TriBench-Ko - Prompt Sensitivity: leaderboard

TriBench-Ko’s prompt-sensitivity dimension tests Korean legal yes/no judgments under leading or reframed instructions that can suggest an incorrect conclusion. This board reports the paper’s macro-F1 over binary judgments on a 0–100 scale; higher is better.

Metric: Macro-F1 (0-100) on the TriBench-Ko Prompt Sensitivity risk items (prompt-framing robustness; detailed variant/repeat mapping is inconsistent between the main and appendix tables): 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-5.492.3#135
2Qwen 3.5 9B87.1#828
3GPT-5.4 Mini87.1#406
4GPT-4o79.7#751
5Midm-2.0-Base-Instruct78.2#1988
6Phi-475.7#1931
7Ministral-3-8B-Instruct-251271.1#1907
8Gemma 3 12B (IT)60.5#1887
9Qwen 3 8B57.5#1873
10EXAONE 3.5 7.8B Instruct57.2#3410
11kanana-1.5-8B-instruct-250545.5#3315
12Llama 3.1 8B Instruct35#4600
13A.X-3.1-Light33.1

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

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