Apple Machine Learning Research

52 articles in the TruthFoundry index. Each links out to the original.

• Weighted State Souping Improves Long-Context Language Modeling · Weighted state souping enhances long-context language modeling by learning query-dependent weights for cached states.
• Gating Mechanism Impacts State Space Models' Performance in In-Context Learning · Research shows that gating mechanisms in state space models influence their ability to perform in-context learning and memorization.
• Evaluating LLM Judges for Personalized Writing Style Assessment · A study examines how well large language model judges evaluate personalized writing style preferences.
• Apple Researchers Develop Neural Networks to Enhance RAG Efficiency with Faster MIPS · Apple researchers present neural networks to improve RAG efficiency by optimizing Maximum Inner Product Search (MIPS) through learned support functions
• GHARP real-time 3D head animation method achieves state-of-the-art performance · Researchers present GHARP, a fast high-quality real-time 3D human head animation system.
• MLX enables distributed trillion-parameter LLM inference on Mac Studio clusters · MLX runs 816B parameter Muse Spark Open across four networked Mac Studio devices.
• New SSR Framework Cuts Multimodal Search Agent Inference Latency Dramatically · Researchers introduce Selection-based Structured Reasoning for efficient multimodal search agents.
• Study compares Plan-then-Retrieve vs Retrieve-then-Plan LLM tool architectures · LLM tool discovery architecture performance varies by query complexity per new research.
• Speech Foundation Models Enable Direct Code-Switching Audio Detection · Researchers present lightweight code-switching detection using pretrained speech encoder hidden states.
• Researchers introduce first scratch-pretrained tri-modal Masked Diffusion Models · Research team presents first tri-modal Masked Diffusion Model pretrained across text, image, audio data.
• Bilingual pre-training outperforms hyperparameter tuning for low-resource LLMs · Research finds bilingual data mixing beats hyperparameter search in data-constrained LLM training.
• SELFCOMPACT Enables Models to Self-Decide Context Compaction · SELFCOMPACT allows language models to autonomously decide when to summarize context, improving efficiency on math and search benchmarks.
• Minions: Distributed GPU Cloud Aggregates Idle Apple Silicon for ML Inference · Minions aggregates idle Apple Silicon GPUs across macOS devices to execute large-scale machine learning inference workloads.
• Study Finds Both LLMs And Humans Struggle With Translation Annotation · Research evaluates LLM performance against humans for machine translation quality annotation.
• Researchers Introduce Normalizing Trajectory Models For Fast Generative Sampling · Normalizing Trajectory Models enable 4-step image generation while retaining exact likelihood training.
• RISED rubric framework improves multi-environment LLM agent training · Researchers introduce RISED, a rubric-based method for training generalist LLM agents.
• Researchers Design Probes to Let Users Define Personal Sensing Systems · Researchers created probes enabling users to define phenomena for personalized machine learning systems.
• Language Discrimination Improves Multilingual Speech Model Performance · Strengthening language discrimination in HuBERT reduces the multilingual performance gap on phonetic and lexical measures.
• Diffusion Models Fail to Match Dependent Token Distributions · Researchers prove discrete diffusion models cannot accurately generate sequences with inherent token dependencies.
• Study finds elaborate MLE agent harnesses offer no performance benefit · Research shows strong LLM backbones outperform complex harnesses for ML engineering agents.
• New RL Method Enables Self-Improvement via Internalized Feedback · Researchers introduce RLTL;DR, a reinforcement learning method that allows agents to self-improve by internalizing feedback from failed attempts.
• SCLATE Framework Simplifies Continual-Learning Agent Training and Evaluation · SCLATE is a new execution substrate that standardizes event scheduling and memory for continual-learning agents.
• Study Reveals Trade-Off Between LLM Conditioning Effectiveness and Fluency · Systematic study finds efficient LLM steering methods often degrade generation fluency.
• Language Models Lose Precision Serializing Tree Structures · A study reveals significant accuracy loss when language models serialize and extract hierarchical arithmetic expressions.
• Researchers present improved rates for federated variational inequalities · New LIPPAX algorithm mitigates client drift for faster federated optimization performance.
• Stabilized Online Teachers Close Gap in Semi-Supervised Federated ASR · New method stabilizes online teachers and server updates to improve semi-supervised federated automatic speech recognition.
• Apple Compresses On-Device Speech Tokenizer via Latent-Space Distillation · Researchers compress Apple's on-device speech tokenizer by 2.8x using latent-space distillation, maintaining accuracy without fine-tuning.
• New Probe Guidance Method Improves Diffusion Language Models · Researchers introduce probe guidance to improve continuous diffusion language models without extra inference passes.
• DSAS Dynamically Scales Activation Steering for LLMs · Researchers introduce DSAS to adaptively modulate activation steering strength based on context and input.
• DACA-GRPO Enhances Diffusion LLMs via Denoising-Aware Credit Assignment · Researchers propose DACA-GRPO to fix credit assignment and bias in reinforcement learning for diffusion language models.