Loading…
Apple
Large American technology company that makes popular consumer electronics, computer software, and online services.
- 40 posts summarised
- latest
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
From their blog
Latest articles
Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings
In data-constrained language model pre-training, two remedies exist: aggressive hyperparameter tuning such as high weight decay, and mixing in data from a high-resource auxiliary language to expand the training distribution. This study systematically compares these levers across four model scales from 150M to 1.43B parameters, using Arabic as the low-resource target and English as the auxiliary, over approximately 1000 pre-training runs. Mixing yields larger improvements than hyperparameter tuning on both validation loss and downstream task accuracy, and the gap grows with model size. Mixing boosts performance by an amount equivalent to 2–3× the unique target data on validation loss and 2–13× on downstream task accuracy, with the gain scaling steeply with model size. The authors recommend mixing in a high-resource language, prioritizing the mixing ratio over hyperparameter tuning, and transferring hyperparameters from a small proxy model via μP, noting that target-language validation loss systematically underestimates mixing’s value.
Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged
The paper evaluates LLMs as alternatives to human judgment for machine translation quality annotation, comparing agreement with human annotators on two schemes: Multidimensional Quality Metrics (MQM) and Error Span Annotation (ESA). Tests cover a long-context set of 70 language pairs plus WMT23 and WMT25 data, examining both score and error span annotation agreement across language pairs and domains. Results show LLM-human agreement sometimes exceeds human-human agreement, but both vary substantially across annotation schemes, language pairs, and domains and remain unreliable for most settings. Humans are particularly challenged by fine-grained MQM annotations and low-resource language pairs, while LLMs struggle with minor errors, wrong language variants, and error span annotation. The authors suggest human-LLM collaborative annotation pipelines as a way to address the reliability issues identified in this work.
Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices
Minions is a dynamic, distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to run large-scale machine learning inference. The work responds to limits in centralized GPU infrastructure, including high cost, limited elasticity, and long wait times when resources are constrained. Demand for GPU-accelerated machine learning inference keeps rising, while modern Apple Silicon macOS devices carry powerful GPUs that stay largely idle during normal usage. Minions offers flexible GPU inference task execution and provides fault tolerance and efficient resource utilization, letting teams use existing hardware in place of additional GPU procurement. The authors report that idle macOS devices can become a practical distributed GPU inference platform in environments where centralized GPU resources are constrained.
Self-Compacting Language Model Agents
Long agent traces of chains of thought and tool calls accumulate stale content that anchors later generations and outgrow context window. SELFCOMPACT lets the model itself decide when and how to compact rather than using fixed-interval compaction triggered at a token threshold. It pairs a compaction tool that summarizes accumulated context with a lightweight rubric for when to fire, such as resolved sub-task or converging trajectory, and when to suppress, such as mid-derivation or stuck. Both are needed: the tool alone is unevenly used across open-weight models, while a rubric closes a meta-cognitive gap without training. On six math and agentic-search benchmarks across seven models, SELFCOMPACT matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over no summarization by up to 18.1 points on math and 5–9 points on agentic search at 30–70% lower per-question cost.
Normalizing Trajectory Models
Diffusion models decompose sampling into many small Gaussian denoising steps, an assumption that breaks down when generation is compressed to a few coarse transitions. Existing few-step methods address this through distillation, consistency training, or adversarial objectives, but sacrifice the likelihood framework. Normalizing Trajectory Models (NTM) models each reverse step as an expressive conditional normalizing flow with exact likelihood training. Architecturally, NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory, forming an end-to-end network trainable from scratch or initializable from pretrained flow-matching models. On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in just four sampling steps while uniquely retaining exact likelihood over the generative trajectory.
RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation
Training a single LLM agent jointly across diverse interactive environments is challenged by scalar rewards that provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. RISED addresses this by using rubrics—textual descriptions of rollout behaviors—to guide both online data selection and policy supervision. An LLM judge tags each rollout with a predefined rubric vocabulary shared across environments, and these profiles select data aligned with the batch’s behavioral composition while limiting overlap with already-selected data. Positive rubrics supply token-level supervision via an on-policy self-distillation teacher, while negative rubrics steer subsequent rollouts away from recurring failure modes. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment.
Negotiating Ontological Boundaries in User-Authored Personal Sensing Systems
Designed artifacts are ontological, shaping what becomes possible, yet systems enabling user authorship are typically evaluated on usability rather than ontological boundary negotiation. Two open-ended probes using a Wizard of Oz technique were created to let participants experience training a personalized machine learning system on phenomena they defined themselves. In a week-long exploratory study, participants used one of two probes during their everyday lives, allowing observation of boundary negotiations. Four sites of ontological boundary negotiation emerged: the boundaries of a phenomenon, the subject as part of relations, distinguishing signal from noise, and the objectivity of data. The paper offers starting points for supporting boundary negotiation through design and proposes open-ended probes as a method for ontological design.
