AI Lab
Official Apple announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
GHARP splits head animation into an offline identity stage and a lightweight runtime network, hitting state-of-the-art quality on the Ava-256 benchmark while running up to 13× faster on an A100 GPU with 8× fewer Gaussians. A body alignment network fixes blur and flicker from underspecified body pose and clothing.
On ToolBench's 46,984 API endpoints, Retrieve-then-Plan wins on single-tool queries while Plan-then-Retrieve wins on multi-tool queries, holding at every retrieval budget k=1 to 20. PtR's extra discovery call roughly doubles inference cost, and a stub-hallucination failure mode invents non-existent API names.
Selection-based Structured Reasoning reformulates agent reasoning as selection from pre-specified natural-language candidates rather than freeform generation. Tested on seven multimodal search benchmarks with 2B and 4B models, it cut per-turn reasoning latency by over 90% and total per-question inference latency by 28-54%.
Apple says a cluster of four Mac Studios running the newly released MLX generates over 36 tokens per second on Muse Spark Open, an 816B-parameter MoE model in NVFP4 precision. Benchmarks also cover Mistral Medium 3.5 (128B) and Qwen 3.8 (27B) on M5 Ultra versus M3 Ultra, with RDMA over Thunderbolt scaling across Macs.
Muse Spark Open is an 816B mixture-of-experts model with 42B active parameters and a 1M token context window, scoring 39 on the Artificial Analysis Intelligence Index — the #1 US open-weights model. Meta trained it on NVIDIA GB300 NVL72 for agentic coding, and it adds audio and video perception to image and document understanding.
Whisper-small encoder hit a mean F1 of 0.798 across 14 language pairs in CS-FLEURS, beating a CNN (0.776) and a HuBERT linear baseline (0.751). The method finds language-selective hidden-state dimensions using monolingual recordings, avoiding reliance on multilingual transcription.
Apple's study pretrains a Masked Diffusion Model from scratch on text, image-text, and audio-text data, reaching 3B parameters over 6.4T tokens with results in text generation, text-to-image, and text-to-speech. It reports the largest systematic open study of multimodal discrete diffusion models to date.
Across ~1,000 pre-training runs at four scales (150M–1.43B params), mixing English into Arabic training beat weight-decay tuning, gaining the equivalent of 2–3x unique target data on validation loss and 2–13x on downstream accuracy.
SelfCompact lets the model itself decide when to summarize its context via a compaction tool plus a lightweight rubric, needing no fine-tuning. Across six benchmarks and seven models it beat a no-summarization baseline by up to 18.1 points on math and 5-9 points on agentic search at 30-70% lower per-question cost.
Apple researchers present Minions, a distributed virtual GPU cloud that pools idle GPUs across macOS devices to run large-scale ML inference without buying new hardware. The system adds fault tolerance and flexible task execution for teams facing constrained centralized GPU capacity.
Apple researchers tested LLMs against human annotators on MQM and Error Span Annotation across 70 language pairs plus WMT23 and WMT25 data. LLM agreement beat human-human agreement on some tasks but varied widely by scheme, language pair, and domain, staying unreliable in most settings.
NTM models each reverse diffusion step as a conditional normalizing flow with exact likelihood training, matching or beating image generation baselines in four sampling steps. A lightweight denoiser self-distilled from the model's own score function produces the four-step samples.
RISED repurposes rubrics as both an online data-selection signal and a supervision source for multi-environment LLM agent training, instead of relying only on scalar rewards. An LLM judge tags each rollout with a shared rubric vocabulary; positive rubrics feed an on-policy self-distillation teacher, negative ones steer later rollouts away from repeated failures.
Apple's paper proves a diffusion step matches the training distribution only when written positions are conditionally independent given fixed tokens, and that no product of per-position distributions can match a dependent group.
Strengthening language discrimination during pretraining cut phone-ABX error from 11.6% to 10.4% in a controlled English/French HuBERT setup, beating the 10.8% monolingual baseline. Lexical sWUGGY rose from 52.1% to 56.7% and ProsAudit prosody from 68.9% to 72.9%.
Apple researchers introduce RLTL;DR: after each failed attempt the policy writes a one-line TL;DR insight from verifier output, and the next rollout is conditioned on all prior insights. On tool-calling and coding sets filtered to Pass@128 = 0, GRPO training of a Qwen 3.5 9B Thinking policy stays at 0-1% Pass@1, while RLTL;DR reaches 14-31% with insights in context and 12-13% without them at eval.
Under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art MLE harnesses showed no advantage over a single session of a minimal-harness coding agent with read, write, and bash primitives. Apple researchers conclude the backbone is the primary driver and hand-crafted harness layers are redundant for current MLE benchmarks.
SCLATE is an execution substrate where benchmarks and unmodified agents add events to one open scheduler on a hybrid simulated clock. Porting seven benchmarks and comparing ten harness/memory configs on ten models showed added memory doesn't reliably beat a harness's native memory.
Apple researchers systematically compared LLM conditioning methods for concept injection and removal, finding efficient activation steering achieves control at a steep cost to fluency. Steering is also far less effective on instruction-tuned models than on base models, while prompting and supervised fine-tuning work for injection but not removal.