MHGraphBench - Fact Checking: leaderboard

Balanced yes or no questions on whether a triple is supported by the subgraph in MHGraphBench (2026), a knowledge-graph-grounded mental-health benchmark built from a curated PrimeKG subgraph, with letter-only multiple-choice answers. Accuracy in percent; higher is better.

Metric: Accuracy (%) on the 4,000 balanced yes/no triple fact checking items, letter-only answers; the OpenAI API models answer at temperature 0 (at most 120 completion tokens) with strict answer-letter parsing, the open models are scored by forced-choice option-letter log-probabilities; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5 Mini64.35#364
2GPT-4o63.45#751
3GPT-5.2 Instant63.35#304
4GPT-4.163.32#505
5GPT-5.1 Instant62.5#472
6Qwen 2.5 32B Instruct58.4#1247
7Qwen 2.5 7B Instruct53.18#2723
8Mistral 7B Instruct (v0.3)52.25#6750
9Llama3-Med42-8B50.3#5084
10Llama3-OpenBioLLM-8B50.25#6364
11Meditron-7B50#8733
12BioMistral-7B49.9#8332
13DeepSeek-R1-Distill-Qwen-7B49.6#8055
14DeepSeek R1 Distill Qwen 32B48.15#1834
15Llama 3.1 8B Instruct45.32#4600

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

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