Elamur Research⌕
☰
home / papers / 2026 / TRACE: Two-Phase RL for Causal Graph Exploration and Deeper Psychological Intervention in Dynamic Counseling Scenarios

TRACE: Two-Phase RL for Causal Graph Exploration and Deeper Psychological Intervention in Dynamic Counseling Scenarios

Paper · 2026 · Shilin Tang, Zunyi Yin, Xuefeng Liang et al.

Summary

Current mental-health LLMs such as CRISP, SoulChat 2.0, SoulChat R1 and CACTUS tend to offer generic empathy or follow fixed CBT scripts. They rarely work out why a client is distressed. TRACE draws on clinical case conceptualization and treats counseling as a task of rebuilding, turn by turn, a hidden causal graph of the client's psychology. The graph runs from observable reactions down to core beliefs, while information comes out only gradually. The authors built a dataset of 2,000 PsyQA-derived client profiles in Chinese and English, each paired with an ideal four-level causal graph generated by DeepSeek-v3.1. They then trained a counselor model with two-phase reinforcement learning. In the first phase, rewards favor uncovering graph nodes and links in a surface-to-deep order. In the second, rewards favor restructuring the irrational beliefs that were found. According to the authors, TRACE beats prior models on graph reconstruction, intervention effectiveness and overall dialogue quality. The work matters because it moves LLM counseling toward interventions grounded in the mechanisms behind distress, instead of reacting only to the emotions a client states.

Key points

  • Causal graph design: The ideal causal graph is a DAG with four node levels: surface (A), intermediate (B), deep (C) and core (D). Edges encode causal or maintenance relations.
  • Dataset:
    • It contains 2,000 cases sampled from PsyQA.
    • A manual check of 500 graphs found that 90.4% met the clinical conceptualization criteria.
    • The data is split into D_cold (900, cold-start training), D_RL (1,000, RL environment simulation) and D_test (100, evaluation). D_RL and D_test were manually verified.
  • Evaluation reliability:
    • DeepSeek-R1-0528 served as the automatic judge.
    • Five psychology graduate students rated 150 dialogues, with inter-annotator Cohen's κ = 0.73.
    • Spearman correlations between human and LLM scores ranged from 0.63 to 0.76 across dimensions, with p < 0.05.
  • Phase-switch threshold: The causal graph filling ratio φ_trans controls when the model moves from exploration to intervention. A value of 0.8 worked best (E_int = 1.37), compared with 1.05, 1.19 and 1.14 for 0.6, 0.7 and 0.9.
    • Switching too early leads to speculative interventions.
    • Switching too late leaves too little room for intervention.
  • Ablation and limitations:
    • Removing intervention-phase RL hurts both intervention efficacy and general dialogue quality.
    • All training and testing used simulated AI clients, so clinical validation is still needed.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Beijing Normal UniversityBeijing · China15 works · 3,177 citations🎓Tohoku UniversitySendai · Japan20 works · 514 citations🎓Xidian UniversityXi'an · China14 works · 1,357 citations🎓Guangzhou Institute of TechnologyGuangzhou · China2 works · 5 citations🎓Beijing Normal University, ZhuhaiZhuhai · China3 works

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksClinical Psychology · 20 worksApplied Psychology · 21 works

Cite

Cite as Shilin Tang, Zunyi Yin, Xuefeng Liang, Guanghui Shi, Song Tong and Chen Guangyu. 2026. TRACE: Two-Phase RL for Causal Graph Exploration and Deeper Psychological Intervention in Dynamic Counseling Scenarios. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1711

@inproceedings{tang2026trace,
  title = {{TRACE: Two-Phase RL for Causal Graph Exploration and Deeper Psychological Intervention in Dynamic Counseling Scenarios}},
  author = {Shilin Tang, Zunyi Yin, Xuefeng Liang, Guanghui Shi, Song Tong and Chen Guangyu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1711},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1711}
}

