# Apple Machine Learning Research

Apple machine learning teams are engaged in state of the art research in machine learning and artificial intelligence. Learn about the latest advancements.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Theoretical Limits of Language Model Alignment

DevFeed: [Theoretical Limits of Language Model Alignment](<https://devfeed.tech/articles/theoretical-limits-of-language-model-alignment-131227.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/theoretical-limits-alignment>)

Published: 2026-11-12T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [rlhf](<https://devfeed.tech/topics/rlhf.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [dpo](<https://devfeed.tech/topics/dpo.md>)

Tags: [alignment](<https://devfeed.tech/tags/alignment.md>), [best-of](<https://devfeed.tech/tags/best-of.md>), [budget](<https://devfeed.tech/tags/budget.md>), [covariance](<https://devfeed.tech/tags/covariance.md>), [error](<https://devfeed.tech/tags/error.md>), [estimator](<https://devfeed.tech/tags/estimator.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [grpo](<https://devfeed.tech/tags/grpo.md>), [hacking](<https://devfeed.tech/tags/hacking.md>), [inference](<https://devfeed.tech/tags/inference.md>), [kl-divergence](<https://devfeed.tech/tags/kl-divergence.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [limits](<https://devfeed.tech/tags/limits.md>), [model-alignment](<https://devfeed.tech/tags/model-alignment.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [theoretical](<https://devfeed.tech/tags/theoretical.md>)

### AI overview

The excerpt derives an information-theoretic upper limit on expected reward gain for language-model alignment under a fixed KL-divergence budget. Its closed-form bound depends on Jeffreys divergence, and a covariance reformulation offers an estimator using base-model samples. In the proxy-reward setting, reward hacking increases with reward error and as the KL penalty decreases; the analysis shows that ensembling rewards mitigates this gap. Experiments on safety and summarization report that best-of-N approaches the theoretical limit, while PPO and GRPO remain substantially below it. The excerpt argues that reaching optimal alignment without high inference costs requires algorithmic improvements.

### Source excerpt

Language model (LM) alignment improves model outputs to reflect human preferences while preserving the capabilities of the base model. The most common alignment approaches are (i) reinforcement learning, which maximizes the expected reward under a KL-divergence constraint, and (ii) best-of-NNN alignment, which selects the highest-reward output among NNN independent samples. Despite their widespread use, the fundamental limits of reward improvement under a KL budget remain poorly understood. We characterize the information-theoretic limits of KL-regularized alignment by deriving the maximum...

## Understanding Private Evolution as Learning-Augmented Clustering

DevFeed: [Understanding Private Evolution as Learning-Augmented Clustering](<https://devfeed.tech/articles/understanding-private-evolution-as-learning-augmented-clustering-136572.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/private-evolution-clustering>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [clustering](<https://devfeed.tech/topics/clustering.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>)

Tags: [clustering](<https://devfeed.tech/tags/clustering.md>), [convergence](<https://devfeed.tech/tags/convergence.md>), [dimension](<https://devfeed.tech/tags/dimension.md>), [error](<https://devfeed.tech/tags/error.md>), [fail](<https://devfeed.tech/tags/fail.md>), [generation](<https://devfeed.tech/tags/generation.md>), [intrinsic](<https://devfeed.tech/tags/intrinsic.md>), [learning](<https://devfeed.tech/tags/learning.md>), [model](<https://devfeed.tech/tags/model.md>), [performance](<https://devfeed.tech/tags/performance.md>), [private](<https://devfeed.tech/tags/private.md>), [recall](<https://devfeed.tech/tags/recall.md>), [sample](<https://devfeed.tech/tags/sample.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>), [synthetic-data-generation](<https://devfeed.tech/tags/synthetic-data-generation.md>)

### AI overview

The paper recasts Private Evolution (PE), a differentially private synthetic-data algorithm, as generative model-augmented Wasserstein learning. Its theoretical bounds improve when the generator captures structure in the true distribution; when generated samples lie in the same low-dimensional space, sample complexity depends on intrinsic rather than ambient dimension. The authors also find that standard PE variants may fail to converge on simple well-clustered instances, and propose a geometry-aware version with provable convergence on those instances. Experiments report competitiveness with standard baselines and improved recall.

