Benchmark RadarAI BENCHMARK PROFILE
G-IdiomAlign
G-IdiomAlign is a gloss-pivoted benchmark for cross-lingual idiom alignment, with protocols for multiple-choice equivalence and gloss-contrastive generation.
- Released
- 2026-06-17
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Idiom translation is challenging due to non-compositionality. This benchmark could support evaluation of multilingual models, but its public availability and scoring contract are not fully clear.
Motivation
Idioms are difficult to transfer across languages due to their non-compositionality and weak surface-form grounding, making literal mappings unreliable.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.