TriBench-Ko - Hallucination: leaderboard

The hallucination risk dimension of TriBench-Ko (2026), a Korean benchmark of attorney-reviewed yes/no items built from court decisions, where false statements carry fabricated case citations, statutory provisions or facts. The score is macro-F1 on a 0-100 scale over the atomic binary judgments, zero-shot at temperature 0; higher is better.

Metric: Macro-F1 (0-100) on the TriBench-Ko Hallucination risk items (fabricated case citations, statutory provisions or facts): yes/no verification of statements about Korean court decisions, computed over the atomic binary judgments, zero-shot, temperature 0, at most 64 output tokens; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.493.3#138
2Qwen 3.5 9B88.9#838
3GPT-5.4 Mini88.3#409
4GPT-4o84.3#754
5Midm-2.0-Base-Instruct83.3#1998
6Phi-482.6#1951
7Ministral-3-8B-Instruct-251277.2#1923
8Qwen 3 8B70.6#1890
9Gemma 3 12B (IT)68.2#1907
10EXAONE 3.5 7.8B Instruct64.8#3453
11kanana-1.5-8B-instruct-250556.6#3297
12Llama 3.1 8B Instruct37.8#4655
13A.X-3.1-Light33.1

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=tribench-ko-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-22.