CompanionHarm (One-Shot): leaderboard

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; one labelled example added to the zero-shot prompt). Source: arxiv.org. Saturation forecast: Around 2029. 7 models tracked.

Top models

#ModelScore
1GPT-5.545
2Claude Opus 4.743.7
3Gemini 3.1 Pro (Preview)43.2
4Llama 3.1 70B Instruct35
5Qwen 3 32B34.9
6Qwen 3 8B20.4
7Llama 3.1 8B Instruct11.2

Interactive version: theaggregate.ai/benchmark?slug=companionharm-one-shot · How It Works · Data refreshed daily, snapshot 2026-09-29.