AI News from Apple ML Research

Latest coverage from Apple ML Research, summarized and scored for signal.

• GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings — Apple ML Research finds that GRPO reasoning training in a model's native language closes most of the gap to English-only training, with strong cross-lingual transfer across languages — but warns that some configurations cause silent capability regressions that only broad evaluation can detect.
• MVICAD2: Multi-View Independent Component Analysis with Delays and Dilations — MVICAD2 extends multi-view ICA for MEG brain data by adding dilation handling alongside delays — critical for neuroscience studies where brain processing speed varies by age and stimulus type, validated on the Cam-CAN aging dataset.
• A Specialized Semismooth Newton Method for Kernel-Based Optimal Transport — Apple ML Research and MIT/Berkeley researchers propose a semismooth Newton method for kernel-based optimal transport that achieves local quadratic convergence, substantially cutting the computational cost that previously made these estimators intractable at scale.
• When Unlearning Is Free: Leveraging Low Influence Points to Reduce Computational Costs — Apple and Harvard researchers propose a machine unlearning framework that skips training data points with negligible model influence, cutting unlearning computation by up to 50%—a practical step toward scalable compliance with GDPR right-to-be-forgotten requirements.
• Arbitrage: Efficient Reasoning via Advantage-Aware Speculation — ARBITRAGE, from Apple and UC Berkeley researchers, cuts LLM reasoning inference latency by up to 2x by dynamically routing generation to a fast draft model or powerful target model step-by-step — avoiding the token-mismatch failures that make standard speculative decoding ineffective on reasoning tasks.
• Beyond Next-Token Prediction: A Performance Characterization of Diffusion versus Autoregressive Language Models — A UC Berkeley / Apple study finds diffusion language models achieve higher arithmetic intensity than autoregressive models but fail to scale to long contexts — and that ARMs outperform DLMs on batched inference throughput. The key fix for DLMs: fewer sampling steps.
• Scaling Categorical Flow Maps — Apple researchers scaled Categorical Flow Maps to 1.7B parameters trained on 2.1T tokens, demonstrating that discrete-data flow matching can generate competitive text in as few as 4 inference steps — the first credible scaling test of this architecture beyond 1B parameters.
• DeepAmbigQA: Ambiguous Multi-hop Questions for Benchmarking LLM Answer Completeness — Apple researchers introduce DeepAmbigQA — 3,600 multi-hop questions where half require name disambiguation; even GPT-5 scores only 0.13 exact match on ambiguous questions, exposing a persistent gap in LLM answer completeness for knowledge-intensive tasks.
• Locking Pretrained Weights via Deep Low-Rank Residual Distillation — Apple researchers propose DLR-Lock, which replaces model MLPs with deep low-rank residual networks to make fine-tuning computationally prohibitive — a defense against unauthorized adaptation of open-weight models that holds against attackers with full knowledge of the strategy.
• Taming Outlier Tokens in Diffusion Transformers — Apple researchers find that Diffusion Transformers develop high-norm outlier tokens in intermediate layers that attract excess attention but carry degraded local semantics — and show that simply masking them fails; their Dual-Stage Registers (DSR) fix both encoder and denoiser stages to improve image generation quality.
• Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers — Apple's production TTS system for Siri Expressive Voices achieves 16x real-time audio synthesis on-device with just 21MB peak runtime memory, using a Diffusion Transformer architecture that decouples temporal and depth processing for constant-memory inference.
• GH-ESD: Grounded Hypothesis-Driven Error Slice Discovery for Instance-Level Vision Tasks — Apple ML Research's GH-ESD (ECCV 2026) introduces a framework that uses LLM-generated hypotheses verified by Vision Language Models to discover systematic failure patterns in object detection and segmentation models, improving Precision@10 by 0.10 over baselines.
• LEAD: Breaking the No-Recovery Bottleneck in Long-Horizon Reasoning — Apple ML Research and EPFL show that over-decomposing long-horizon tasks in LLMs creates a "no-recovery bottleneck" where a handful of hard steps become irreversible — their LEAD method fixes this, enabling o4-mini to solve Checkers Jumping at complexity n=13, up from n=11 with extreme decomposition.
