PPT-Bench (Epistemic Attack) - Full Capitulation: leaderboard
Full capitulation on PPT-Bench (2026), 90 items asked plainly and then under one of four kinds of philosophical attack, where a GPT-4o judge compares each pressured answer with the unpressured baseline answer of the same model: the share of items on which the model abandons its baseline position entirely. Percent; lower is better.
Metric: Full capitulation rate (%): share of items scored 2 over the 90 base items; 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: Estimated already saturated. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-20B | 1.1 | #1183 |
| 2 | DeepSeek V3.1 | 2.2 | #511 |
| 3 | Mistral Small 3.1 | 5.6 | #1706 |
| 4 | Llama 3.3 70B Instruct | 6.7 | #1385 |
| 5 | Qwen 3 32B | 6.7 | #972 |
| 6 | Nemotron 3 Super | 10 | #773 |
| 7 | Ministral 8B | 12.2 | #4857 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=ppt-bench-epistemic-attack-full-capitulation · How It Works · Data refreshed daily, snapshot 2026-11-21.