-4o judge CTRS scores over three configurations (Appendix H).
  • resistance update bounds = [-0.10, +0.15]
    Tuned on GPT-4o judge CTRS scores; asymmetric bounds motivated by negativity bias (Appendix H).
  • initial attitude midpoints = 0.165, 0.495, 0.83
    Midpoints of Positive, Neutral, Negative intervals; chosen by construction, not fitted.
assumptions (4)
  • domain assumption Beck (2020) CBT protocol is the gold-standard structure for therapy sessions.
    The five-stage pipeline and techniques derive from this text; no independent clinical validation of the adapted single-session protocol.
  • domain assumption Single-session adaptation omitting session bridge, homework review, and agenda setting preserves therapeutic fidelity.
    Appendix B documents the omissions; impact on CBT completeness is not empirically measured.
  • domain assumption GPT-4o LLM-as-a-judge scores are valid proxies for clinical quality.
    Used for CTRS, behavioral alignment, and hyperparameter tuning; the paper itself acknowledges length bias.
  • ad hoc to paper Exponential smoothing with mean reversion models emotional homeostasis.
    Equation 1 postulates the mechanism without external validation against real patient resistance traces.
invented entities (1)
  • patient resistance variable r
    purpose: Internal latent state governing patient attitude, action distribution, and steering prompt.
    Defined by the framework; the ground-truth resistance is the framework's own update, never measured against real patients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance." pith.science (2026). https://pith.science/paper/I3CMQA6Z

@misc{pith2026260804524,
  author       = {Pith},
  title        = {Pith review of: ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I3CMQA6Z}},
  note         = {Machine review of arXiv:2608.04524}
}
read the original abstract

Synthetic generation of Cognitive Behavioral Therapy (CBT) sessions is challenged by two competing demands: adhering to strict therapeutic structure while modeling the resistant, unpredictable behavior of real patients. Existing script-based methods fail to capture dynamic therapeutic interactions, while multi-agent approaches struggle to adhere to CBT's sequential structure; both suffer from sycophancy, producing overly compliant patients that misrepresent real clinical settings. In this work we introduce ODRA, a novel framework for synthesizing therapy dialogues through a Chain-of-Thought (CoT) strategy grounded in foundational CBT guidelines (Beck, 2020). ODRA further incorporates a resistance orchestrator to solve patient sycophancy, which employs steering techniques to elicit behaviors aligned with their resistance level. Automated and expert evaluations show that ODRA significantly outperforms existing methods across therapeutic skills, CBT alignment, and patient behavioral fidelity, with licensed psychologists preferring ODRA sessions across 12 of 13 clinical metrics. Furthermore, models fine-tuned on our dataset demonstrate superior therapeutic performance against both cooperative and resistant patients, validating that explicit resistance modeling in synthetic training data directly translates to downstream clinical robustness.

Figures

Figures reproduced from arXiv: 2608.04524 by the authors.

Figure 1
Figure 1. The ODRA framework synthesizes CBT coun [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architectural overview of ODRA. (A) The CBT Aligner controls session execution across a five-stage pipeline. (B) The intake form models patient profile and attitude. The patient updates resistance in response to therapist input, samples an action, and generates an utterance guided by the action, updated resistance, and steering prompt. On the other side, the therapist evaluates patient utterances for Therapy-Interfe… view at source ↗
Figure 3
Figure 3. Execution example of Therapist and Patient [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: Execution example of the stage-specific rea [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Ablation study of attitudinal prompt steering [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: MIRROR dialogue excerpt where the therapist overuse the reframing technique. We highlight in [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: MIRROR dialogue excerpt where the therapist overuse the reframing technique. We highlight in [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: System prompt used for the patient model during simulation. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: System prompt used for the therapist model during simulation. [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Prompt used by the patient to generate the contextual action distribution. [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]
Figure 11
Figure 11. Figure 11: Prompt for generating patient utterances. [PITH_FULL_IMAGE:figures/full_fig_p028_11.png]
Figure 12
Figure 12. Figure 12: Prompt used by the patient to update their internal resistance. [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Prompt used by the therapist to infer patient resistance. [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: Prompt used by the therapist to perform the TIBs classification. [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]
Figure 15
Figure 15. Figure 15: Attitudinal steering prompt used for negative patients. [PITH_FULL_IMAGE:figures/full_fig_p031_15.png]
Figure 16
Figure 16. Figure 16: Attitudinal steering prompt used for neutral patients. [PITH_FULL_IMAGE:figures/full_fig_p032_16.png]
Figure 17
Figure 17. Figure 17: Attitudinal steering prompt used for positive patients. [PITH_FULL_IMAGE:figures/full_fig_p033_17.png]
Figure 18
Figure 18. Figure 18: Qualitative dialogue excerpts from an ODRA session where the patient exhibits a negative attitude. [PITH_FULL_IMAGE:figures/full_fig_p034_18.png]
Figure 19
Figure 19. Figure 19: Qualitative dialogue excerpts from an ODRA session where the patient exhibits a neutral attitude. [PITH_FULL_IMAGE:figures/full_fig_p035_19.png]
Figure 20
Figure 20. Figure 20: Qualitative dialogue excerpts from an ODRA session where the patient exhibits a positive attitude. [PITH_FULL_IMAGE:figures/full_fig_p036_20.png]
Figure 21
Figure 21. Figure 21: Qualitative reasoning traces from a sampled therapist turn within the CBT Work stage (Part I). [PITH_FULL_IMAGE:figures/full_fig_p037_21.png]
Figure 22
Figure 22. Figure 22: Qualitative reasoning traces from a sampled therapist turn within the CBT Work stage (Part II). [PITH_FULL_IMAGE:figures/full_fig_p038_22.png]
Figure 23
Figure 23. Figure 23: Qualitative reasoning traces from a sampled patient turn exhibiting a negative attitude. [PITH_FULL_IMAGE:figures/full_fig_p039_23.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 30 canonical work pages

