FinED-Bench: leaderboard
FinED-Bench (2026) tests whether LLMs can find and classify errors injected into real 2025 Chinese financial documents from nine scenarios. F1 in percent over all errors: a detection counts when the extracted sentence contains the erroneous one and the error type is classified correctly; higher is better. The errors span general knowledge, financial domain knowledge and financial reasoning.
Metric: F1 (%; harmonic mean of pooled precision and recall over all injected errors in 973 Chinese financial documents from 2025, nine real-world scenarios; 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 | 48.34 | #741 |
| 2 | Qwen 3 14B | 43.15 | #1235 |
| 3 | Qwen 3 8B | 39.99 | #1838 |
| 4 | Qwen 3 14B (Non-reasoning) | 27.19 | #1235 (Qwen 3 14B) |
| 5 | DeepSeek R1 0528 Qwen3 8B | 23.31 | #2101 |
| 6 | GPT-4o Mini | 18.9 | #1627 |
| 7 | Qwen 3 8B (Non-reasoning) | 17.44 | #1838 (Qwen 3 8B) |
| 8 | Qwen 2.5 7B Instruct | 9.85 | #2583 |
| 9 | GPT-3.5 Turbo | 7.17 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.