### Source excerpt

Private Evolution (PE) is a differentially private algorithm for synthetic data generation. While it can be viewed as a Wasserstein learning algorithm, it performs much better in practice than worst-case Wasserstein analyses would predict. We recast PE as generative model-augmented Wasserstein learning. We show theoretically that when we take into account the use of a generative model that is able to capture something about the true distribution, then we can obtain much better performance bounds. For example, if the generator gives samples in the same low-dimensional space as the distribution...

## Complexity of Differentially Private Selection via Federated APIs

DevFeed: [Complexity of Differentially Private Selection via Federated APIs](<https://devfeed.tech/articles/complexity-of-differentially-private-selection-via-federated-apis-145276.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/differentially-private-selection>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Differential Privacy](<https://devfeed.tech/topics/differential-privacy.md>), [computational complexity](<https://devfeed.tech/topics/computational-complexity.md>)

Tags: [academic](<https://devfeed.tech/tags/academic.md>), [aggregation](<https://devfeed.tech/tags/aggregation.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [differential-privacy](<https://devfeed.tech/tags/differential-privacy.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [federated](<https://devfeed.tech/tags/federated.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [frameworks](<https://devfeed.tech/tags/frameworks.md>), [private](<https://devfeed.tech/tags/private.md>), [selection](<https://devfeed.tech/tags/selection.md>), [utility](<https://devfeed.tech/tags/utility.md>)

### AI overview

The excerpt identifies a utility limit for private selection using standard noisy aggregation: it requires Ω(√d/ε) samples. The authors propose adding a primitive that implements a version of the Sparse Vector Technique (SVT), with simple implementations they say are possible under various trust models. Their ε-differentially private selection algorithm has additive error O(log(dn)/ε), matching the central model when n is polynomially bounded in d, and uses an expected O(d log* d) SVT queries. Alongside existing shuffle-model lower bounds, the result establishes an exponential separation between the SVT model and shuffle or aggregation-based models. The excerpt also names Fp loss minimization, heavy-hitter identification, and sparse mean estimation as applications.

### Source excerpt

Differential privacy (DP) has become the standard for privacy-preserving data analysis, yet its practical implementation in federated or distributed settings often faces a significant utility gap compared to the central model. While existing frameworks, including local, shuffle, and secure aggregation models, offer vetted primitives, they can incur high error for important tasks such as selection. Our first contribution is to show that this is inherent in the aggregation model: any private selection algorithm relying on the standard noisy aggregation primitive requires Ω(d/ϵ)Ω(\sqrt{d} /...

## Toward a Theory of Evaluability

DevFeed: [Toward a Theory of Evaluability](<https://devfeed.tech/articles/toward-a-theory-of-evaluability-136573.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/theory-evaluability>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [benchmark overfitting machine learning](<https://devfeed.tech/topics/benchmark-overfitting-machine-learning.md>), [computational complexity](<https://devfeed.tech/topics/computational-complexity.md>)

Tags: [algorithms](<https://devfeed.tech/tags/algorithms.md>), [apple](<https://devfeed.tech/tags/apple.md>), [area](<https://devfeed.tech/tags/area.md>), [assessment](<https://devfeed.tech/tags/assessment.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [foundation-models](<https://devfeed.tech/tags/foundation-models.md>), [generative](<https://devfeed.tech/tags/generative.md>), [lower-bounds](<https://devfeed.tech/tags/lower-bounds.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [models](<https://devfeed.tech/tags/models.md>), [protocols](<https://devfeed.tech/tags/protocols.md>), [safety](<https://devfeed.tech/tags/safety.md>), [test](<https://devfeed.tech/tags/test.md>), [theory](<https://devfeed.tech/tags/theory.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

The paper formalizes model evaluation when test samples may have appeared in training. For evaluators who only observe model outputs on the test set, accurate evaluation is possible if and only if the distribution is close to having small support; the paper also gives tight sample-complexity bounds. With access to additional model queries, evaluation can become more powerful for natural distributions and concept classes, drawing on connections to self-correctors in program testing. The excerpt does not specify the exact bounds or query requirements.

### Source excerpt

Machine learning models, particularly large foundation models, are increasingly evaluated along multiple axes that are relevant for their usage. In addition to accuracy measures appropriate for the specific task of interest, measures capturing safety, calibration and fairness properties are often of interest. The classical approach to these problems assumes that the test set has been held out from the training process, an assumption that is increasingly harder to justify for large foundation models. In this work, we formalize a notion of evaluability without this assumption, and ask if it is...

