SciTaRC (Program-of-Thought): leaderboard
Metric: LLM-judge accuracy (%) with program-of-thought prompting (the model writes Python that is executed to produce the answer) on the 371 expert-authored SciTaRC questions over tables from recent arXiv papers (LaTeX table sources as input), zero-shot; an answer counts only when a Llama-3.3-70B-Instruct judge gives full credit against the gold answer (partial matches count as wrong); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 45 | #138 |
| 2 | DeepSeek V3.2 (Non-reasoning) | 36.1 | #198 (DeepSeek V3.2) |
| 3 | Qwen 3 Coder 30B A3B Instruct | 19.7 | #399 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=scitarc-program-of-thought · How It Works · Data refreshed daily, snapshot 2026-10-11.