Multilingual Information Retrieval on Belebele (nDCG@20 & Recall@20 per Language) benchmark leaderboard · SOTA2 Research

Multilingual Information Retrieval on Belebele (nDCG@20 & Recall@20 per Language)

0.681Arabic nDCG@20LangSAE Editing

Evaluation Results

Method	Links
LangSAE Editing2026.01		0.681	0.6719	0.7635	0.7461	0.75	0.7288	0.4483	0.4103	0.2766	0.196	0.6947	0.6821	0.7304	0.7099	0.7485	0.7276	0.7119	0.7008	0.7292	0.7068	0.6534	0.628
SAE Reconstructed2026.01		0.493	0.4844	0.737	0.7298	0.6888	0.6633	0.3884	0.3393	0.2749	0.1804	0.3461	0.2731	0.6918	0.6656	0.6603	0.6308	0.5171	0.4836	0.6165	0.5712	0.5414	0.5022
multilingual-e5-large2026.01		0.4853	0.475	0.7322	0.7246	0.6857	0.66	0.3836	0.3329	0.2738	0.1794	0.3397	0.2649	0.6847	0.658	0.6522	0.6227	0.5116	0.4769	0.6107	0.5634	0.5359	0.4958
All-but-the-Top2026.01		0.4194	0.3909	0.7087	0.7028	0.6692	0.6458	0.3727	0.3209	0.2631	0.1674	0.4175	0.3728	0.6842	0.6569	0.6416	0.6134	0.5503	0.5344	0.6642	0.6229	0.5391	0.5028