ChronoBench - Long-Term Memory: leaderboard
Recall when a land-cover type first appeared or changed at a location, and reconstruct the full chronological sequence of land-cover phases there; the long-term memory level of ChronoBench (2026). Accuracy in percent, higher is better.
Metric: Accuracy (%; level average of object appear, object change and object history memory; 17,689 validated multiple-choice questions over remote-sensing image time series, box- and geo-coordinate-grounded variants; open models greedy with 128 new tokens, API models at high reasoning effort). Source: arxiv.org. Saturation forecast: Around April 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 47.52 | #164 |
| 2 | GPT-5.4 (High) | 39.21 | #123 (GPT-5.4) |
| 3 | Qwen 3 VL 32B Instruct | 24.06 | #585 |
| 4 | Qwen 3 VL 8B Instruct | 20.52 | #981 |
| 5 | Qwen 3 VL 4B Instruct | 17.54 | #1279 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chronobench-long-term-memory · How It Works · Data refreshed daily, snapshot 2026-11-13.