References

44 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. John R. Anderson. 2014. Rules of the Mind. Psychology Press. find
  2. Judith S. Beck. 2020. Cognitive Behavior Therapy: Basics and Beyond, 3rd edition. Guilford Publications. find
  3. Peter J. Bieling and Willem Kuyken. 2003. Is cognitive case formulation science or science fiction? Clinical Psychology: Science and Practice, 10(1):52. find
  4. Denny Borsboom and Angélique O. J. Cramer. 2013. Network analysis: An integrative approach to the structure of psychopathology. Annual Review of Clinical Psychology, 9(1):91–121. find
  5. ByteDance Seed Team. 2025. Introduction to techniques used in Seed1.6. ByteDance Seed Blog. find
  6. Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of ACL, pages 2318–2335.
  7. Mingyu Chen, Jingkai Lin, Zhaojie Chu, Xiaofen Xing, Yirong Chen, and Xiangmin Xu. 2025. CATCH: A novel data synthesis framework for high therapy fidelity and memory-driven planning chain of thought in AI counseling. In Findings of EMNLP, pages 10254–10286. find
  8. Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. SoulChat: Improving LLMs’ empathy, listening, and comfort abilities through fine-tuning with multi-turn empathy conversations. In Findings of EMNLP, pages 1170–1183. find
  9. Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V Le, Sergey Levine, and Yi Ma. 2025. SFT memorizes, RL generalizes: A comparative study of foundation model post-training. In Proceedings of ICML, volume 267, pages 10818–10838.
  10. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  11. DeepSeek. 2025. DeepSeek-V3.1 release. DeepSeek API Docs. find
  12. Eiko I. Fried, Claudia D. van Borkulo, Angélique O. J. Cramer, Lynn Boschloo, Robert A. Schoevers, and Denny Borsboom. 2017. Mental disorders as networks of problems: A review of recent insights. Social Psychiatry and Psychiatric Epidemiology, 52(1):1–10. find
  13. Xing Fu, Haozhen Li, Bichen Wang, Hao Yang, Yanyan Zhao, and Bing Qin. 2025. Look beyond feeling: Unveiling latent needs from implicit expressions for proactive emotional support. In Proceedings of EMNLP, pages 21582–21609. find
  14. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 543 o… — The Llama 3 Herd of Models
  15. Marco Guevara, Shan Chen, Spencer Thomas, Tafadzwa L Chaunzwa, Idalid Franco, Benjamin H Kann, Shalini Moningi, Jack M Qian, Madeleine Goldstein, Susan Harper, and Hugo J W L Aerts. 2024. Large language models to identify social determinants of health in electronic health records. NPJ Digital Medici… find
  16. Stephen N. Haynes, William H. O’Brien, and Antonio Godoy. 2020. A proposed model for the psychometric evaluation of clinical case formulations with quantified causal diagrams. Psychological Assessment, 32(6):541. find
  17. Stephen N. Haynes, William H. O’Brien, Joseph Keawe’aimoku, Cilia Witteman, and Ph D Ka-holokula. 2012. Concepts of causality in psychopathology: Applications in clinical assessment, clinical case formulation and functional analysis. Journal of Unified Psychotherapy and Clinical Science, 1(1):87–103… find
  18. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In Proceedings of ICLR.
  19. He Hu, Yucheng Zhou, Juzheng Si, Qianning Wang, Hengheng Zhang, Fuji Ren, Fei Ma, Laizhong Cui, and Qi Tian. 2025. Beyond empathy: Integrating diagnostic and therapeutic reasoning with large language models for mental health counseling. arXiv preprint arXiv:2505.15715. — [arXiv:2505.15715](https://arxiv.org/abs/2505.15715) find
  20. Philip N. Johnson-Laird. 1983. Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness. Harvard University Press. find
  21. Juhee Kim, Chunghu Mok, Jisun Lee, Hyang Sook Kim, and Yohan Jo. 2025. Dialogue systems for emotional support via value reinforcement. In Proceedings of ACL (Volume 1: Long Papers), pages 28733–28766. find
  22. Willem Kuyken, Christine A. Padesky, and Robert Dudley. 2011. Collaborative Case Conceptualization: Working Effectively with Clients in Cognitive-Behavioral Therapy. Guilford Press. find
  23. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with PagedAttention. In Proceedings of SOSP, pages 611–626.
