Note Multilingual synthetic corpus for translation. Over 1 trillion tokens of parallel text in English and 500+ languages by translating data from FineWeb2 into English using Gemma3 27B

Note General overview of popular inference engines performance across different workload scenarions and GPU platforms.

Note This paper is about proper and systematic evaluation to appropriately adjust claims, revealing hidden fragilities and failure modes—such as reliance on masking artifacts, shortcut learning, or flawed reasoning—and highlighting how models may perform well on specific benchmarks while failing to generalize.

Note A survey on multilingual LLMs with design recommendations for tokenization, sampling, architectures, and evaluation to support multilingual LMs.

google/translategemma-27b-it

Image-Text-to-Text • Updated 7 days ago • 35.3k • 292

Note Multimodal translation model for around 55 langauges and optimized from Gemma family models for MT task.

Note MMFMs still exhibit surprising deficiencies in spatial intelligence. In this paper, they explore scaling up MMFMs to cultivate spatial intelligence within the SenseNova-SI family, built upon established multimodal foundations including visual understanding models (Qwen3-VL and InternVL3) and unified understanding and generation models.

Note Most diagnostic models still process different modalities in isolation. This limits their ability to capture early, cross-modal disease signatures. This work introduces a MMFM built on a transformer architecture that integrates heterogeneous clinical data through modality-specific encoders and cross-modal attention.

Note Geo-foundation models show promise for flood mapping, offering modest but consistent improvements over traditional approaches while reducing data and computational requirements

Note Split memory and reasoning in Large Language Models.

Note Parallel image translation dataset spanning 840k images across 14 languages. The images are sourced from 8 languages from 28 categories. GPT-4 is used to translate the OCR-regognized text, verified using similarity to Google Translate and random human verification.

Note Work on Benchmarking LLMs in Medical Domain; Introduces a simulated clinical environment in which multimodal agents interact with patients, gather incomplete information, use tools, and reason over time.

Note Interesting recent work on a multimodal LLM for scientific discovery incorporating a wide variety of scientific information and showing consistently high results on benchmarks.

Note An evaluation framework that efficiently tracks datasets, model variants, aggregation choices, and downstream tasks while remaining fast, reproducible, and scalable. keywords: evaluation, medical imaging, reproducibility