LLMTabBench (4-Shot) - Post-Cutoff Datasets: leaderboard
LLMTabBench (Grushina et al., 2026) tests LLMs as in-context binary classifiers of serialized table rows, prompted with the task description and column names and answering by generating a label. This board covers the 8 datasets published after October 2024, one per knowledge domain, with 4 labelled example rows in the prompt, averaged over five draws of the examples; ROC-AUC from 0 to 1 averaged over datasets, higher is better.
Metric: ROC-AUC (0-1; 4 labelled demonstrations; 8 binary classification datasets published after October 2024; generation mode, prompt with column names and task description; mean over datasets and five demonstration draws). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o Mini | 0.71 | #1627 |
| 2 | Qwen 3 14B | 0.71 | #1235 |
| 3 | Qwen 3 8B | 0.7 | #1838 |
| 4 | Qwen 3 1.7B | 0.6 | #5551 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=llmtabbench-4-shot-post-cutoff-datasets · How It Works · Data refreshed daily, snapshot 2026-11-12.