HarmThoughts (Few-shot): leaderboard

Sentence-level classification of harmful behaviors inside jailbroken reasoning traces on HarmThoughts (2026), using a 16-category taxonomy of how harm is suppressed, rationalized, executed or concealed, with few-shot human-annotated examples. Macro-F1 times 100 over the parseable predictions; higher is better.

Metric: Macro-F1 (x100) of 16-way sentence-level behavior classification (harm-propagation, safety-preserving and neutral behaviors) on jailbroken reasoning traces of HarmThoughts, the model prompted with the taxonomy and the full trace, few-shot prompting with human-annotated examples; scored over the parseable predictions; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro56.21#276
2Gemini 2.0 Flash52.52#723
3GPT-4o40.85#751
4Llama 3.3 70B40.41#1514

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

Interactive version: theaggregate.ai/benchmark?slug=harmthoughts-few-shot · How It Works · Data refreshed daily, snapshot 2026-11-21.