15 stories
Apple ML Research researchers show that information accessible via a model's top-k logits can reveal image content that model owners expect to be inaccessible, sometimes matching what is exposed by direct residual stream projections.
A new benchmark from Apple ML Research probes LLMs on four distinct context understanding tasks across nine datasets, finding that pretrained dense models struggle with nuanced contextual features and that 3-bit quantization causes variable performance drops.
Apple presents five research highlights at ICLR 2026 in Rio de Janeiro, covering parallel RNN training, SSM length generalization, unified image understanding and generation, real-time 3D scene synthesis, and a new protein folding approach.
Apple's ParaRNN framework uses Newton's method to parallelize nonlinear RNN training, achieving a 665x speedup over sequential approaches and enabling billion-scale RNN language models for the first time.
Apple ML researchers model scene dynamics by working directly in a compressed motion embedding space derived from large-scale trajectory data, enabling efficient motion generation conditioned on text or spatial inputs.
A first-place finish in the March 2026 Kaggle Playground churn prediction competition came from a four-level stack of 150 models selected from 850 runs, generated by GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6 in a human-in-the-loop workflow with GPU-accelerated execution.
Microsoft Research and collaborators introduce ADeLe, a framework that characterizes both benchmarks and LLMs using shared capability scores, achieving ~88% prediction accuracy on unseen tasks for models like GPT-4o and LLaMA-3.1-405B.
An open-source framework from Microsoft Research automates the selection between RAG and fine-tuning approaches, plans adaptation strategies against real deployment constraints, and replaces manual hyperparameter search with a budgeted refinement loop.
A research experiment built with NYU uses an Executive LLM to steer multi-party AI conversations toward targeted skill assessment, with AI scoring accuracy matching human expert agreement rates in a 188-person study.
Google's Simula framework decomposes synthetic dataset creation into independently controllable axes — diversity, complexity, and quality — and has already been deployed in Gemma safety models, Android scam detection, and spam filtering.
Google Cloud's ICLR paper introduces a structured memory framework that distills both failed and successful agent trajectories into transferable reasoning strategies, beating memory-free baselines by up to 8.3% on web benchmarks.
Google's Auto frame feature uses 3D scene reconstruction and generative inpainting to re-render photos from a different viewpoint, fixing parallax and perspective distortion without a reshoot.
Google DeepMind's new text-to-speech model scores 1,211 on the Artificial Analysis TTS Elo leaderboard, introduces natural-language audio tags for vocal direction, and watermarks all output with SynthID.
Accenture, Bain, BCG, Deloitte, and McKinsey gain early access to Gemini models and direct DeepMind technical engagement in a push to close a significant gap between enterprise AI ambition and production deployment.
A new distributed training architecture from Google DeepMind uses asynchronous compute islands to train large models across distant locations — more than 20x faster than conventional synchronization, with self-healing fault tolerance.