Apple researchers introduce split-and-scale, a method that treats 3D scenes as continuous functions of 3D coordinates so a single pretrained tokenizer can be applied to arbitrarily small sub-regions at inference time, expanding latent capacity and detail without raising training costs.
This paper characterizes the information-theoretic limits of KL-regularized alignment, deriving a closed-form expression for optimal reward improvement governed by Jeffreys divergence rather than √KL, and shows best-of-N approaches the limit while PPO and GRPO are suboptimal.
Researchers show that private selection built on the standard noisy aggregation primitive inherently needs Ω(√d/ε) samples, and propose a new SVT-based primitive that yields an O(log(dn)/ε) additive-error selection algorithm, establishing an exponential separation from shuffle and aggregation models.
This paper formalizes a notion of evaluability without the held-out assumption, investigating whether black-box models can be accurately evaluated when test samples may have been seen during training. It fully characterizes when accurate evaluation is possible for output-only access and shows that additional query access can enhance evaluator power, linking to self-correctors in program testing.
This paper introduces weighted state souping, which learns query-dependent weights to combine cached state encodings from state space models, substantially improving long-context retrieval, multi-hop QA, variable tracking, and tool use. A sparse variant achieves near-constant encoding cost with respect to context length. The work was accepted at the NeurIPS 2026 PALM workshop.
This work shows that the gating mechanism in State Space Models causes them to first learn an in-weights memorization solution, delaying or preventing convergence to correct in-context learning, even without architectural limitations. However, gating often benefits generalization to long sequence lengths.
Researchers from Apple and collaborating institutions evaluate LLM judges for personalized writing style tasks. They find judge quality correlates only weakly with general LLM ability (Pearson ρ ≤ 0.2), varies by writing task and evaluation criteria, and that paired comparisons outperform independent scoring. Notably, the strongest LLMs do not most reliably reproduce human evaluations.
Apple researchers present a method at ICML 2026 that trains small neural networks to learn support functions for maximum inner product search, speeding up retrieval in RAG and other applications while maintaining high recall.
Apple researchers propose a two-stage distillation method that first distills a Transformer into a linearized attention model and then into a Mamba architecture, achieving a perplexity of 14.01 close to the teacher's 13.86 on Pythia-1B and largely preserving downstream performance, offering a new pathway for adopting SSMs.
Apple researchers propose an online reinforcement learning framework that trains UI navigation agents directly on live iOS applications. The framework decouples the agent into a planning agent and a UI grounding model, uses an LLM-as-a-Judge to provide dense step-level and terminal rewards, and optimizes with PPO and step-level advantage estimation. They introduce GiGPPO for critic stability and an infrastructure orchestrating hundreds of concurrent virtual iOS devices. Their 8B RL agents achieve nearly a sevenfold improvement over SFT baselines in out-of-distribution generalization to Android and outperform Gemini-3.5-Flash while tying Gemini-3.1-Pro on real iOS apps.
GHARP animates 3D human heads in real time from a few input images and a driving expression signal. It decouples the problem into an offline identity stage and a lightweight runtime animation stage, and uses a body alignment network to resolve ambiguity from body pose and clothing, reducing blur and flicker. It achieves state-of-the-art quality on Ava-256 while running up to 13× faster with 8× fewer Gaussians.
TEAL is the first pixel-level active learning framework built on ViT representations, performing diversity-based selection directly on the native token lattice and refining with margin-based uncertainty on a finer decoder grid. It improves up to +24.36 mIoU over the strongest baseline with one labeled pixel per image and reaches 97% of fully supervised performance with just 10 pixels.
RISED uses LLM-generated rubrics to guide data selection and self-distillation when training a single agent across multiple interactive environments. It improves mean pass rates and per-environment performance over baselines.