News Hub

Apple research improves maximum inner product search efficiency for RAG and recommendation systems.

Apple Machine Learning Research demonstrated local and distributed LLM inference using MLX and RDMA on Mac Studio.

Apple research introduces Selection-based Structured Reasoning to reduce inference cost in multimodal agents.

Study compares Plan-then-Retrieve versus Retrieve-then-Plan approaches for tool discovery in LLM agents.

Apple researchers detect code-switching in multilingual speech using language-selective dimensions in speech foundation models.

Apple introduces tri-modal Masked Diffusion Models pretrained on text, image-text, and audio-text data.

Apple research shows bilingual pre-training outperforms hyperparameter tuning for low-resource language models.

Apple researchers found LLMs struggle to reliably replace human judgment for machine translation quality evaluation across diverse language pairs, domains, and annotation granularities.

Minions is a distributed virtual GPU cloud that pools idle Apple Silicon GPUs from macOS devices to provide elastic machine learning inference capacity.

Apple proposes SELFCOMPACT, a method for language model agents to compress long traces by removing stale content while preserving partial results mid-derivation.

Normalizing Trajectory Models enable diffusion-based generation in few steps while preserving the likelihood framework.

RISED selects training data for LLM agents across multiple environments using rubrics and cross-environment relationships.

Research examines how users can control ontological design choices in personal sensing systems rather than having them predetermined by designers.

Discrete diffusion models have inherent limitations due to token dependencies in domains like pixels, phonemes, and words.

Apple research shows language discrimination during pretraining reduces performance gaps between multilingual and monolingual speech models.

Apple proposes RLTL;DR, a method for agents to self-improve by internalizing self-generated feedback without external rewards.

Apple presents SCLATE, a substrate for training and evaluating continual-learning agents.

Apple researchers systematically studied trade-offs between LLM output control effectiveness and generation quality.

Apple researchers studied how much tree-structured information survives when language models serialize it into natural language during reasoning.

Apple researchers improved convergence rates for federated optimization solving stochastic variational inequalities problems.