AI BriefingApple ML news
The latest AI and developer news from Apple ML, with short summaries.
Visit website →Apple Unveils REFACTOR-VLA: Developing Reusable Robot Motion Skills via Unsupervised Learning
Apple has unveiled the REFACTOR-VLA system, which develops reusable robot motion skills through unsupervised learning.
Apple Releases 'Agent Seer' to Automatically Generate AI Agent Evaluation Scenarios Using Only MCP Specifications
Apple has released Agent Seer, a tool that automatically generates AI agent evaluation scenarios without manual effort, using only MCP specifications.
LLM Probabilistic Belief Updates Diverge from Bayesian Principles, with Non-Bayesian Heuristics Performing Better
LLM belief updates diverge from Bayesian principles, and non-Bayesian heuristics actually demonstrated better performance.
Apple Improves Open-Domain QA Alignment with Search-Distilled Rubric-Reward
Apple improved performance by enhancing open-domain QA alignment using a search-distilled Rubric-Reward framework.
Luce: Relightable Gaussians for 3D Asset Generation
Apple released the Luce model for 3D asset generation and relighting.
Integrated Expansion-Pruning Pipeline for Pre-training Generative Language Models: IDEA Prune
The IDEA Prune pipeline, which integrates pre-training of expanded models with pruning, improved LLM compression performance.
PROOF-Gen: Better Distillation from Optimized Data
Apple released PROOF-Gen, which recovers failed tool-calling scenarios to significantly improve model distillation performance.
STARFlow2 Achieves Unified Multimodal Generation by Connecting Language Models and Normalizing Flows
Apple released STARFlow2, which combines LLMs and normalizing flows to generate text and images in a unified manner.
Beyond Visual CoT: Internalized Visual Thinking for Active Video Reasoning
Apple proposed the IVT framework, which internalizes visual thinking during inference to reduce latency by more than 5x.
Progressive Refinement: An Iterative Pseudo-Labeling Approach for Mandarin-English Code-Switching ASR
Applying iterative pseudo-labeling significantly improved the accuracy of Mandarin-English code-switching speech recognition.
Scaling Laws for Mixture Pretraining Under Data Constraints
Sparse data is optimally reused 15–20 times, leading to the proposal of a scaling law that reflects this.
Lexical Intervention for Multilingual Knowledge Transfer in Data-Constrained Environments
Apple proposed LINK, a data intervention technique leveraging bilingual lexicons, to significantly improve multilingual LLM training efficiency.
P-Completeness of Inverted Index Traversal: On the Complexity of Boolean Query DAG Evaluation
Apple proves the P-completeness of inverted index-based Boolean query evaluation and proposes the efficient ComputePN algorithm.
Multidimensional Analysis of Human-like Behavior in LLMs: Models, User Factors, and System Prompts
Analysis of 21,000 conversations across 4 LLM models verified the appropriateness and controllability of human-like behaviors.
Specialized Semismooth Newton Method for Kernel-Based Optimal Transport
A specialized Semismooth Newton (SSN) method is proposed to address the high computational cost of kernel-based optimal transport.
MVICAD2: Multi-view Independent Component Analysis Considering Delay and Dilation
The MVICAD2 model is proposed to account for both delay and dilation in order to reflect individual differences in brain signal analysis.
Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings
A large-scale analysis of reasoning performance and cross-lingual transfer effects in non-English and multilingual environments using GRPO.
Reducing Unlearning Costs: Leveraging Low-Impact Data Points to Cut Computational Expenses
This study proposes a framework that identifies data points with negligible impact on model training, reducing unlearning computational costs by up to 50%.
Arbitrage: Efficient Inference via Speculative Decoding Considering Model Superiority
ARBITRAGE predicts model superiority to reduce mathematical reasoning latency by up to 2x.
Beyond Next-Token Prediction: Performance Characteristics of Diffusion and Autoregressive Language Models
Diffusion language models enable parallel generation but showed limitations in long contexts and batch processing.