LongDS v1 - Geoscience: leaderboard
LongDS (Xu et al., Zhejiang University and Ant Group, EMNLP 2026) long-horizon data-analysis tasks built from Kaggle notebooks, in which later turns update, perturb, roll back or combine analytical states set up many turns earlier. A ReAct agent writes Python in a persistent Jupyter kernel (DSGym, up to 40 steps per turn) and a DeepSeek-V4-Pro judge checks each turn against executable reference code. The score is per-task turn accuracy macro-averaged over the 19 geoscience tasks of the original v1 release, in percent; higher is better.
Metric: Accuracy (%; macro-average over 19 geoscience tasks of per-task turn accuracy; DSGym ReAct agent in a persistent Jupyter kernel, up to 40 steps per turn, DeepSeek-V4-Pro judge). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 42.2 | #91 |
| 2 | Claude Sonnet 4.6 | 31.92 | #158 |
| 3 | GPT-5.4 | 28.9 | #121 |
| 4 | Kimi K2.6 | 28.83 | #173 |
| 5 | DeepSeek V4 Pro | 16.6 | #181 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longds-v1-geoscience · How It Works · Data refreshed daily, snapshot 2026-11-12.