A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
ONERULER is a multilingual benchmark that adapts the English-only RULER to evaluate long-context language models across 26 languages.
Unknown
MR3 is a multilingual, rubric-agnostic reward reasoning model trained to produce more accurate reasoning and evaluations in target languages.
Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca, et al.
Tiny Aya L2-Thinker achieves over 93% in-language reasoning across 60 languages via data mixing in SFT, without reasoning supervision in every target language.
Parinthapat Pengpun, Simran Khanuja, Graham Neubig
Introduces knowledge-graph structural rarity metrics for multimodal entity linking, a training-free reasoning+retrieval framework, and MERLIN-Rare, improving SOTA by 6.9% overall and up to 23.3% on rare entities.
Unknown
This paper introduces M-RewardBench, a benchmark for evaluating reward models in multilingual settings, revealing that current reward models exhibit limited robustness across languages.
Unknown
Proposes MRRE, a training-free inference-time method using representation engineering to enhance multilingual reasoning in LLMs and LVLMs without additional training data.
Melvin Johnson, Mike Schuster, Quoc V. Le, et al.
Proposes a single multilingual NMT model using a target-language token for zero-shot translation across multiple language pairs.
Unknown
Creates a 2.8M-pair multilingual parallel corpus for scientific/technical MT in eight low-resource Indian languages from NPTEL lecture transcriptions.
Unknown
Leverages Gemini LLM to generate generalizable multilingual text and code embeddings via synthetic data and filtering.
Unknown
SigLIP 2 improves multilingual vision-language encoders with captioning pretraining, self-supervised losses, and online data curation, offering native aspect ratio preservation.
Unknown
mmE5 is a multimodal multilingual embedding model trained on synthetic data via a framework that ensures broad task/language coverage, robust cross-modal alignment, and high fidelity through self-evaluation and refinement.
Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, et al.
Jina Embeddings v3 is a 570M parameter multilingual embedding model supporting 8192-token contexts, task-specific LoRA adapters, and flexible dimension reduction via Matryoshka learning.