  24. Suyeon Lee, Sunghwan Kim, Minju Kim, Dongjin Kang, Dongil Yang, Harim Kim, Minseok Kang, Dayi Jung, Min Hee Kim, Seungbeen Lee, Kyong-Mee Chung, Youngjae Yu, Dongha Lee, and Jinyoung Yeo. 2024. Cactus: Towards psychological counseling conversations using cognitive behavioral theory. In Findings of E… find
  25. Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. Towards emotional support dialog systems. In Proceedings of ACL (Volume 1: Long Papers), pages 3469–3483.
  26. William R. Miller and Gary S. Rose. 2009. Toward a theory of motivational interviewing. American Psychologist, 64(6):527. find
  27. Clarissa W. Ong, Hiba Arnaout, Kate Sheehan, Estella Fox, Eugen Owtscharow, and Iryna Gurevych. 2025. Using large language models to create personalized networks from therapy sessions. arXiv preprint arXiv:2512.05836. — [arXiv:2512.05836](https://arxiv.org/abs/2512.05836) find
  28. OpenAI. 2024. GPT-4o mini: Advancing cost-efficient intelligence. OpenAI Blog. find
  29. OpenAI. 2025. GPT-5. OpenAI Product Page. find
  30. Jacqueline B. Persons. 2008. The Case Formulation Approach to Cognitive-Behavior Therapy. Guilford Press. find
  31. Huachuan Qiu, Hongliang He, Shuai Zhang, Anqi Li, and Zhenzhong Lan. 2024. SMILE: Single-turn to multi-turn inclusive language expansion via Chat-GPT for mental health support. In Findings of EMNLP, pages 615–636. find
  32. Qwen Team. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. find
  33. Alberto Sanfeliu and King-Sun Fu. 2012. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, (3):353–362. find
  34. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y K Li, Yang Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  35. Haojie Xie, Yirong Chen, Xiaofen Xing, Jingkai Lin, and Xiangmin Xu. 2025. PsyDT: Using LLMs to construct the digital twin of psychological counselor with personalized counseling style for psychological counseling. In Proceedings of ACL (Volume 1: Long Papers), pages 1081–1115.
  36. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  37. Y. Yang, P. Achananuparp, H.Y. Huang, J. Jiang, P.L. Kit, N.G. Lim, C.T.S. Ern, and E.P. Lim. 2025b. CAMI: A counselor agent supporting motivational interviewing through state inference and topic exploration. In Proceedings of ACL (Volume 1: Long Papers), pages 21037–21081. find
  38. Xianrong Yao, Dong She, Chenxu Zhang, Yimeng Zhang, Yueru Sun, Noman Ahmed, Yang Gao, and Zhanpeng Jin. 2025. Empathy-r1: A chain-of-empathy and reinforcement learning framework for long-form mental health support. arXiv preprint arXiv:2509.14851. — [arXiv:2509.14851](https://arxiv.org/abs/2509.14851) find
  39. Jiahao Yuan, Zhiqing Cui, Hanqing Wang, Yuansheng Gao, Yucheng Zhou, and Usman Naseem. 2025. Kardia-r1: Unleashing LLMs to reason toward understanding and empathy for emotional support via rubric-as-judge reinforcement learning. arXiv preprint arXiv:2512.01282. — [arXiv:2512.01282](https://arxiv.org/abs/2512.01282) find
  40. Jinfeng Zhou, Yuxuan Chen, Jianing Yin, Yongkang Huang, Yihan Shi, Xikun Zhang, Libiao Peng, Rongsheng Zhang, Tangjie Lv, Zhipeng Hu, Hongning Wang, and Minlie Huang. 2025. Crisp: Cognitive restructuring of negative thoughts through multi-turn supportive dialogues. In Proceedings of EMNLP, pages 324…
  41. 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. In Findings of ACL-IJCNLP, pages 1489– 1503. find
  42. Takehiro Takayanagi, Masahiro Suzuki, Ryotaro Kobayashi, Hiroki Sakaji, and Kiyoshi Izumi. 2024. Is ChatGPT the future of causal text mining? A comprehensive evaluation and analysis. In Proceedings of IEEE BigData, pages 6651–6660. find
  43. Team GLM. 2024. ChatGLM: A family of large language models from GLM-130B to GLM-4 all tools. arXiv preprint arXiv:2406.12793.
  44. Song Tong, Kai Mao, Zhen Huang, Yukun Zhao, and Kaiping Peng. 2024. Automating psychological hypothesis generation with AI: When large language models meet causal graph. Humanities and Social Sciences Communications, 11(1):1–14. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tang-2026-trace-two-phase-rl-causal/