G-IdiomAlign - Gloss-Contrastive Generation (With Gloss): leaderboard
G-IdiomAlign (Ye et al., 2026): Given a source idiom and its English gloss, the model writes one meaning-equivalent idiom in the target language. A prediction counts as correct when its Qwen3-Embedding-8B cosine similarity to the reference idiom is at least 0.80, a semantic proxy that can miss valid synonymous idioms; accuracy in percent (higher is better), macro-averaged over 72 translation directions.
Metric: Acc@0.80 (%; the model writes one target-language idiom for a source idiom, correct when its Qwen3-Embedding-8B cosine similarity to the reference idiom is at least 0.80, greedy decoding, macro-average over 72 translation directions; source idiom plus its English gloss). Source: arxiv.org. Saturation forecast: Around 2030. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 (Thinking) | 29.63 | #403 (DeepSeek V3.2) |
| 2 | Gemini 2.5 Pro | 27.11 | #262 |
| 3 | Claude Haiku 4.5 | 25.66 | #568 |
| 4 | DeepSeek V3.2 (Non-reasoning) | 23.57 | #403 (DeepSeek V3.2) |
| 5 | Ministral-8B-Instruct-2410 | 21.22 | #4720 |
| 6 | Qwen 3 8B (Thinking) | 20.86 | #1838 (Qwen 3 8B) |
| 7 | Qwen 3 8B (Non-reasoning) | 18.8 | #1838 (Qwen 3 8B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=g-idiomalign-gloss-contrastive-generation-with-gloss · How It Works · Data refreshed daily, snapshot 2026-11-12.