Researchers at Apple trained a 1.7B-parameter categorical flow model on 2.1T tokens and self-distilled it into a Categorical Flow Map (CFM) that generates text in as few as 4 inference steps while mai…
A study by researchers from Seoul National University, University of California, Berkeley, ICSI, LBNL, and University of Texas at Austin finds that Diffusion Language Models (DLMs) achieve higher arit…
UC Berkeley, ICSI, and LBNL researchers introduced ARBITRAGE, a step-level speculative decoding framework that uses a lightweight router to dynamically choose between draft and target model steps, red…
Researchers introduced DeepAmbigQA, a dataset of 3,600 questions requiring multi-hop reasoning with half containing explicit name ambiguity, to benchmark LLM answer completeness. Tests showed that eve…
Researchers from Apple and academic institutions introduced Dual-Stage Registers (DSR), a register-based intervention that reduces outlier tokens in Diffusion Transformers (DiTs) for image generation,…
Researchers from Georgia Tech and Apple have shown that applying standard graph algorithms to UMAP's internal k-nearest-neighbor (kNN) graph can enhance data sensemaking, outperforming or complementin…
Meta researchers introduced MoMo, a two-stage imitation-learning framework that enables robots to dial motion modes—steady, dynamic, or intermediate—during manipulation tasks. In tests across six real…
Apple's Siri Expressive Voices, powered by the AFM 3 Core Advanced on-device foundation model, now synthesize rich, configurable speech in real time using a memory-efficient audio synthesis architectu…
A new method called Lookahead-Enhanced Atomic Decomposition (LEAD) breaks the no-recovery bottleneck in long-horizon reasoning for Large Language Models, enabling the o4-mini model to solve Checkers J…
Researchers from Apple propose an environment-free synthetic data generation method for training API-calling LLM agents, using LLMs as digital world models to generate trajectories without executable …
Researchers at an undisclosed institution introduced CalibAtt, a training-free method that accelerates text-to-video generation by up to 1.58× via calibrated sparse attention. The method identifies an…
Researchers from University of California, Santa Barbara, Carnegie Mellon University, LMSYS Org, and University of Wisconsin–Madison introduced the Length Value Model (LenVM), a token-level framework …
Apple researchers have developed an efficient unlearning framework that reduces computational costs by up to ~50% by identifying and excluding low-influence data points before the unlearning process, …
Researchers introduce Visual Concept Inference from Sets (VICIS), a task that evaluates vision-language models' ability to infer shared concepts from example images and apply them to new inputs. State…
Apple researchers developed a personalization system for incremental video search on Apple TV that combines text-based and ID-based embeddings, achieving a 2.99% improvement in NDCG@10 and 3.30% in MR…
Simple self-distillation (SSD) improves LLM code generation by fine-tuning models on their own sampled outputs, boosting Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6, with gains c…
Researchers at Apple and other institutions published the first evaluation of uncertainty quantification (UQ) methods for LLM function-calling, finding that multi-sample UQ methods like Semantic Entro…
Researchers from Apple propose FAE (Feature Auto-Encoder), a framework that adapts pre-trained visual encoders like DINO and SigLIP for image generation using as little as a single attention layer. On…
Researchers at Apple propose CLaRa (Continuous Latent Reasoning), a unified framework that bridges retrieval and generation in retrieval-augmented generation (RAG) by performing embedding-based compre…
Researchers from the University of California, Santa Barbara, the University of Washington, and an independent researcher, with work done while at Apple, introduced the Proactive Agent Research Enviro…