News Hub

Apple uses latent-space distillation to compress neural audio encoders for on-device dictation.

Apple introduced probe guidance, a method to guide flow matching models using frozen diffusion model states without extra forward passes.

Apple introduced REVERSAL-BENCH, a benchmark measuring reinforcement learning performance in non-reversible real-world manipulation tasks.

DACA-GRPO improves reinforcement learning for diffusion language models by addressing temporal credit assignment and likelihood bias.

Value induction in LLMs post-trains models on language expressing behavioral traits and values like helpfulness and honesty.

Glyph is a production system using AI agents to generate column descriptions and assign governance labels in enterprise data catalogs.

Apple researchers proposed selective persistent memory to preserve context across multi-turn agentic LLM sessions.

Energy-navigated distillation trains discrete flow matching models to generate text in few steps using multi-step trajectories.

Apple proposes evaluating video captions using multiple-choice question answering instead of text matching.

Apple presents SimpleDesign, a model for jointly generating protein sequences and structures.

Apple proposes REFACTOR-VLA, which learns reusable motor programs for vision-language-action models.

Study quantifies internal inconsistencies in LLM probabilistic beliefs using information processing analysis.

Agent Seer synthesizes realistic test scenarios for tool-using AI agents by extracting specifications from tool documentation.

Apple proposes rubric-based reward framework that decomposes answer quality into multiple dimensions for question answering alignment.

Apple introduced Luce, a 3D representation unifying geometry and physically-based rendering materials for image-to-3D generation.

Apple's PROOF-Gen optimizes data generation for efficient distillation of tool-calling capabilities into smaller models.

Apple researchers proposed IDEA Prune, an enlarge-and-prune pipeline for efficient language model pretraining within inference budget constraints.

STARFlow2 enables unified multimodal generation of interleaved text-image sequences with improved visual fidelity.

Models can learn internalized visual thinking for video reasoning without generating intermediate reasoning images.

Apple research uses lexical interventions for cross-lingual knowledge transfer with scarce target language data.