ConflictQA - Text Conflict (Non-Complementary): leaderboard

ConflictQA (2026) asks questions whose evidence, knowledge-graph triples plus Wikipedia text, contains a contradiction injected into one source; the model is told the sources may conflict and answers zero-shot, scored by exact match of the answer set and averaged over two evidence orders. This board covers single-hop questions with falsified text beside correct triples. Exact match in percent; higher is better.

Metric: Exact match (%) on 802 single-hop (non-complementary) questions when the provided evidence contains conflicting text passages (TextConf); zero-shot with a conflict-aware prompt, the mean of two evidence orders; exact match of the generated answer set; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1DeepSeek V3.2Non-reasoning53.43#419
2O3 Mini52.56#546
3Qwen 3 8B52.44#1873
4GPT-5.151.88#240
5Llama 3.1 70B Instruct48.96#1571
6ConflictQA Qwen3-30B-A3B-Thinking (checkpoint unspecified)48.69
7open-mistral-7B47.07#7710
8GPT-4o46.57#751
9GPT-3.5 Turbo (0125)43.77#3550
10Llama 3.1 8B Instruct41.15#4600
11Mistral Large 332.17#832

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

Interactive version: theaggregate.ai/benchmark?slug=conflictqa-text-conflict-non-complementary · How It Works · Data refreshed daily, snapshot 2026-11-21.