PSI-Bench (Depression Simulation) - Roleplay-doh: leaderboard

Fidelity of LLM depression-patient simulators on PSI-Bench (2026), run inside the Roleplay-doh simulator framework: simulated counselling sessions are compared with real patient conversations on narrative-emotion progression, emotional valence, lexical diversity, response length and depression markers. Mean alignment score in percent; higher is better.

Metric: Overall fidelity score (%) of the model as a depressed-patient simulator inside the Roleplay-doh framework: mean of five alignment scores with human depressed-patient conversations (narrative-emotion stage progression, valence, lexical diversity, response length, depression markers), simulated sessions with a gpt-oss-20b counsellor; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Llama 3.3 70B Instruct55.72#1385
2Llama 3.1 8B Instruct52.13#4600
3Qwen 2.5 72B Instruct51.71#1117
4Qwen 3 30B A3B 2507 Instruct47.76#1100
5GPT-4.1 Mini45.36#789
6GPT-OSS-20B38.72#1183
7GPT-OSS-120B33.31#704

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=psi-bench-depression-simulation-roleplay-doh · How It Works · Data refreshed daily, snapshot 2026-11-21.