FinReportBench: leaderboard
FinReportBench (Tang et al., 2026) evaluates institution-grade financial research report generation from a public query, a reconstructed research trajectory and a hidden source packet, over 244 Chinese and English tasks. Rendered reports are graded by a GPT-5.6 Luna judge against 35 expert-derived rubric items in three ordered layers (deliverability, report identity and institutional completeness), combined into a hierarchical score from 0 to 100 that is zero when a deliverability check fails; higher is better.
Metric: Provisional hierarchical score (0-100): mean over the 244 bilingual report-generation tasks of S = 100 (g1 + (S1/100) g2) / (d1 + d2), where g1, g2 are the weighted rubric points earned on the 4 report-identity (G1) and 27 institutional-completeness (G2) items, d1, d2 the points available and S1 the G1 percentage, and S = 0 when any of the 4 deliverability (G0) preflight checks fails; pass, partial and fail items earn 1, 0.5 and 0; GPT-5.6 Luna at medium reasoning effort judges extracted text and up to four rendered pages against the frozen 35-item rubric; no-skill condition. Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniMax-M2.7 | 22.6 | #518 |
| 2 | GLM-5.2 | 21.8 | #138 |
| 3 | MiniMax-M2.5 | 21.1 | #582 |
| 4 | DeepSeek V4 Flash | 20.2 | #258 |
| 5 | Qwen 3.7 Max | 19.1 | #145 |
| 6 | DeepSeek V4 Pro | 18.8 | #188 |
| 7 | MiniMax-M3 | 18.6 | #269 |
| 8 | Kimi K2.6 | 17.7 | #178 |
| 9 | Qwen 3.6 27B | 15.1 | #341 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finreportbench · How It Works · Data refreshed daily, snapshot 2026-11-13.