MBZUAI ·
A Mixture of Experts (MoE) layer is a sparsely activated deep learning layer. It uses a router network to direct each token to one of the experts. Yuanzhi Li, an assistant professor at CMU and affiliated faculty at MBZUAI, researches deep learning theory and NLP. Why it matters: This highlights MBZUAI's engagement with cutting-edge deep learning research, specifically in efficient model design.
arXiv ·
Arabic-DeepSeek-R1 is an application-driven, open-source Arabic Large Language Model (LLM) that has achieved a new state-of-the-art (SOTA) across the Open Arabic LLM Leaderboard (OALL). The model utilizes a sparse Mixture-of-Experts (MoE) backbone and a four-phase Chain-of-Thought (CoT) distillation scheme, which incorporates Arabic-specific linguistic verification and regional ethical norms. It records the highest average score on the OALL suite and outperforms proprietary frontier systems like GPT-5.1 on a majority of benchmarks evaluating comprehensive Arabic language-specific tasks. Why it matters: This work offers a validated and cost-effective framework for developing high-performing, culturally-grounded AI for under-represented languages, addressing the digital equity gap.
arXiv ·
The paper introduces Sparse-Quantized Representation (SpQR), a new compression format and quantization technique for large language models (LLMs). SpQR identifies outlier weights and stores them in higher precision while compressing the remaining weights to 3-4 bits. The method achieves less than 1% accuracy loss in perplexity for LLaMA and Falcon LLMs and enables a 33B parameter LLM to run on a single 24GB consumer GPU. Why it matters: This enables near-lossless compression of LLMs, making powerful models accessible on resource-constrained devices and accelerating inference without significant accuracy degradation.
arXiv ·
The paper introduces Yet another Policy Optimization (YaPO), a reference-free method for learning sparse steering vectors in the latent space of a Sparse Autoencoder (SAE) to steer LLMs. By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Experiments show YaPO converges faster, achieves stronger performance, exhibits improved training stability and preserves general knowledge compared to dense steering baselines.
arXiv ·
The paper introduces Duet, a hybrid neural relation understanding method for cardinality estimation. Duet addresses limitations of existing learned methods, such as high costs and scalability issues, by incorporating predicate information into an autoregressive model. Experiments demonstrate Duet's efficiency, accuracy, and scalability, even outperforming GPU-based methods on CPU.