T2S-Bench - Node Extraction: leaderboard
Metric: Node score (%) on T2S-Bench-E2E (87 human-verified text-structure pairs): given the text and every link, the model labels the nodes, scored by mean semantic similarity to the reference node labels, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 45 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 58.09 | #145 |
| 2 | Claude Sonnet 4.5 | 55.97 | #138 |
| 3 | Claude Sonnet 4 (20250514) | 54.11 | #211 |
| 4 | GPT-5.2 | 50.57 | #105 |
| 5 | Qwen 3 235B A22B 2507 Instruct | 49.39 | #291 |
| 6 | DeepSeek R1 0528 | 49.24 | #217 |
| 7 | DeepSeek V3 (0324) | 47.32 | #332 |
| 8 | Claude Haiku 4.5 (20251001) | 47.06 | #251 |
| 9 | DeepSeek V3.2 | 46.98 | #198 |
| 10 | Gemini 2.5 Flash | 46.9 | #237 |
| 11 | Qwen 3 14B | 46.85 | #524 |
| 12 | DeepSeek V3.1 | 46.59 | #260 |
| 13 | Qwen 3 235B A22B 2507 (Thinking) | 45.97 | #253 (Qwen 3 235B A22B 2507) |
| 14 | Mistral Small 3.2 | 45.74 | #587 |
| 15 | Mistral Large 2 (Nov) Instruct (2411) | 45.71 | #458 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=t2s-bench-node-extraction · How It Works · Data refreshed daily, snapshot 2026-10-11.