  1. [1]

    A Survey of Large Language Models in Psychotherapy: Current Landscape and Future Directions

    Na, Hongbin and Hua, Yining and Wang, Zimu and Shen, Tao and Yu, Beibei and Wang, Lilin and Wang, Wei and Torous, John and Chen, Ling. A Survey of Large Language Models in Psychotherapy: Current Landscape and Future Directions. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.385

    doi:10.18653/v1/2025.findings-acl.385 2025

  2. [2]

    Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory

    Lee, Suyeon and Kim, Sunghwan and Kim, Minju and Kang, Dongjin and Yang, Dongil and Kim, Harim and Kang, Minseok and Jung, Dayi and Kim, Min Hee and Lee, Seungbeen and Chung, Kyong-Mee and Yu, Youngjae and Lee, Dongha and Yeo, Jinyoung. Cactus: Towards Psychological Counseling Conversations using Cognitive Behavioral Theory. Findings of the Association fo...

    doi:10.18653/v1/2024.findings-emnlp.832 2024

  3. [3]

    A Synthetic Data Generation Framework for Grounded Dialogues

    Bao, Jianzhu and Wang, Rui and Wang, Yasheng and Sun, Aixin and Li, Yitong and Mi, Fei and Xu, Ruifeng. A Synthetic Data Generation Framework for Grounded Dialogues. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.608

    work page doi:10.18653/v1/2023.acl-long.608 2023

  4. [4]

    and Sherrill, Andrew M

    Bn, Suhas and Mattioli, Dominik O. and Sherrill, Andrew M. and Arriaga, Rosa I. and Wiese, Christopher and Abdullah, Saeed. How Real Are Synthetic Therapy Conversations? Evaluating Fidelity in Prolonged Exposure Dialogues. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.1144

    work page doi:10.18653/v1/2025.findings-emnlp.1144 2025

  5. [5]

    2026 , eprint=

    Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Client Psychological Graphs , author=. 2026 , eprint=

    work page 2026

  6. [6]

    CP sy C oun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for C hinese Psychological Counseling

    Zhang, Chenhao and Li, Renhao and Tan, Minghuan and Yang, Min and Zhu, Jingwei and Yang, Di and Zhao, Jiahao and Ye, Guancheng and Li, Chengming and Hu, Xiping. CP sy C oun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for C hinese Psychological Counseling. Findings of the Association for Computational Linguistics: ACL 2024. ...

    doi:10.18653/v1/2024.findings-acl.830 2024

  7. [7]

    2026 , eprint=

    MAGneT: Coordinated Multi-Agent Generation of Synthetic Multi-Turn Mental Health Counseling Sessions , author=. 2026 , eprint=

    work page 2026

  8. [8]

    CAMI : A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration

    Yang, Yizhe and Achananuparp, Palakorn and Huang, Heyan and Jiang, Jing and Kit, Phey Ling and Lim, Nicholas Gabriel and Ern, Cameron Tan Shi and Lim, Ee-Peng. CAMI : A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics ...

    work page doi:10.18653/v1/2025.acl-long.1024 2025

Show all 59 references
  1. [9]

    2026 , eprint=

    CALM-IT: Generating Realistic Long-Form Motivational Interviewing Dialogues with Dual-Actor Conversational Dynamics Tracking , author=. 2026 , eprint=

    2026

  2. [10]

    H eal M e: Harnessing Cognitive Reframing in Large Language Models for Psychotherapy

    Xiao, Mengxi and Xie, Qianqian and Kuang, Ziyan and Liu, Zhicheng and Yang, Kailai and Peng, Min and Han, Weiguang and Huang, Jimin. H eal M e: Harnessing Cognitive Reframing in Large Language Models for Psychotherapy. Proceedings of the 62nd Annual Meeting of the Association ...

