AI BriefingApple ML news
The latest AI and developer news from Apple ML, with short summaries.
Visit website →Scaling Categorical Flow Maps
Apple researchers implemented 4-step text generation using a 1.7B CFM model.
DeepAmbigQA: Ambiguous Multi-Hop Questions for Evaluating LLM Answer Completeness
DeepAmbigQA combines ambiguity and multi-hop reasoning to evaluate the answer completeness of LLMs.
Locking Pre-trained Weights via Deep Low-Rank Residual Distillation
DLR-Lock makes unauthorized modification of pre-trained weights difficult while maintaining model performance.
Taming Outlier Tokens in Diffusion Transformers
Researchers proposed the DSR technique to reduce outlier tokens in DiTs, improving image generation quality.
Understanding Alignment in Multimodal LLMs: A Comprehensive Study
The study comparatively analyzed preference alignment methods and datasets for multimodal LLMs and proposed BDHS.
MoMo: Controlling Robot Manipulation Motion Modes via Spatiotemporal Action Tokenization
We propose MoMo, a framework that allows robots to adjust their motion style according to the task environment through spatiotemporal action tokenization.
Where Dimensionality Reduction Meets Network Science: Understanding Data Using UMAP's kNN Graph
The study demonstrates that applying standard graph algorithms to the kNN graph that UMAP builds internally can effectively reveal the structure of data.
Memory-Efficient Audio Synthesis Using a Decoupled Temporal-Depth Diffusion Transformer
Apple has unveiled a new audio synthesis architecture that generates high-quality speech in real time while minimizing memory usage.
GH-ESD: Grounded Hypothesis-Driven Error Slice Discovery for Instance-Level Vision Tasks
The paper proposes the GH-ESD framework, which leverages LLMs and VLMs to discover instance-level error patterns in vision models.
Solving the 'No-Recovery Bottleneck' in Long-Horizon Reasoning
Proposes the LEAD methodology to solve the problem of irrecoverable errors caused by excessive decomposition during long-horizon reasoning in LLMs.
Accelerating Text-to-Video Generation via Calibrated Sparse Attention
CalibAtt is a technique that speeds up video generation models by up to 1.58x by identifying unnecessary connections during attention computation.
Environment-free Synthetic Data Generation Technique for API-Calling Agents
This proposes a technique that generates training data for agents using only API specifications, without a real execution environment.
Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling
LenVM optimizes LLM inference efficiency and performance by predicting the remaining generation length at the token level.
LVSum: A Benchmark for Timestamp-Aware Long Video Summarization
A human-annotated LVSum benchmark has been released to evaluate temporal accuracy in long video summarization.
RayRoPE: Ray Positional Encoding for Multi-View Attention
We propose RayRoPE, a positional encoding scheme with SE(3) invariance and geometric adaptability for multi-view transformers.
Inferring Visual Concepts from Image Sets
A model called VICIS has been proposed that enables Vision-Language Models to infer visual concepts from example images and apply them.
Making Unlearning More Efficient: Cutting Compute Costs with Low-Impact Points
Removing low-influence data first in machine learning models can cut unlearning costs by up to 50%.
Simple Self-Distillation Improves LLM Code Generation Performance
LLMs can significantly improve their code generation ability using only their own outputs, without needing a verifier or separate training.
Personalizing Progressive Video Search: Hybrid Text and ID Embeddings
Apple TV combined Text and ID Embeddings for personalized search, improving accuracy by 8% on short queries.
Doubly Sub-linear Interactive Proofs of Proximity (dsIPPs) for Approximate Verification
We built a doubly sub-linear interactive proof of proximity (dsIPPs) scheme that can verify properties of data by reading only a tiny fraction of the input data.