JFinBench: leaderboard
Models answer Japanese financial examination questions. Untooled results, reasoning settings and controlled agent harnesses stay separate; the best result across different harnesses is not treated as a common setup.
Source: arxiv.org.
Benchmark results
Results are published separately for each benchmark or evaluation setting.
- JFinBench - No Tools - CPA
- JFinBench - No Tools - Financial Counter
- JFinBench - No Tools - FP
- JFinBench - No Tools - Nissho
- JFinBench - No Tools - Overall
- JFinBench - No Tools - Tax Accountant
- JFinBench - No Tools - Zensho
- JFinBench - Tools - Claude Code 2.1.306
- JFinBench - Tools - Codex 0.163.0
- JFinBench - Tools - Gemini CLI 0.65.0
- JFinBench - Tools - Kimi Code 2.3.0
- JFinBench - Tools - MiMo Code 0.1.15
- JFinBench - Tools - mini-SWE-agent 2.4.6
- JFinBench - Tools - Mistral Vibe 3.0.3
- JFinBench - Tools - Qwen Code 0.25.3
- JFinBench - Tools - Terminus-2 Harbor 0.24.0
Interactive version: theaggregate.ai/benchmark?slug=jfinbench · How It Works · Data refreshed daily, snapshot 2026-11-17.