G-IdiomAlign - Multiple-Choice Equivalence - Other Targets: leaderboard
G-IdiomAlign (Ye et al., 2026) multiple-choice idiom equivalence over the 14 directions into the other seven languages. Given a source idiom, the model picks the meaning-equivalent target-language idiom among four options whose three distractors are a literal translation, a lexical look-alike and a contextually associated idiom; accuracy in percent (higher is better), macro-averaged over the directions.
Metric: Accuracy (%; source idiom to its meaning-equivalent target-language idiom among four options, the other three typed distractors (literal translation, lexical cue, contextual association), greedy decoding; macro-average over the 14 directions into the other seven languages). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 62.82 | #262 |
| 2 | DeepSeek V3.2 (Thinking) | 53.92 | #403 (DeepSeek V3.2) |
| 3 | DeepSeek V3.2 (Non-reasoning) | 47.39 | #403 (DeepSeek V3.2) |
| 4 | Claude Haiku 4.5 | 44.03 | #568 |
| 5 | Qwen 3 8B (Thinking) | 31.29 | #1838 (Qwen 3 8B) |
| 6 | Ministral-8B-Instruct-2410 | 27.78 | #4720 |
| 7 | Qwen 3 8B (Non-reasoning) | 24.02 | #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-multiple-choice-equivalence-other-targets · How It Works · Data refreshed daily, snapshot 2026-11-12.