Limits of Confidence in Diffusion
The post analyzes limitations of confidence-based ranking in discrete diffusion, where sequences are generated by writing multiple token positions per step from per-position distributions. It shows that a step matches the training distribution only when the written positions are conditionally independent given fixed tokens, and no product of per-position distributions can match a dependent group. Additionally, per-position distributions do not determine group dependence: two joint distributions can have identical per-position marginals but differ in value combinations. On the synthetic ScanAndAdd task, every group of two or more undetermined positions written by a confidence ranking is dependent, and the generated distribution is 29 times the sampling-noise floor in total variation, with per-sample metrics at 1.0. The post also includes updates on two related readings: on-policy distillation for reasoning models and self-supervised learning with Transformers.
Language Discrimination Improves Linguistic Learning in Multilingual Speech Models
Multilingual self-supervised speech models underperform monolingual models under a matched total pretraining data budget, but strengthening language discrimination during pretraining reduces this gap while preserving cross-language sharing. In a controlled English/French HuBERT setting, two interventions—an auxiliary language classifier and per-language k-means targets—improve language discrimination. Continuous-feature phone discrimination error (phone-ABX) decreased from 11.6% in the bilingual baseline to 10.4%, approaching the monolingual 10.8%. Lexical performance (sWUGGY) rose from 52.1% to 56.7% (monolingual: 58.5%), and prosodic performance (ProsAudit lexical subtask) from 68.9% to 72.9% (monolingual: 72.6%). The strongest gains occur when language discrimination is introduced in the first HuBERT iteration; later or repeated interventions yield smaller improvements and increase language-wise segregation, supporting a causal role for language discrimination in reducing the cost of multilingual learning.
RLTL;DR: Self-Improvement by Internalizing Self-Generated Feedback
Standard reinforcement learning with verifiable rewards (RLVR) fails when tasks are so difficult that agents have no chance of success and no teacher models exist. RLTL;DR addresses this by having the policy write its own TL;DR insight after each failed attempt, conditioning the next rollout on all previous insights. Backpropagation on the in-context insights internalizes a direct task-to-insight mapping. On challenging tool-calling and coding datasets filtered to Pass@128=0, standard GRPO training stays flat at 0-1% Pass@1, while RLTL;DR achieves 14-31% Pass@1 with insights in context during training and 12-13% without at eval time. Reducing to SFTL;DR, training on only 4k (task, insight) tuples without rollouts, recovers almost full performance of RLTL;DR and classical SFT, demonstrating a promising compacted training paradigm.
How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?
Recent autonomous machine learning engineering agents increasingly rely on elaborate machinery such as multi-agent orchestrators and dedicated retrieval subagents, motivated by progress stagnation over long-horizon cycles and limited Large Language Model primitives. This paper finds that under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline that uses direct read, write, and bash primitives. The authors argue via large-scale systematic ablation studies that the machinery layers become redundant in the coding agent setting. The backbone, not the harness complexity, is the primary driver for performance. Consequently, the effort spent elaborating hand-crafted harnesses around strong models yields poor returns for current MLE benchmarks.
On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study
Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet current conditioning methods are often evaluated narrowly on effectiveness without considering generation quality. This study systematically investigates a range of conditioning methods in both concept injection and removal scenarios, including efficient steering, activation steering, prompting, and supervised fine-tuning. The authors find that efficient steering methods frequently achieve their conditioning goal at a steep cost to fluency, degrading generation quality. They identify a critical previously overlooked interaction: activation steering methods are far less effective on instruction-tuned models than on their base counterparts. Simple prompting and supervised fine-tuning are viable options for concept injection but are not as good at concept removal, while cheaply computed textual metrics highly correlate with costly LLM-as-judge scores, providing practical insights for evaluation.
Faster Rates for Federated Variational Inequalities
This paper studies federated optimization for solving stochastic variational inequalities (VIs), a field with growing attention but a significant performance gap compared to federated convex optimization. The authors address this gap by establishing a series of improved convergence rates. They first refine the analysis of the classical Local Extra SGD algorithm, providing tighter guarantees for smooth and monotone VIs. However, they identify an inherent limitation of Local Extra SGD: excessive client drift. To overcome this, they propose the Local Inexact Proximal Point Algorithm with Extra Step (LIPPAX), which mitigates drift and achieves better rates under bounded Hessian, bounded operator, and low-variance regimes. The work also extends these improved guarantees to federated composite variational inequalities.
A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization
Semi-supervised federated learning (SSFL) for ASR uses a teacher to generate pseudo-labels on unlabeled client data while the server trains on a small labeled seed. The paper finds that pseudo-label errors compound across the output sequence and training rounds, causing divergence and a large gap to fully-supervised FL. Two coupled design axes close this gap: the teacher type (per-client online, global broadcast, or transitioning) and the anchor (server-side labeled-data updates that stabilize training). A per-client online teacher diverges on its own but, once stabilized, matches or beats the global teacher in-domain and under domain shift. The server must keep training on labeled data between rounds; otherwise the online teacher drifts, and this interleaving governs convergence more than the seed model does. The proposed guidelines improve over the strongest prior method on 9 of 11 pairs, by 20.8% in-domain and 10.0% cross-domain on average.
