The post above shares just a handful of highlights, and a comprehensive schedule of Apple's contributions at #NeurIPS2025 can be found here: machinelearning.apple.com/updates/appl... 7/7

And at booth 1103, attendees will be able to see demos of Apple research - including distributed compute using MLX to run a 1T model on cluster of 4 Mac Studios, and a demo of FastVLM: machinelearning.apple.com/research/fas... 6/7

A Principled Approach to Determining Training Data Mixtures:
"Scaling Laws for Optimal Data Mixtures" machinelearning.apple.com/research/opt... 5/7

Innovative Approaches to Generative AI:
"STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis" machinelearning.apple.com/research/sta...
"LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss" machinelearning.apple.com/research/end... 4/7

Understanding the Strengths and Limitations of Reasoning Models:
"The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity" machinelearning.apple.com/research/ill... 3/7

Advancing Privacy-Preserving ML:
"Instance-Optimality for Private KL Distribution Estimation" machinelearning.apple.com/research/ins...
"Privacy Amplification by Random Allocation" machinelearning.apple.com/research/pri... 2/7

tl;dr: some parameters are much more important than others, and in some cases removing just 1 can turn an LLM's output to nonsense

The inference code, model checkpoints, and an iOS/macOS demo app based on MLX are available here: github.com/apple/ml-fas...

How fast is it? Here's the demo app running FastVLM 0.5B model on iPhone 16 Pro. Time to first token is shown on the screen, highlighting near real-time performance.

And for a comprehensive overview of Apple research at the conference - including the complete schedule of orals, posters, workshops, booth programming and more - see this post: machinelearning.apple.com/updates/appl...

Accepted as a Spotlight at @iclr-conf.bsky.social the work shares a new method for fine-grained control over #genAI output - without the computational overhead, complexity, and volume of data needed by #RLHF or fine-tuning, and with more reliable results than prompt engineering.