PPT-Bench (Epistemic Attack): leaderboard

Capitulation under philosophical pressure on PPT-Bench (2026): 90 factual, ethical, safety-critical, social and self-description items are asked plainly and then under one of four kinds of philosophical attack, and a GPT-4o judge compares each pressured answer with the unpressured baseline answer of the same model. Share of items with partial or full capitulation, in percent; lower is better.

Metric: Binary capitulation rate (%): share of items scored 1 or 2 over the 90 base items (four philosophical pressure types, five domains); a gpt-4o pairwise judge compares each single-turn pressured answer (L1) with the model's own unpressured baseline answer (L0) and scores 0 (held), 1 (partial) or 2 (full capitulation); temperature 0, one generation per item; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Nemotron 3 Super23.3#815
2GPT-OSS-20BLow28.9#1188
3DeepSeek V3.145.6#511
4Mistral Small 3.155.6#1734
5Llama 3.3 70B Instruct61.1#1394
6Qwen 3 32B68.9#983
7Ministral 8B86.7#4866

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ppt-bench-epistemic-attack · How It Works · Data refreshed daily, snapshot 2026-11-22.