Jiahe Liu

Research profile

Research summary

Jiahe Liu is a researcher at Monash University (Australia) whose work, profiled here from OpenAlex bibliographic data, sits at the intersection of natural language processing, mental health conversational AI, and LLM safety evaluation. The snapshot is early-stage: six works in the trailing three-year window, all dated 2026, no citation indicators yet (h-index and i10-index recorded as 0), and 23 distinct coauthors. Topic counts are strongly concentrated — 'Ethics and Social Impacts of AI' (4), with AI in Healthcare and Education, Explainable AI, and Speech and dialogue systems at 2 each, and Schizophrenia research at 1. Two Evolve-reviewed 2026 arXiv preprints define the visible record. StratCBT ('Steering the Compass') contributes the first large strategy-aligned cognitive behavioral therapy counseling dataset of roughly 9.7K sessions, generated synthetically with LLMs and judged by GPT-4o — valuable, though its realism claims warrant caution. SDARE-Bench is presented as the first benchmark probing how large language models detect and respond to stigma in dyadic and group dialogue; its group-pressure stress tests are notable, while synthetic self-labelled data and moderate classifier F1 temper the conclusions. Across both papers runs a consistent signature: constructing synthetic dialogue corpora and evaluation benchmarks for psychologically sensitive conversations, with this researcher positioned as a middle author within larger teams.

Research keywords

Highlights

• StratCBT: first large strategy-aligned CBT counseling dataset

  The 2026 arXiv preprint 'Steering the Compass' contributes a dataset of roughly 9.7K sessions aligning dynamic psychological counseling conversations with cognitive behavioral therapy strategies. Evolve's verdict: valuable as the first resource of its kind, but fully LLM-synthetic and GPT-4o-judged, so realism claims need caution.

• SDARE-Bench: first LLM benchmark for stigma in dyadic and group dialogue

  This 2026 benchmark evaluates how large language models detect and respond to conversational stigma across dyadic and multi-party dialogue. Evolve notes that the group-pressure stress test results are striking, but synthetic self-labelled data and moderate classifier F1 weaken the persuasiveness of its conclusions.

• A consistent methodology: LLM-synthesized corpora, LLM-judged evaluation

  Both reviewed 2026 papers share a methodological signature — synthetically generated sensitive-dialogue corpora evaluated with LLM judges. This enables rapid benchmark construction for psychologically sensitive topics, and both Evolve verdicts flag the same trade-off between scale and realism.

Timeline

1. 2026Released 'Steering the Compass' on arXiv, introducing StratCBT, a large strategy-aligned cognitive behavioral therapy counseling dataset of roughly 9.7K sessions.
2. 2026Released SDARE-Bench on arXiv, the first benchmark evaluating large language models on stigma detection and response in dyadic and group dialogue.
3. 2026OpenAlex snapshot records 6 works in the recent window, all dated 2026, with 23 distinct coauthors — the most concentrated year in the visible record.

Papers in the Evolve catalog

• 2026-09-18

  Steering the Compass: Aligning Dynamic Psychological Counseling Conversations with Cognitive Behavioral Therapy Strategies

• 2026-09-01

  SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue

Frequent co-authors