AI BriefingApple ML news
The latest AI and developer news from Apple ML, with short summaries.
Visit website →Text-Conditioned JEPA for Learning Semantically Rich Visual Representations
The paper proposes TC-JEPA, which learns semantically rich visual representations by conditioning on captions.
The core of practical learning-based image compression
A new learning-based image codec cut bitrate by up to 3x compared to existing codecs.
Normalizing Flows Combined with Iterative Denoising
Apple researchers proposed iTARFlow, showing competitive results on ImageNet 64, 128, and 256.
SpecMD: A Comprehensive Study on MoE Expert Prefetching and Caching
SpecMD benchmarks MoE expert cache policies and proposes Least-Stale.
From Where Things Are to What They're For: A Spatial-Functional Intelligence Benchmark for Multimodal LLMs
SFI-Bench is a new video benchmark that evaluates MLLMs' spatial and functional reasoning with over 1,700 questions.
Probabilistic KV Routing: Adaptive Layer-wise Cache Sharing
Random cross-layer attention enables layer-wise KV cache sharing.
PORTool: Reward Tree-Based Importance-Aware Policy Optimization for Multi-Tool Integrated Reasoning
PORTool improved both accuracy and tool-call efficiency together through step-wise importance estimation.
Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents
Adding inference-time review to tool-calling agents reduced errors before execution.
International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026
Apple will present three new research papers at ICASSP 2026 in Barcelona from May 4-8.
STARFlow-V: End-to-End Video Generation Modeling Based on Normalizing Flows
STARFlow-V demonstrates high-quality autoregressive video generation without diffusion.
Automatic Generation of Sign Language Annotations Using Sign Language Models
A pseudo-annotation pipeline was proposed that automatically generates annotation candidates based on sign language video and English.
DSO: Direct Steering Optimization for Bias Mitigation
DSO uses RL to optimize steering transformations to reduce bias in VLMs and LLMs.
Adaptive Thinking: LLMs Judge When to Think in Latent Space
Sonata optimizes an LLM's reasoning budget using self-consistency.
Local Mechanisms of Compositional Generalization in Conditional Diffusion
This work showed on CLEVR and SDXL that compositional generalization in Conditional Diffusion corresponds exactly to local conditional scores.
StereoFoley: Video-Based Object-Aware Stereo Audio Generation
StereoFoley generated stereo audio that reflects object positions from video.
Efficient Kinematics Generation via Long-Horizon Motion Embedding
It generates long motions using a motion embedding compressed 64x from large-scale trajectories.
ParaRNN: Large-Scale Nonlinear RNNs That Can Be Trained in Parallel
Apple's ParaRNN parallelizes training of nonlinear RNNs for LLMs, achieving a 665× speedup.
Apple's Machine Learning Research at ICLR 2026
Apple unveiled research on RNNs, SSMs, multimodal, 3D, and protein folding at ICLR 2026.
Can Large Language Models Understand Context
A benchmark of 4 tasks and 9 datasets verified the limits of LLMs' context understanding.
What Do Your Logits Know? (The Answer May Surprise You!)
Even the final **logits** alone can leak unnecessary information about an image query.