Arxiv Papers

Mila developed Milabench, a custom benchmarking suite for deep learning workloads in HPC, based on extensive research and community input, featuring 26 primary and 16 optional benchmarks. https://arxiv.org/abs//2411.11940 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Spotify...

Mila developed Milabench, a custom benchmarking suite for deep learning workloads in HPC, based on extensive research and community input, featuring 26 primary and 16 optional benchmarks. https://arxiv.org/abs//2411.11940 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Spotify...

This paper investigates how different prompt templates impact the performance of Large Language Models, revealing significant variations in effectiveness, particularly in code translation tasks. https://arxiv.org/abs//2411.10541 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016...

This paper investigates how different prompt templates impact the performance of Large Language Models, revealing significant variations in effectiveness, particularly in code translation tasks. https://arxiv.org/abs//2411.10541 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016...

The paper explores using sparse autoencoders to steer language model activations for safer responses, improving refusal behavior while noting potential negative impacts on overall performance. https://arxiv.org/abs//2411.11296 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Sp...

The paper explores using sparse autoencoders to steer language model activations for safer responses, improving refusal behavior while noting potential negative impacts on overall performance. https://arxiv.org/abs//2411.11296 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Sp...

LLaVA-o1 is a novel Vision-Language Model that enhances reasoning in visual question-answering through structured multistage processes, outperforming larger models with fewer training samples. https://arxiv.org/abs//2411.10440 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Sp...

LLaVA-o1 is a novel Vision-Language Model that enhances reasoning in visual question-answering through structured multistage processes, outperforming larger models with fewer training samples. https://arxiv.org/abs//2411.10440 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016 Sp...

The paper introduces affine concept editing (ACE) for controlling language model behavior through activation manipulation, demonstrating improved precision in managing refusal responses across various prompts. https://arxiv.org/abs//2411.09003 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-paper...

The paper introduces affine concept editing (ACE) for controlling language model behavior through activation manipulation, demonstrating improved precision in managing refusal responses across various prompts. https://arxiv.org/abs//2411.09003 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-paper...

The paper introduces Cut Cross-Entropy (CCE), a method that significantly reduces memory usage during training of large language models by optimizing cross-entropy loss computation without sacrificing performance. https://arxiv.org/abs//2411.09009 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-p...

The paper introduces Cut Cross-Entropy (CCE), a method that significantly reduces memory usage during training of large language models by optimizing cross-entropy loss computation without sacrificing performance. https://arxiv.org/abs//2411.09009 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-p...

Add-it is a training-free approach for semantic image editing that seamlessly integrates objects into images using a weighted extended-attention mechanism, achieving state-of-the-art results without fine-tuning. https://arxiv.org/abs//2411.07232 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-pap...

Add-it is a training-free approach for semantic image editing that seamlessly integrates objects into images using a weighted extended-attention mechanism, achieving state-of-the-art results without fine-tuning. https://arxiv.org/abs//2411.07232 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-pap...

The SPA framework enhances user experience by generating diverse, high-quality responses from foundation models using synthetic data and data attribution methods, improving performance in code generation and natural language tasks. https://arxiv.org/abs//2411.06722 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/...

The SPA framework enhances user experience by generating diverse, high-quality responses from foundation models using synthetic data and data attribution methods, improving performance in code generation and natural language tasks. https://arxiv.org/abs//2411.06722 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/...

This paper introduces AIOLI, a new method for dynamically adjusting data mixture proportions during training, outperforming existing methods and improving language model performance across multiple datasets. https://arxiv.org/abs//2411.05735 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/...

This paper introduces AIOLI, a new method for dynamically adjusting data mixture proportions during training, outperforming existing methods and improving language model performance across multiple datasets. https://arxiv.org/abs//2411.05735 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/...

This paper addresses imbalanced computation and memory in pipeline parallelism for large language models by partitioning vocabulary layers, reducing communication barriers, and achieving improved throughput and memory balance. https://arxiv.org/abs//2411.05288 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/po...

This paper addresses imbalanced computation and memory in pipeline parallelism for large language models by partitioning vocabulary layers, reducing communication barriers, and achieving improved throughput and memory balance. https://arxiv.org/abs//2411.05288 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/po...

This study explores whether pre-trained transformer models of chemical structures align with human olfactory perception, demonstrating their ability to predict expert labels and human ratings of odorants. https://arxiv.org/abs//2411.03038 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1...

This study explores whether pre-trained transformer models of chemical structures align with human olfactory perception, demonstrating their ability to predict expert labels and human ratings of odorants. https://arxiv.org/abs//2411.03038 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1...

The paper introduces Mixtures of In-Context Learners (MOICL), enhancing in-context learning by optimizing demonstration subsets, improving performance, and reducing memory usage in Transformer LLMs. https://arxiv.org/abs//2411.02830 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476...

The paper introduces Mixtures of In-Context Learners (MOICL), enhancing in-context learning by optimizing demonstration subsets, improving performance, and reducing memory usage in Transformer LLMs. https://arxiv.org/abs//2411.02830 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476...

OpenAI's Sora evaluates video generation models' ability to learn physical laws, revealing limitations in generalization and suggesting scaling alone isn't enough for uncovering fundamental principles. https://arxiv.org/abs//2411.02385 YouTube: https://www.youtube.com/@ArxivPapers TikTok: https://www.tiktok.com/@arxiv_papers Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692...