HalluScore - Adversarial: leaderboard

HalluScore (2026) is an Arabic question-answering benchmark of 827 hallucination-prone Modern Standard Arabic questions, with human annotation of model answers. This board is the factual hallucination rate, the percent of answers that state an unsupported fact with confidence on the 360 adversarial questions with misleading phrasing, zero-shot at temperature 0 (expressed uncertainty or refusal does not count); lower is better.

Metric: Factual hallucination rate (%) on the 360 adversarial HalluScore questions (misleading or carefully crafted phrasing), zero-shot short-answer Arabic question answering at temperature 0 with no retrieval, responses annotated by humans; an answer that states an unsupported fact with confidence counts, an expressed uncertainty or refusal does not; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-531.11#178
2Claude Opus 434.72#326
3Claude Sonnet 4.538.89#261
4GPT-4o44.72#751
5O4 Mini51.11#368
6Qwen3-235B-A22B-Instruct-2507-tput53.61#548
7Grok 4 FastReasoning55.56#447
8Qwen 3 Next 80B A3B Instruct58.61#778
9DeepSeek R160.83#482
10DeepSeek V362.5#729
11Grok 465.28#319
12Llama 4 Maverick Instruct FP870.56#1023
13ALLaM-7B-Instruct-preview75#2001
14Fanar-1-9B80.83#1913
15O381.67#237

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

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