    2024 doi

  3. [11]

    2025 , eprint=

    Roleplaying with Structure: Synthetic Therapist-Client Conversation Generation from Questionnaires , author=. 2025 , eprint=

    2025

  4. [12]

    2025 , issn =

    Toward large reasoning models: A survey of reinforced reasoning with large language models , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.patter.2025.101370 , url =

    2025

  5. [13]

    CATCH : A Novel Data Synthesis Framework for High Therapy Fidelity and Memory-Driven Planning Chain of Thought in AI Counseling

    Chen, Mingyu and Lin, Jingkai and Chu, Zhaojie and Xing, Xiaofen and Chen, Yirong and Xu, Xiangmin. CATCH : A Novel Data Synthesis Framework for High Therapy Fidelity and Memory-Driven Planning Chain of Thought in AI Counseling. Findings of the Association for Computational Li...

    2025 doi

  6. [14]

    Multimodal Cognitive Reframing Therapy via Multi-hop Psychotherapeutic Reasoning

    Kim, Subin and Kim, Hoonrae and Do, Heejin and Lee, Gary. Multimodal Cognitive Reframing Therapy via Multi-hop Psychotherapeutic Reasoning. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Languag...

    2025 doi

  7. [15]

    P sy A dvisor: A Plug-and-Play Strategy Advice Planner with Proactive Questioning in Psychological Conversations

    Hu, Yuxin and Liu, Danni and Liu, Bo and Chen, Yida and Cao, Jiuxin and Liu, Yan. P sy A dvisor: A Plug-and-Play Strategy Advice Planner with Proactive Questioning in Psychological Conversations. Proceedings of the 63rd Annual Meeting of the Association for Computational Lingu...

    2025 doi

  8. [16]

    Towards Understanding Sycophancy in Language Models , url =

    Sharma, Mrinank and Tong, Meg and Korbak, Tomek and Duvenaud, David and Askell, Amanda and Bowman, Sam and DURMUS, Esin and Hatfield-Dodds, Zac and Johnston, Scott and Kravec, Shauna and Maxwell, Timothy and McCandlish, Sam and Ndousse, Kamal and Rausch, Oliver and Schiefer, N...

  9. [17]

    CONSENSAGENT : Towards Efficient and Effective Consensus in Multi-Agent LLM Interactions Through Sycophancy Mitigation

    Pitre, Priya and Ramakrishnan, Naren and Wang, Xuan. CONSENSAGENT : Towards Efficient and Effective Consensus in Multi-Agent LLM Interactions Through Sycophancy Mitigation. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings...

    2025 doi

  10. [18]

    Measuring Sycophancy of Language Models in Multi-turn Dialogues

    Hong, Jiseung and Byun, Grace and Kim, Seungone and Shu, Kai. Measuring Sycophancy of Language Models in Multi-turn Dialogues. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.121

    2025 doi

  11. [19]

    MIRROR : Multimodal Cognitive Reframing Therapy for Rolling with Resistance

    Kim, Subin and Kim, Hoonrae and Lee, Jihyun and Jeon, Yejin and Lee, Gary. MIRROR : Multimodal Cognitive Reframing Therapy for Rolling with Resistance. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.751

    2025 doi

  12. [20]

    2023 , issn =

    Are mental health awareness efforts contributing to the rise in reported mental health problems? A call to test the prevalence inflation hypothesis , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.newideapsych.2023.101010 , url =

    2023

  13. [21]

    and Mei, Cristina and Dalal, Naeem and Alvarez-Jimenez, Mario and Blakemore, Sarah-Jayne and Browne, Vivienne and Dooley, Barbara and others , title=

    McGorry, Patrick D. and Mei, Cristina and Dalal, Naeem and Alvarez-Jimenez, Mario and Blakemore, Sarah-Jayne and Browne, Vivienne and Dooley, Barbara and others , title=. The Lancet Psychiatry , year=. doi:10.1016/S2215-0366(24)00163-9 , url=

    doi

  14. [22]

    and Stirman, Shannon Wiltsey and Ungar, Lyle H

    Stade, Elizabeth C. and Stirman, Shannon Wiltsey and Ungar, Lyle H. and Boland, Cody L. and Schwartz, H. Andrew and Yaden, David B. and Sedoc, Jo. Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation , jo...

