MortarBench: leaderboard
Metric: F1 (%; all 188 cases, mean of 3 trials over MortarBench's 188 synthetic mortgage files (47 real loan-officer questions x 4 mutated profiles, each a ULAD application plus a bank statement); Boolean answers score 1 or 0 and transaction or account lists get set-overlap F1; baseline agent without the authors' CRIT confidence filter, temperature 1 (GPT-5.5 at its fixed default)). Source: arxiv.org. Saturation forecast: Around December 2026. 7 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 80.8 |
| 2 | DeepSeek V4 Pro | 80.7 |
| 3 | GPT-5.5 | 79.9 |
| 4 | Kimi K2.6 | 64 |
| 5 | Claude Sonnet 4.6 | 56.1 |
| 6 | Qwen 3 32B | 51.2 |
Interactive version: theaggregate.ai/benchmark?slug=mortarbench · How It Works · Data refreshed daily, snapshot 2026-09-26.