Publisher
Stories with reporting from Apple Machine Learning Research, ranked by what is gaining ground now. Each story links to the original articles.
Apple Machine Learning ResearchRISED proposes an approach to selecting training data and environments for a single LLM agent trained across diverse interactive environments. It focuses on relationships among current rollouts across environments rather than relying only on environment-level allocation or local reward signals.
Apple Machine Learning ResearchResearchers report that strengthening language discrimination during pretraining improves multilingual self-supervised speech models under a matched total data budget. The approach reduces, and on some measures closes, the gap with monolingual models on phonetic and higher-level linguistic measures.
Apple Machine Learning ResearchA systematic study investigates trade-offs between how effectively LLM conditioning injects or removes a target concept and the fluency of generated text. It examines a range of conditioning approaches to assess both concept control and generation quality.
Apple Machine Learning ResearchSCLATE proposes a substrate for training and evaluating continual-learning agents across long, multi-session tasks. It schedules agent-side events such as session stops and starts, cron jobs and memory consolidation alongside benchmark tasks.
Apple Machine Learning ResearchResearchers study federated optimization methods for stochastic variational inequalities and establish new convergence bounds. Their work addresses a gap between existing results for these problems and state-of-the-art rates in federated convex optimization.
Apple Machine Learning ResearchResearchers propose a round-trip test to measure how much tree-structured information survives conversion into language-model text and back. Their study uses procedurally generated arithmetic expressions to examine this loss in structured content.
Apple Machine Learning ResearchResearchers analyze confidence limits in discrete diffusion methods that generate multiple token positions per step using per-position distributions. They show that a step matches the training process when tokens have dependencies, including in domains such as pixels, phonemes, and words.
Apple Machine Learning ResearchA study examines how much infrastructure autonomous machine-learning engineering agents need, against a backdrop of progress on public leaderboards and concerns about stagnation over long task cycles. It considers whether increasingly elaborate setups, including multi-agent orchestrators and retrieval subagents, are necessary.
Apple Machine Learning ResearchRLTL;DR examines self-improvement in reinforcement learning with verifiable rewards, where agents may have little or no chance of solving difficult tasks and no teacher models or example solutions to learn from. The work explores internalizing feedback generated by the model itself as an alternative.
Apple Machine Learning ResearchA research paper examines user-authored personal sensing systems and how design choices shape or constrain what users can create. It argues for evaluating such systems beyond usability, usefulness and technical feasibility.