FinED-Bench - Financial Reasoning Errors: leaderboard
FinED-Bench (2026) tests whether LLMs can find and classify errors injected into real 2025 Chinese financial documents from nine scenarios. This board scores the reasoning errors (conflicting expressions, time contradictions, numerical inconsistencies, calculation errors and clause conflicts), which need document-level reasoning. F1 in percent; higher is better.
Metric: F1 (%; financial reasoning errors: conflicting expression, time contradiction, numerical inconsistency, calculation error, clause conflict; a predicted error counts as correct when the extracted sentence matches or contains the annotated sentence and the error type is classified correctly; the paper's error-detection prompt over each document, reasoning models with thinking on unless marked no thinking). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 38 | #741 |
| 2 | Qwen 3 14B | 29.1 | #1235 |
| 3 | Qwen 3 8B | 27.96 | #1838 |
| 4 | Qwen 3 14B (Non-reasoning) | 16.63 | #1235 (Qwen 3 14B) |
| 5 | DeepSeek R1 0528 Qwen3 8B | 13.71 | #2101 |
| 6 | GPT-4o Mini | 10.98 | #1627 |
| 7 | Qwen 3 8B (Non-reasoning) | 8.6 | #1838 (Qwen 3 8B) |
| 8 | Qwen 2.5 7B Instruct | 3.79 | #2583 |
| 9 | GPT-3.5 Turbo | 2.41 | #2773 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fined-bench-financial-reasoning-errors · How It Works · Data refreshed daily, snapshot 2026-11-12.