AI Briefing#apple-research
The latest AI and developer news about #apple-research, with the original source and a short summary.
The latest AI and developer news about #apple-research, with the original source and a short summary.
The study shows that best-of-N alignment closely approaches the theoretical maximum reward gain, while PPO and GRPO remain substantially suboptimal.
A new recipe distills Pythia-1B into a Mamba model, achieving a perplexity of 14.01 compared to the teacher's 13.86.
The method learns query-dependent weights to compose cached state encodings, improving performance on document QA, needle-in-a-haystack retrieval, and tool use.
The study demonstrates that gating mechanisms in SSMs cause models to learn in-weights memorization solutions, delaying or preventing convergence to correct in-context learning even without architectural memory limits.
Study shows judge quality varies by task and criteria, with paired comparisons outperforming independent scoring and top models exhibiting exaggerated sensitivity to subtle details.
The method achieves state-of-the-art quality on the Ava-256 benchmark while using 8x fewer Gaussians.
The study shows a complexity-dependent crossover where Plan-then-Retrieve excels on multi-tool queries while Retrieve-then-Plan wins on single-tool tasks.
The study presents the first tri-modal Masked Diffusion Models pretrained on text, image-text, and audio-text data, achieving strong results at 3B scale.
Mixing high-resource auxiliary data boosts performance equivalent to 2–3× more unique target data on validation loss.
The method improves math performance by up to 18.1 points and agentic search by 5–9 points while reducing token costs by 30–70%.
The study evaluated LLMs against human annotators across 70 language pairs using MQM and ESA schemes, finding both unreliable in most settings.
The study identifies four key sites where users negotiate ontological boundaries when defining their own phenomena for machine learning training.
Apple research shows confidence-based sampling in discrete diffusion models selects dependent token groups, causing a 29x deviation from the training distribution despite perfect per-sample metrics.
Interventions like auxiliary classifiers reduced phone discrimination error from 11.6% to 10.4% in English/French HuBERT models.
Apple researchers introduce RLTL;DR, a method that boosts Pass@1 scores from near-zero to 12–13% on difficult tasks by internalizing self-generated feedback.
The study shows activation steering is less effective on instruction-tuned models and often degrades fluency.
Fine-tuning with ~3,600 examples enables open-weight models to surpass untrained Gemini-3.1-Pro in expression recovery.
Researchers from Apple and Georgia Tech introduce LIPPAX to mitigate client drift and improve convergence rates in federated variational inequalities.
The new method improves accuracy by 20.8% in-domain and 10.0% cross-domain over prior methods by coupling online teachers with server-side update stabilization.
The method achieves 2.8× compression while staying within 1.9% relative WER of the teacher model on five of six pairs.
Apple researchers have unveiled DSAS, a dynamic scaling Activation Steering technique designed to reduce unnecessary performance degradation.
Apple researchers confirmed that injecting specific values into LLMs improves safety but also increases sycophancy.
A study featured on the Apple Research blog proposes TS-DFM to improve the trajectory quality of discrete flow matching, reducing perplexity by 32% compared to a 1,024-step teacher model with 8-step inference and improving speed by 128x.
LLM belief updates diverge from Bayesian principles, and non-Bayesian heuristics actually demonstrated better performance.