EmoHarbor: leaderboard
Personalised emotional-support dialogue in Chinese from EmoHarbor (Ye, Xiang, Zhang and Zong, 2026; Table 3): a simulated user built from one of 100 real-world profiles (GPT-4o thinker and talker agents) talks with the model for up to 15 turns, then a Qwen3-235B user-evaluator rates the conversation on ten dimensions such as problem resolution, mood improvement, empathy, safety and redundancy. The score is the mean of the ten 1-5 Likert ratings; higher is better.
Metric: Mean of 10 Rubric Dimensions (1-5 Likert). Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B (Thinking) | 4.13 | #622 (Qwen 3 235B A22B) |
| 2 | Gemini 2.5 Pro | 4.12 | #262 |
| 3 | Doubao-Seed-1.6 (Thinking) | 4.02 | |
| 4 | DeepSeek R1 | 3.9 | #467 |
| 5 | QwQ-32B | 3.89 | #909 |
| 6 | Claude Sonnet 4 (Thinking) | 3.84 | #397 (Claude Sonnet 4) |
| 7 | GPT-5 | 3.77 | #163 |
| 8 | Claude 3.7 Sonnet | 3.65 | #455 |
| 9 | DeepSeek V3.1 (Thinking) | 3.58 | #500 (DeepSeek V3.1) |
| 10 | GPT-4o (2024-11-20) | 3.5 | #864 |
| 11 | GLM-4.5 (Thinking) | 3.36 | #525 (GLM-4.5) |
| 12 | O3 Mini | 2.97 | #536 |
| 13 | Qwen 3 32B (Thinking) | 2.82 | #960 (Qwen 3 32B) |
| 14 | Qwen 2.5 72B Instruct | 2.63 | #1093 |
| 15 | Qwen 2.5 32B Instruct | 2.46 | #1216 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=emoharbor · How It Works · Data refreshed daily, snapshot 2026-11-12.