    2024 doi

  15. [23]

    and Birnbaum, Michael L

    Nguyen, Viet Cuong and Taher, Mohammad and Hong, Dongwan and Possobom, Vinicius Konkolics and Gopalakrishnan, Vibha Thirunellayi and Raj, Ekta and Li, Zihang and Soled, Heather J. and Birnbaum, Michael L. and Kumar, Srijan and De Choudhury, Munmun. Do Large Language Models Ali...

    2025 doi

  16. [24]

    When Can We Trust LLM s in Mental Health? Large-Scale Benchmarks for Reliable LLM Evaluation

    Badawi, Abeer and Rahimi, Elahe and Laskar, Md Tahmid Rahman and Grach, Sheri and Bertrand, Lindsay and Danok, Lames and Dhanesh, Prathiba and Huang, Jimmy and Rudzicz, Frank and Dolatabadi, Elham. When Can We Trust LLM s in Mental Health? Large-Scale Benchmarks for Reliable L...

    2026 doi

  17. [25]

    2020 , publisher=

    Cognitive behavior therapy: Basics and beyond , author=. 2020 , publisher=

    2020

  18. [26]

    Harnessing the power of synthetic data in healthcare: innovation, application, and privacy , journal=

    Giuffr. Harnessing the power of synthetic data in healthcare: innovation, application, and privacy , journal=. 2023 , month=. doi:10.1038/s41746-023-00927-3 , url=

    2023 doi

  19. [27]

    2018 , issn =

    Cognitive distortions in patients with social anxiety disorder: Comparison of a clinical group and healthy controls , journal =. 2018 , issn =. doi:https://doi.org/10.1016/j.ejpsy.2017.08.004 , url =

    2018 doi

  20. [28]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Cognitive reframing of negative thoughts through human-language model interaction , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  21. [29]

    Metamo: Empowering Large Language Models with Psychological Distortion Detection for Cognition-aware Coaching

    Hotta, Hajime and Le, Huu-Loi and Phan, Manh-Cuong and Nguyen, Minh-Tien. Metamo: Empowering Large Language Models with Psychological Distortion Detection for Cognition-aware Coaching. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Syst...

    2025 doi

  22. [30]

    World Psychiatry , volume =

    Cuijpers, Pim , title =. World Psychiatry , volume =. doi:https://doi.org/10.1002/wps.20661 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/wps.20661 , year =

    doi

  23. [31]

    BMC psychiatry , volume=

    Associations of negative cognitions, emotional regulation, and depression symptoms across four continents: International support for the cognitive model of depression , author=. BMC psychiatry , volume=. 2020 , publisher=

    2020

  24. [32]

    and Zhi, Jiayin and Eack, Shaun M

    Wang, Ruiyi and Milani, Stephanie and Chiu, Jamie C. and Zhi, Jiayin and Eack, Shaun M. and Labrum, Travis and Murphy, Samuel M and Jones, Nev and Hardy, Kate V and Shen, Hong and Fang, Fei and Chen, Zhiyu. PATIENT - : Using Large Language Models to Simulate Patients for Train...

    2024 doi

  25. [33]

    Journal of medical Internet research , volume=

    Designing a chatbot for a brief motivational interview on stress management: qualitative case study , author=. Journal of medical Internet research , volume=. 2019 , publisher=

    2019

  26. [34]

    Review of general psychology , volume=

    Bad is stronger than good , author=. Review of general psychology , volume=. 2001 , publisher=

    2001

  27. [35]

    The Scandinavian Psychoanalytic Review , volume=

    Homeostasis and self-regulation , author=. The Scandinavian Psychoanalytic Review , volume=. 2024 , publisher=

    2024

  28. [36]

    Automatic robust estimation for exponential smoothing: Perspectives from statistics and machine learning , journal =

    Devon Barrow and Nikolaos Kourentzes and Rickard Sandberg and Jacek Niklewski , keywords =. Automatic robust estimation for exponential smoothing: Perspectives from statistics and machine learning , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.eswa.2020.113637 , url =

    2020

  29. [37]

    Advances in Neural Information Processing Systems , volume=

    Improved representation steering for language models , author=. Advances in Neural Information Processing Systems , volume=

  30. [38]

    Advances in Neural Information Processing Systems , volume=

    Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization , author=. Advances in Neural Information Processing Systems , volume=

  31. [39]

    2025 , eprint=

    Persona Vectors: Monitoring and Controlling Character Traits in Language Models , author=. 2025 , eprint=

    2025

  32. [40]

    Steering off Course: Reliability Challenges in Steering Language Models

    Da Silva, Patrick Queiroz and Sethuraman, Hari and Rajagopal, Dheeraj and Hajishirzi, Hannaneh and Kumar, Sachin. Steering off Course: Reliability Challenges in Steering Language Models. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V...

