Sand-Glass (Prompted Compression) - Chinese-Spanish Translation Density: leaderboard
Sand-Glass (Cui, Yu et al., Bilibili, 2026) tests translation for dubbing under duration limits: 1,000 colloquial Mandarin video-transcript segments, in the prompted-compression setting where the prompt gives the source syllable count and the target band for the target/source syllable ratio. Translation into Spanish targets the 1.0-1.1 band; translation density (BLEU-rho) is back-translation BLEU divided by the achieved syllable ratio, averaged over three runs; higher is better.
Metric: BLEU-rho (back-translation BLEU divided by the target/source syllable ratio; Zh-Es, target band 1.0-1.1; 1,000 Mandarin video-transcript segments, prompt gives the source syllable count and target ratio band, mean of 3 runs). Source: arxiv.org. Saturation forecast: Around May 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Low) | 0.27 | #163 (GPT-5) |
| 2 | Claude Opus 4.1 | 0.26 | #268 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sand-glass-prompted-compression-chinese-spanish-translation-density · How It Works · Data refreshed daily, snapshot 2026-11-12.