Nov 5, 2025

• Apple uses 3D Gaussian splatting for Personas and 3D conversions of photos

  • Apple's Vision Pro introduces Personas, advanced virtual avatars created through 3D photo scans, enabling real-time interaction among users, marking a significant leap in VR technology.

• Optimizing Datalog for the GPU

  • The paper introduces a hash-indexed sorted array for efficient Datalog execution on GPUs, enhancing performance through optimized memory access patterns and reducing redundant computations during Semi-naïve Evaluation.

• Kosmos: An AI Scientist for Autonomous Discovery

  • Kosmos is an advanced AI scientist that automates data-driven discovery, executing up to 42,000 lines of code and analyzing 1,500 papers in a single run, significantly enhancing research efficiency.

• Learning from Failure to Tackle Hard Problems

  • BaNEL (Bayesian Negative Evidence Learning) leverages failed attempts to train generative models, addressing the challenge of extremely sparse rewards in complex problem-solving scenarios, such as drug discovery and theorem proving.

• Building blobd: single-machine object store with sub-ms reads and 15 GB/s upload

  • blobd is a newly designed single-machine object store that achieves sub-millisecond read latencies and 15 GB/s upload speeds by leveraging direct I/O, atomic writes, and io_uring for efficient asynchronous operations, significantly outperforming traditional solutions like S3 and MinIO.

• Deutsche Telekom and NVIDIA Launch Industrial AI Cloud — a ‘New Era’ for Germany’s Industrial Transformation

  • Deutsche Telekom and NVIDIA have launched the world’s first Industrial AI Cloud, a sovereign platform designed to enhance Germany's industrial capabilities, set to go live in early 2026.

• Trajectory Distillation for Foundation Models

  • Trajectory distillation offers a leaner alternative to traditional reinforcement learning (RL) for post-training foundation models, achieving comparable performance at a 10× lower cost.

• DeepInverse Joins the PyTorch Ecosystem: the library for solving imaging inverse problems with deep learning

  • DeepInverse is an open-source library that simplifies deep learning for imaging across various domains, including medical imaging and computational photography, by providing tools for image reconstruction and state-of-the-art neural networks.

• Moral Uncertainty Around Emerging AI Introspection

  • Emerging AI models exhibit functional self-modeling, suggesting they can describe their reasoning and detect internal inconsistencies, raising questions about the empirical nature of consciousness in AI.

• Oolong: Evaluating Long Context Reasoning and Aggregation Capabilities

  • Oolong introduces a benchmark for long-context reasoning tasks, emphasizing the need for models to analyze and aggregate information from large text chunks rather than relying on selective retrieval methods.

• MemSearcher: Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learning

  • MemSearcher innovatively combines current user queries with a compact memory, enhancing reasoning and search efficiency while minimizing computational costs, thus addressing the limitations of traditional search agents.

• Using Span Queries to Optimize for Cache and Attention Locality

  • Span queries generalize inference server interfaces, enabling diverse workloads like chat and RAG to be expressed as expression trees with commutativity constraints, enhancing flexibility and performance.

• Apriel-H1: Towards Efficient Enterprise Reasoning Models

  • The Apriel-H1 family of hybrid LLMs combines transformer attention and State Space Models (SSMs), achieving 15B model size with linear inference complexity and improved throughput for reasoning tasks.

• Federated Attention: A Distributed Paradigm for Collaborative LLM Inference over Edge Networks

  • Federated Attention (FedAttn) integrates the federated paradigm into self-attention mechanisms, enabling privacy protection and efficiency in collaborative LLM inference over edge networks.

• The Sequential Edge: Inverse-Entropy Voting Beats Parallel Self-Consistency at Matched Compute

  • Sequential scaling outperforms parallel self-consistency in language model reasoning, achieving accuracy gains of up to 46.7% across various configurations, as demonstrated through evaluations on five state-of-the-art models and three reasoning benchmarks.