Items tagged with “Apple”.
A new RL approach uses instruction-derived checklists to guide alignment, outperforming fixed-criteria reward models across multiple benchmarks on Qwen2.5-7B-Instruct, presented at ICLR 2025.
Apple ML Research introduces SlowFast-LLaVA-1.5 (SF-LLaVA-1.5), a family of token-efficient video LLMs designed for long-form video understanding. It leverages SlowFast two-streams and public data to achieve state-of-the-art results at 1B–7B scales, with mobile-friendly implications.
Apple researchers identify 'super weights'—an extremely small subset of LLM parameters—that can decisively influence model behavior, enabling compression ideas and raising questions about internal dynamics.
An in-depth look at intersectional bias in LLMs through a new benchmark and a confidence-based fairness metric, revealing reliability gaps in decision-support scenarios.
This Apple ML Research summary outlines a new approach that extends Non-Negative Matrix Factorization (NMF) beyond regularly sampled data by using learnable functions, enabling analysis on irregular time-frequency representations commonly used in audio.
OCAT fine-tunes a CAT model by freezing most parameters and tuning a small corpus-related set, delivering stronger translation quality on WMT23 EN→ZH and EN→DE while maintaining robustness to hyperparameters.
Apple Research shows a joint ASR and pitch accent detection model enhances semi-supervised speech representations, achieving a 41% improvement in F1-score for pitch accent detection and a 28.3% WER reduction on LibriSpeech under limited-resource fine-tuning.
Misty introduces a UI prototyping workflow inspired by conceptual blending, enabling rapid incorporation of design examples into ongoing UIs, backed by a first-use study with 14 frontend developers.
UICoder studies finetuning large language models to generate UI code using automated feedback from compilers and multimodal models, reducing reliance on human feedback and nearing the performance of larger proprietary models.
A deep dive into In-Context Retrieval and Reasoning (ICR2) for long-context LLMs, including benchmarks, methods, and implications for retrieval-augmented generation (RAG).
This work introduces a vector-quantization based approximation to cross-attention for contextual biasing in ASR, enabling scalable, memory-efficient use of large bias catalogs with notable accuracy gains.