CompanionBench - English - IRT Theta: leaderboard
Overall IRT theta for English-condition emotional-companionship dialogues in CompanionBench (arXiv 2608.02046v2, Table 3): 28 agents in multi-session conversations with a simulator trained on de-identified real data, rated on ten capabilities by a cross-family judge panel with judge severity separated by item response theory. Rating-scale units (1-10); higher is better; rankable rather than absolute.
Metric: IRT theta (judge-severity-corrected, 1-10 scale). Source: arxiv.org. Saturation forecast: Around February 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 8.27 | #33 |
| 2 | Claude Opus 4.8 | 8.24 | #39 |
| 3 | GPT-5.4 | 8.21 | #83 |
| 4 | DeepSeek V4 Pro | 8.06 | #112 |
| 5 | Claude Sonnet 4.6 | 8.04 | #94 |
| 6 | GLM-5 | 7.98 | #149 |
| 7 | Kimi K2.6 | 7.95 | #113 |
| 8 | GLM-5.1 | 7.94 | #109 |
| 9 | GPT-5.1 | 7.93 | #145 |
| 10 | Qwen 3.7 Max | 7.93 | #85 |
| 11 | Kimi K2.5 | 7.88 | #152 |
| 12 | GLM-5.2 | 7.86 | #80 |
| 13 | DeepSeek V4 Flash | 7.82 | #155 |
| 14 | Qwen 3.5 397B A17B | 7.77 | #154 |
| 15 | Gemma 4 31B (IT) | 7.77 | #178 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=companionbench-english-irt-theta · How It Works · Data refreshed daily, snapshot 2026-11-05.