## Weighted State Souping for Efficient Long-Context Retrieval

DevFeed: [Weighted State Souping for Efficient Long-Context Retrieval](<https://devfeed.tech/articles/weighted-state-souping-for-efficient-long-context-retrieval-129818.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/weighted-state-souping>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [context](<https://devfeed.tech/topics/context.md>), [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-systems](<https://devfeed.tech/tags/ai-systems.md>), [context](<https://devfeed.tech/tags/context.md>), [fast](<https://devfeed.tech/tags/fast.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language](<https://devfeed.tech/tags/language.md>), [like](<https://devfeed.tech/tags/like.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [memory](<https://devfeed.tech/tags/memory.md>), [modeling](<https://devfeed.tech/tags/modeling.md>), [models](<https://devfeed.tech/tags/models.md>), [neurips](<https://devfeed.tech/tags/neurips.md>), [palm](<https://devfeed.tech/tags/palm.md>), [paper](<https://devfeed.tech/tags/paper.md>), [performance](<https://devfeed.tech/tags/performance.md>), [queries](<https://devfeed.tech/tags/queries.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

Weighted state souping addresses the decline of state space models and standard state souping on long, crowded contexts by learning query-dependent weights to combine cached states. The paper reports substantial improvements across single- and multi-hop document question answering, needle-in-a-haystack retrieval, variable tracking, and tool use. It also adapts the method into a sparse variant with near-constant encoding cost as context length grows, while preserving or sometimes improving performance. The authors offer theoretical arguments about why souping works and when it may fail.

### Source excerpt

This paper was accepted at the Workshop on Personalized, Aligned, Long-term Memory for AI Systems (PALM) at NeurIPS 2026. State space models (SSMs) combine fast inference and fixed state size with competitive language-modeling performance. Recently, state souping has been proposed as a procedure whereby an SSM independently encodes chunks of tokens like documents, tool definitions, or memory records. The resulting state encodings may be cached and recombined to cheaply construct a context for downstream queries. However, both state souping and vanilla SSMs deteriorate on long, crowded...

## On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models

DevFeed: [On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models](<https://devfeed.tech/articles/on-the-importance-of-gating-memorization-vs-in-context-learning-in-state-space-models-129817.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/gating-memorization-in-context>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Mamba](<https://devfeed.tech/topics/mamba.md>), [Model Development](<https://devfeed.tech/topics/model-development.md>)

Tags: [context](<https://devfeed.tech/tags/context.md>), [learning](<https://devfeed.tech/tags/learning.md>), [mechanism](<https://devfeed.tech/tags/mechanism.md>), [memory](<https://devfeed.tech/tags/memory.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [sequence](<https://devfeed.tech/tags/sequence.md>), [sequences](<https://devfeed.tech/tags/sequences.md>), [state-space-models](<https://devfeed.tech/tags/state-space-models.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

The study argues that gating shapes both what State Space Models learn and how they generalize. Theory and experiments indicate that gating can lead SSMs to first learn an in-weights memorization strategy, delaying or even preventing correct in-context learning, including in cases where the architecture and its memory capacity pose no fundamental barrier. Gating can also improve generalization to longer sequences. The excerpt presents this tension as a way to understand SSM training and generalization; it does not provide specific experiments or recommendations for changing the gating mechanism.

### Source excerpt

State Space Models (SSMs) have emerged as a compelling alternative to Transformers, enabling sequence modeling with constant memory and linear compute. Although SSMs exhibit reasonable performance and favorable computational characteristics, they continue to lag behind Transformers on tasks that require in-context learning and precise retrieval, slowing their adoption for large-scale language modeling. In this work, we demonstrate that both the success and failure of SSMs in these domains can be explained by studying the role of the gating mechanism, a prevalent component in modern recurrent...