Compressing Streaming Neural Audio Encoders via Latent-Space Distillation
Apple’s on-device dictation tokenizer maps waveform to latent representations for a sparsely activated language model, and compressing it reduces memory and latency demands. The authors compress it via latent-space distillation, using the teacher’s pre-quantizer latent as the supervision target rather than tokens or output distributions. A student encoder is trained to regress the teacher’s per-frame latent with a squared-error objective, and an affine layer absorbs width mismatch. This recipe works for both token interfaces and for tokenizers pretrained alone or jointly with a language model, requiring no fine-tuning. At 2.8x compression, the student stays within 1.9% relative WER of its teacher on five of six pairs and outperforms an independently trained tokenizer of identical capacity by 3.9% relative.
How to Guide Your Language Flow
Probe guidance, a novel method for guiding flow matching models, is introduced. It uses the frozen internal states of an existing diffusion model to construct a guidance signal, operating on a similar principle as autoguidance but removing the need for an additional forward pass at inference time. The method provides a reliable path ensuring the weak and strong models share similar dynamics. Probe guidance sets a new state-of-the-art performance on unconditional generation when benchmarked on continuous diffusion language models. Applied to a 1.7B diffusion language model, it consistently improves multiple choice question answering benchmarks.
Dynamically Scaled Activation Steering
Activation steering has emerged as a way to guide generative models toward desired behaviors, but uniform interventions can reduce performance when steering is unnecessary. DSAS, or Dynamically Scaled Activation Steering, addresses this by separating when to steer from how to steer and adapting intervention strength across layers and inputs. At generation time, it computes context-dependent scaling factors that modify existing steering transformations, and it can also be jointly optimized end-to-end with the steering function. Combined with existing methods, DSAS improves the reported Pareto front between toxicity mitigation and utility preservation, while adding minimal computational overhead; experiments also apply it to a text-to-image diffusion model for modulating specific concepts and show token-level interpretability through the strength and location of steering.
REVERSAL-BENCH: A Reversibility Axis and Reset Oracle for Measuring the Reset-Free RL Cliff
REVERSAL-BENCH addresses the challenge of evaluating reset-free reinforcement learning when environments contain irreversible transitions. It controls reversibility with continuous parameter ρ ∈ [0, 1] and supplies a reset oracle, a ground-truth mechanism for verifying state recoverability across eight manipulation settings and five physics engines. Evaluations spanning standard actor-critic, safe reinforcement learning, and specialized reset-free methods find a sharp reversibility cliff: reset-free agents become absorbed into irrecoverable states as ρ increases, while episodic agents maintain steady learning. Geometrically identical reversible counterparts indicate the breakdown is caused by irreversibility rather than obstacle complexity, and the effect persists under learned policies in full physics simulations. The released suite includes a multi-simulator dataset with recoverability labels and a reset oracle; a safety shield predicts recoverability accurately, but active recovery mainly works when agents can steer away from traps.
DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models
DACA-GRPO addresses two weaknesses in reinforcement learning for diffusion language models: existing methods do not assign credit across denoising steps, and mean-field likelihood estimates are biased and high-variance. It is a lightweight, plug-and-play enhancement for GRPO-style trainers that combines Denoising Progress Scores with Stratified Masking Likelihood. The progress scores derive per-token importance weights from intermediate predictions without an additional forward pass, while stratified masking partitions token positions so each token is predicted with most of the sequence as context. Applied to three GRPO base methods, the approach improves results across seven benchmarks covering mathematical reasoning, code generation, constraint satisfaction, and constrained generation. Reported gains reach 5.6 percentage points on math reasoning, 7.4 on code generation, 36.3 on constraint satisfaction, and 5.9 on JSON schema adherence.
Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation
Discrete flow matching can parallelize text generation across positions, but reaching high quality may require hundreds or thousands of model evaluations. The paper argues that few-step distillation fails primarily because teacher trajectories are built from blind stochastic jumps: an early poor midpoint decision propagates, and the student is trained to imitate it. Its Trajectory-Shaped Discrete Flow Matching (TS-DFM) uses a lightweight energy compass during training to score candidate continuations at each midpoint and select the most coherent path, without adding inference cost. In experiments with 170M-parameter language models, an eight-step shaped student achieves 32% lower perplexity than a 1,024-step teacher and runs 128 times faster, with gains across source distributions and three evaluators; it also outperforms the compared discrete-generation baselines, including systems using more data or larger models.
Topics they write about
All topics →Weekly digest
The best engineering reads, once a week.