Benchmark explainer
LLM capability on extracting structured tables from unstructured documents (PDF, HTML, mixed layouts).
DTBench tests LLM capability on extracting structured tables from unstructured documents — PDF, HTML and mixed layouts. Developed by Vals AI.
Discrimination
95.9/100Does it still separate models?
Saturation headroom
98/100How far from ceiling / clustered at the top?
Contamination resistance
86/100Public vs held-out; training-leak risk.
Harness comparability
100/100Is it apples-to-apples, or mixed / vendor-optimized?
Freshness
100/100How old is the benchmark?
Contamination history: Synthetic, published 2026; some templates drawn from public document datasets (notable: authors flag specific overlap risks in §6).
Limitations: Synthetic generation may not match real-world document noise/distribution; LLM extraction accuracy is the only published metric — interpretability scoring absent.
| Model | Score | Evidence |
|---|---|---|
| Claude Fable 5.2 | 98.67 | unverified |
| Claude Opus 5.5 | 98.22 | unverified |
| Claude Fable 5 | 97.33 | unverified |
| Claude Opus 5 | 96.45 | unverified |
| Gemini 4 Argon | 96 | unverified |
DTBench: DTBench tests LLM capability on extracting structured tables from unstructured documents — PDF, HTML and mixed layouts. Developed by Vals AI.
Claude Fable 5.2 leads DTBench at 98.67. The full leaderboard above lists every recorded measurement, not just the headline number.
150 models have recorded scores on DTBench, spanning a score spread of 95.92.
tensor.news grades DTBench A for integrity (score 96.1/100), ranking #5 of 62 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
Yes — top scores are clustering near the ceiling (saturation ratio 0.02), so DTBench no longer separates leading models well. Treat small gaps at the top with caution.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on DTBench are reasonably apples-to-apples.
Every DTBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2602.13812.