• Environment-free Synthetic Data Generation for API-Calling Agents — Apple ML Research proposes generating synthetic training trajectories for API-calling agents without any real executable environment — an LLM simulator mimics API responses from specs alone. Models fine-tuned on this data show significant gains on AppWorld and OfficeBench benchmarks.
• Accelerating Text-to-Video Generation with Calibrated Sparse Attention — Apple researchers introduce CalibAtt, a training-free sparse attention method that speeds up video diffusion models by up to 1.58x on Wan 2.1 14B and Mochi 1 by identifying and skipping token-pair connections that consistently produce negligible attention scores.
• RayRoPE: Projective Ray Positional Encoding for Multi-View Attention — Apple ML Research's ECCV 2026 paper introduces RayRoPE, a positional encoding for multi-view transformers that uses predicted 3D scene points instead of ray directions, yielding 15% relative improvement on LPIPS in CO3D novel-view synthesis and consistent gains on stereo depth estimation.
• LVSum: A Benchmark for Timestamp-Aware Long Video Summarization — Apple ML Research's LVSum benchmark — 72 videos across 13 domains averaging 16 minutes — reveals that transcripts outperform visual frames for MLLM summarization and that all tested models fail systematically at temporal grounding, exposing a critical gap in long video AI evaluation.
• Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling — Apple's LenVM framework assigns a constant negative reward per generated token to model remaining generation length at the token level, lifting a 7B model's exact-length matching score from 30.9 to 64.8 on LIFEBench — above frontier closed-source models — while sustaining 63% GSM8K accuracy at a 200-token budget versus 6% for the baseline.
• When Unlearning Is Free: Leveraging Low Influence Points to Reduce Computational Costs — Apple researchers cut machine unlearning compute by up to ~50% by identifying low-influence training points that can be safely skipped during data removal — a practical step toward affordable GDPR-scale privacy compliance for large ML models.
• Show Me Examples: Inferring Visual Concepts from Image Sets — Apple ML Research reveals that state-of-the-art VLMs fail at inferring shared visual concepts from example image sets — they ignore visual context or default to biased outputs. The VICIS paper (ECCV 2026) introduces both the benchmark task and a training framework that learns concept-specific embeddings from image sets.
• Interactive Proofs for General Distribution Properties — Apple/Weizmann researchers publish a FOCS 2026 paper constructing interactive proof systems for any distribution property decidable by a bounded-depth circuit, achieving verifier complexity Õ(D+N^0.99) — enabling efficient statistical verification without rerunning the full analysis and extending far beyond prior label-invariant-only results.
• Location-Invariant Properties of Functions Versus Properties of Distributions: United in Testing but Separated in Verification — Weizmann/Apple researchers show that location-invariant function properties and distribution properties — nearly identical in testing complexity — diverge sharply in interactive proof verification, and construct new doubly-sublinear IPPs for two natural symmetric function classes where the corresponding distribution properties provably have no efficient IPP.
• Doubly Sub-linear Interactive Proofs of Proximity — Weizmann/Apple researchers introduce doubly sub-linear interactive proofs of proximity (dsIPPs) at ITCS 2026 — a class where both proof generation and verification are faster than reading the full input — demonstrated for branching program properties, Hamming weight approximation, and approximate graph bipartiteness.
• Personalizing Incremental Video Search with Hybrid Text and ID Embeddings — Apple's ML Research team published a RecSys 2026 paper on personalizing Apple TV's keystroke-by-keystroke video search using dual embeddings — a multilingual text encoder and a collaborative ID model — achieving +8.63% NDCG lift on 1–3 character queries and +1.14% tap-through rate in a live A/B test.
• Embarrassingly Simple Self-Distillation Improves Code Generation — Apple researchers show LLMs can boost code generation by ~30% relative using only their own sampled outputs — no teacher model, no RL, just temperature-varied SFT — lifting Qwen3-30B-Instruct from 42.4% to 55.3% pass@1 on LiveCodeBench v6 and challenging assumptions that external feedback is required for self-improvement.