LongDS v1.1 Full: leaderboard
LongDS (Xu et al., Zhejiang University and Ant Group, EMNLP 2026) tests long-horizon agentic data analysis on multi-turn tasks built from Kaggle notebooks and datasets. This board is the authors' audited v1.1 release on all 68 tasks (2,225 turns), where each model runs under one vendor's agentic coding harness (Claude Code, Codex, Kimi Code or Qoder). Later turns update, perturb, roll back or combine analytical states set up earlier; a judge scores each turn against executable reference code, and the score is turn accuracy macro-averaged over tasks, in percent.
Metric: Accuracy (%; judged turn accuracy macro-averaged over 68 tasks), each model under one agentic coding harness. Source: zjunlp.github.io. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 64.42 | #29 |
| 2 | Kimi K3 | 58.59 | #53 |
| 3 | Qwen 3.8 Max (0902) | 56.88 | #49 |
| 4 | GLM-5.2 | 54.25 | #137 |
| 5 | DeepSeek V4 Pro | 40.96 | #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-v1-1-full · How It Works · Data refreshed daily, snapshot 2026-11-12.