FinLifeBench - Life-Event History - Exact History Match: 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. Share of checkpoints whose predicted event-anchor history matches the gold history exactly; percent, higher is better.
Metric: Checkpoints with the exact history (%). Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (Low) | 11.3 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | Claude Sonnet 4.6 (Low) | 8.3 | #94 (Claude Sonnet 4.6) |
| 3 | GPT-5.6 Sol (Low) | 6.8 | #22 (GPT-5.6 Sol) |
| 4 | GPT-5.6 Luna (Low) | 6.5 | #79 (GPT-5.6 Luna) |
| 5 | GPT-5.6 Terra (Low) | 6.5 | #40 (GPT-5.6 Terra) |
| 6 | Gemini 3.5 Flash (Low) | 5.3 | #66 (Gemini 3.5 Flash) |
| 7 | Llama 4 Maverick | 5 | #486 |
| 8 | Claude Opus 4.8 (Low) | 5 | #39 (Claude Opus 4.8) |
| 9 | GPT-OSS-120B (Low) | 1.8 | #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-life-event-history-exact-history-match · How It Works · Data refreshed daily, snapshot 2026-11-05.