AI News from Apple ML Research

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

• Putting Captions to the Test: Evaluating Video Caption Quality through Multiple-Choice Question Answering — Apple researchers introduce CapQuiz at ACL 2026, a reference-free video captioning benchmark that tests quality by measuring whether captions help answer human-verified multiple-choice questions — outperforming BLEU/SPICE-style metrics on human judgment correlation and exposing systematic blind spots in current VLLMs.
• DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation — Apple researchers introduce DiscoSign at EMNLP 2026 — the first framework for discourse-aware text-to-ASL gloss translation, addressing spatial coreference, question-answer clause structures, and concept-gloss consistency that sentence-level systems miss entirely and that directly affect comprehension for Deaf and Hard-of-Hearing users.
• SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign — Apple researchers challenge the standard two-stage protein design pipeline with SimpleDesign: trained end-to-end directly on 2M+ sequence-structure pairs, it matches multi-stage latent-space models on codesign benchmarks — suggesting the autoencoder pretraining stage most prior work depends on is unnecessary.
• REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs — Apple researchers introduce REFACTOR-VLA, a robot policy system that learns reusable motor skills via a wake/sleep architecture, finding that InfoNCE contrastive loss during world-model training dramatically improves skill clustering on the LIBERO benchmark.
• Agent Seer: Synthesizing Scenarios from Specification Understanding — Apple's Agent Seer auto-generates graded evaluation scenarios for tool-calling agents from MCP specifications alone — no manual curation or live tool access needed — with strong quality across 7 diverse domains. Key finding: argument value accuracy, not tool selection, is the failure mode that coarse benchmarks miss.
• LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs’ Probabilistic Beliefs — Apple ML Research finds LLMs inconsistently update beliefs under new evidence — and counterintuitively, non-Bayesian heuristic updates often outperform exact Bayesian processing. The diagnosis: LLMs' underlying probabilistic world models are misspecified, not just their reasoning. Direct implications for AI deployed in medicine, science, and law.
• From Preferences to Principles: Rubric-Based Alignment for Grounded Knowledge Answers — Apple researchers introduce a rubric-based reward framework for open-domain QA that improves over instruction-tuned baselines by 6.5%, using query-specific rubrics grounded in retrieved evidence across three quality dimensions — composition, grounding, and instruction-following.
• Luce: Relightable Gaussians for 3D Asset Generation — Apple's Luce converts a single image into a relightable 3D asset with physically-based rendering materials, improving FID by 28% over the prior best on the Toys4K benchmark and recording a CLIP alignment score of 0.8519 vs. 0.8299 on a new AI-image benchmark.
• PROOF-Gen: From Optimized Data to Better Distillation — Apple researchers' PROOF-Gen recovers 93% of failed teacher trajectories during model distillation by using a per-scenario reflector to generate corrective guidance, boosting Qwen3-4B Pass^1 on tool-calling from 0.132 to 0.529 — a 4x gain — with positive transfer to deployed on-device models across all locales.
• IDEA Prune: An Integrated Enlarge-and-Prune Pipeline in Generative Language Model Pretraining — Apple ML Research presents IDEA Prune: an integrated enlarge-and-prune pipeline that outperforms training target-size models from scratch by pretraining a larger model then compressing it via iterative structured pruning under a single cosine annealing schedule — validated at production scale compressing 2.8B to 1.3B parameters with up to 2T pretraining tokens.
• STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation — Apple's STARFlow2 achieves unified text-image generation using autoregressive normalizing flows — architecturally identical to causal Transformers — eliminating the structural mismatch between LLM text generation and diffusion-based image synthesis in a single causal forward pass.
• Beyond Visual CoT: Internalized Visual Thinking for Proactive Video Reasoning — Apple ML Research introduces Internalized Visual Thinking (IVT), a post-training method that bakes visual foresight into model weights during training — eliminating inference-time image generation and cutting latency by over 5× versus Visual Chain-of-Thought.
• Scaling Laws for Mixture Pretraining Under Data Constraints — Apple ML Research finds that in mixture pretraining, scarce domain-specific data can be repeated 15–20 times before hitting diminishing returns, and introduces a repetition-aware scaling law to help practitioners compute optimal data mix configurations — validated across 2,000+ training runs.
• Multilingual Knowledge Transfer under Data Constraints via Lexical Interventions — Apple researchers propose LINK, a pretraining data intervention that improves multilingual knowledge transfer by randomly swapping English words with target-language translations — requiring only a bilingual vocabulary and achieving up to 2x faster convergence for low-resource languages.
• Progressive Refinement: An Iterative Pseudo-Labeling Approach for Mandarin-English Code-Switching ASR — Apple researchers apply iterative pseudo-labeling to Mandarin-English code-switching speech recognition for the first time, achieving 6.35% and 8.29% Mix Error Rate reductions on SEAME benchmarks by bootstrapping from unlabeled audio.
• Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts — Apple researchers analyzed 21,000 multi-turn conversations across GPT-4o, GPT-4.1-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash, finding human evaluators consider AI self-referential and relationship-building behaviors less appropriate than from humans — but boundary-setting more appropriate from AI than from people.
• The P-Completeness of Inverted Index Traversal: On the Complexity of Evaluating Boolean Query DAGs — Apple researchers prove that evaluating boolean query DAGs over inverted indices is P-Complete, then introduce ComputePN — an algorithm that bounds evaluation to O(|Q| · |U_active|) time by decoupling logical negation from universe-wide scans, directly relevant to AI agent search pipelines.
• 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.