LoFa - BoolQ: leaderboard
LoFa (2026) tests whether a language model keeps a correct answer when an opponent argues for a false claim over up to three rounds with arguments built on one of ten logical fallacies, written by DeepSeek-R1. The score is LFR@3 in percent, the share of questions the model knew and still answers correctly after the third attack, averaged over the ten fallacies; this board uses BoolQ yes or no questions. Higher is better.
Metric: Logical Fallacy Resistance after three attack rounds (LFR@3, %): among the questions a model first answers correctly and whose plain false claim it rejects, the share it still answers correctly after up to three rounds of a DeepSeek-R1-written argument using one of ten logical fallacies, mean over the ten fallacy types, three runs; higher is better; 491 BoolQ yes or no items. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 91.08 | #996 |
| 2 | DeepSeek V3 | 79.17 | #716 |
| 3 | Llama 3.1 405B | 73.61 | #1076 |
| 4 | GPT-3.5 Turbo | 66.01 | #2753 |
| 5 | Llama 3.1 70B | 64.7 | #1490 |
| 6 | Llama 3.1 8B | 62.14 | #4861 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lofa-boolq · How It Works · Data refreshed daily, snapshot 2026-11-13.