    2025 doi

  33. [41]

    arXiv preprint arXiv:2501.17148 , year=

    Axbench: Steering llms? even simple baselines outperform sparse autoencoders , author=. arXiv preprint arXiv:2501.17148 , year=

    arXiv

  34. [42]

    arXiv preprint arXiv:2510.04484 , year=

    Psychological steering in llms: An evaluation of effectiveness and trustworthiness , author=. arXiv preprint arXiv:2510.04484 , year=

    arXiv

  35. [43]

    A Survey on In-context Learning

    Dong, Qingxiu and Li, Lei and Dai, Damai and Zheng, Ce and Ma, Jingyuan and Li, Rui and Xia, Heming and Xu, Jingjing and Wu, Zhiyong and Chang, Baobao and Sun, Xu and Li, Lei and Sui, Zhifang. A Survey on In-context Learning. Proceedings of the 2024 Conference on Empirical Met...

    2024 doi

  36. [44]

    Are Emergent Abilities in Large Language Models just In-Context Learning?

    Lu, Sheng and Bigoulaeva, Irina and Sachdeva, Rachneet and Tayyar Madabushi, Harish and Gurevych, Iryna. Are Emergent Abilities in Large Language Models just In-Context Learning?. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1...

    2024 doi

  37. [45]

    The Learnability of In-Context Learning , url =

    Wies, Noam and Levine, Yoav and Shashua, Amnon , booktitle =. The Learnability of In-Context Learning , url =

  38. [46]

    Why people get in their own way

    Chapman, Alexander L and Rosenthal, M Zachary. Why people get in their own way. Managing therapy-interfering behavior: Strategies from dialectical behavior therapy

  39. [47]

    arXiv preprint arXiv:2412.19437 , year=

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    arXiv

  40. [48]

    2025 , eprint=

    Self-Exploring Language Models for Explainable Link Forecasting on Temporal Graphs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  41. [49]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    arXiv

  42. [50]

    Behavior Therapy , volume=

    The structure of competence: Evaluating the factor structure of the Cognitive Therapy Rating Scale , author=. Behavior Therapy , volume=. 2020 , publisher=

    2020

  43. [51]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  44. [52]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  45. [53]

    Cognitive Therapy and Research , volume=

    Using motivational interviewing to manage process markers of ambivalence and resistance in cognitive behavioral therapy , author=. Cognitive Therapy and Research , volume=. 2018 , publisher=

    2018

  46. [54]

    CARFT : Boosting LLM Reasoning via Contrastive Learning with Annotated Chain-of-Thought-based Reinforced Fine-Tuning

    Zhu, Wenqiao and Liu, Ji and Zhang, Rongjunchen and Wu, Haipang and Zhang, Yulun. CARFT : Boosting LLM Reasoning via Contrastive Learning with Annotated Chain-of-Thought-based Reinforced Fine-Tuning. Proceedings of the 2025 Conference on Empirical Methods in Natural Language P...

    2025 doi

  47. [55]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    arXiv

  48. [56]

    arXiv preprint arXiv:2603.25196 , year=

    A Decade-Scale Benchmark Evaluating LLMs' Clinical Practice Guidelines Detection and Adherence in Multi-turn Conversations , author=. arXiv preprint arXiv:2603.25196 , year=

  49. [57]

    Training Models to Generate, Recognize, and Reframe Unhelpful Thoughts

    Maddela, Mounica and Ung, Megan and Xu, Jing and Madotto, Andrea and Foran, Heather and Boureau, Y-Lan. Training Models to Generate, Recognize, and Reframe Unhelpful Thoughts. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo...

    2023 doi

  50. [58]

    Advances in neural information processing systems , volume=

    Qlora: Efficient finetuning of quantized llms , author=. Advances in neural information processing systems , volume=

  51. [59]

    Sentence-bert: Sentence embeddings using siamese bert-networks , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

Pith tools

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