FinLifeBench - Financial State - Evidence Recall: leaderboard
FinLifeBench (arXiv 2609.01198v2, Table 3): 11 models reconstruct customer records from up to 300 cumulative Korean banking sessions per persona (20 personas, 400 checkpoints per task), full context, low reasoning setting where available. Recall of gold evidence session identifiers for changed state paths; surplus citations are not penalized; percent, higher is better.
Metric: Recall of gold evidence sessions (%). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Low) | 88.1 | #22 (GPT-5.6 Sol) |
| 2 | Claude Sonnet 4.6 (Low) | 86.6 | #94 (Claude Sonnet 4.6) |
| 3 | GPT-5.6 Terra (Low) | 85.5 | #40 (GPT-5.6 Terra) |
| 4 | Gemini 3.1 Pro (Preview) (Low) | 85.1 | #64 (Gemini 3.1 Pro (Preview)) |
| 5 | Claude Opus 4.8 (Low) | 84.8 | #39 (Claude Opus 4.8) |
| 6 | GPT-5.6 Luna (Low) | 78.6 | #79 (GPT-5.6 Luna) |
| 7 | Gemini 3.5 Flash (Low) | 77.3 | #66 (Gemini 3.5 Flash) |
| 8 | Llama 4 Maverick | 39.3 | #486 |
| 9 | GPT-OSS-120B (Low) | 3.6 | #346 (GPT-OSS-120B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=finlifebench-financial-state-evidence-recall · How It Works · Data refreshed daily, snapshot 2026-11-05.