CompanionHarm (Full Codebook): leaderboard
CompanionHarm (2026) asks LLMs to label each AI utterance in real multi-turn Replika companion conversations with one of 13 socio-emotional harm categories or no harm, using the preceding context; this board uses that prompt plus the category definitions, boundary cases and decision guidance of the annotation codebook. Macro-F1 in percent; higher is better.
Metric: Macro-F1 (%; 14-way single-label classification of annotated Replika AI companion utterances into 13 socio-emotional harm categories or no harm, each with its preceding multi-turn context; temperature 0; the annotation codebook's category definitions, boundary cases and decision guidance added). Source: arxiv.org. Saturation forecast: Around October 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 45.3 | #49 |
| 2 | Claude Opus 4.7 | 43.4 | #79 |
| 3 | Gemini 3.1 Pro (Preview) | 42.8 | #94 |
| 4 | Qwen 3 32B | 37.3 | #960 |
| 5 | Llama 3.1 70B Instruct | 36.7 | #1546 |
| 6 | Qwen 3 8B | 29.7 | #1830 |
| 7 | Llama 3.1 8B Instruct | 13.3 | #4246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=companionharm-full-codebook · How It Works · Data refreshed daily, snapshot 2026-11-13.