## Evaluating the Evaluators: Investigating LLM Judges for Personalized Writing Style Assessment

DevFeed: [Evaluating the Evaluators: Investigating LLM Judges for Personalized Writing Style Assessment](<https://devfeed.tech/articles/evaluating-the-evaluators-investigating-llm-judges-for-personalized-writing-style-assessment-129816.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/evaluating-evaluators-llm-judges>)

Published: 2026-11-10T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-evaluation](<https://devfeed.tech/tags/ai-evaluation.md>), [annotations](<https://devfeed.tech/tags/annotations.md>), [assessment](<https://devfeed.tech/tags/assessment.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [generator](<https://devfeed.tech/tags/generator.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llms](<https://devfeed.tech/tags/llms.md>), [neurips](<https://devfeed.tech/tags/neurips.md>), [performance](<https://devfeed.tech/tags/performance.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [style](<https://devfeed.tech/tags/style.md>), [workshop](<https://devfeed.tech/tags/workshop.md>), [writing](<https://devfeed.tech/tags/writing.md>)

### AI overview

The paper finds that LLM judges assess personalized writing unevenly: their quality varies by writing task and by whether the criterion is general helpfulness or style personalization, and it correlates only weakly with MMLU ability (Pearson's ρ ≤ 0.2). In an evaluation using text from four generator LLMs across three long-form tasks, human pairwise labels, and judge annotations, judges were more consistent and performant in paired comparisons than when scoring responses independently. The strongest models did not most reliably match human evaluations; the authors hypothesize that they may overreact to subtle details that humans and weaker models overlook.

### Source excerpt

This paper was accepted at the Can We Trust AI Evaluation? Workshop at NeurIPS 2026. We evaluate how good large language model (LLM) judges are for personalized writing tasks. This is important when a user's preferred writing style is known (e.g., "upbeat") and an LLM judge is used to evaluate whether generated text adheres to this preference. We examine how the writing task, the combination of generator and judge LLM, the evaluation objective, and general commonsense and reasoning ability impact LLM judge performance. To this end, we generate text with four LLMs for three long-form writing...

## Faster Retrieval for Better, More Efficient RAG

DevFeed: [Faster Retrieval for Better, More Efficient RAG](<https://devfeed.tech/articles/faster-retrieval-for-better-more-efficient-rag-119799.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/faster-retrieval>)

Published: 2026-11-05T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [AI-driven information retrieval](<https://devfeed.tech/topics/ai-driven-information-retrieval.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>)

Tags: [cost](<https://devfeed.tech/tags/cost.md>), [database](<https://devfeed.tech/tags/database.md>), [exhaustive](<https://devfeed.tech/tags/exhaustive.md>), [inner-product](<https://devfeed.tech/tags/inner-product.md>), [key](<https://devfeed.tech/tags/key.md>), [keys](<https://devfeed.tech/tags/keys.md>), [mips](<https://devfeed.tech/tags/mips.md>), [query](<https://devfeed.tech/tags/query.md>), [question](<https://devfeed.tech/tags/question.md>), [rag](<https://devfeed.tech/tags/rag.md>), [recommendation-systems](<https://devfeed.tech/tags/recommendation-systems.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [search](<https://devfeed.tech/tags/search.md>), [vector](<https://devfeed.tech/tags/vector.md>)

### AI overview

Amortized MIPS trains small neural networks offline for a fixed key database and an expected query distribution, trading upfront work for faster retrieval at inference. SupportNet learns cluster support scores to route queries more accurately than centroid-based IVF routing; on Natural Questions with 128 clusters, a small model reached 95% routing recall about 1.9x faster, while the mid-size model found the correct partition 92.6% of the time when scanning 5% of the database, versus 87.5% for the centroid baseline. KeyNet predicts a likely best-match key and rewrites the query for an existing IVF index. On the same benchmark, the largest KeyNet reached 99% recall in 0.23 ms versus 1.8 ms for raw queries, and scanning 0.001% of the database retrieved the true key up to 96.7% of the time versus 86%. The authors suggest KeyNet when inference speed is the priority and SupportNet when calibrated scores or small-model compute efficiency matter. They report robustness to mild query-distribution shifts; the approach requires offline training for the fixed database and depends on representative queries.

### Source excerpt

Maximum inner product search (MIPS) sits at the core of recommendation systems, search, and retrieval-augmented generation (RAG). Given a query vector, MIPS systems search within a database of millions of candidate key vectors for the one that best matches the query. The original data point associated with that best-matching key is then used to handle the query: in RAG, the document that is best suited to answer the question is typically prefixed to the query. As the number of keys grows to millions and vector dimensions reach thousands, the computational cost of running an exhaustive...

## Attention to Mamba: A Recipe for Cross-Architecture Distillation

DevFeed: [Attention to Mamba: A Recipe for Cross-Architecture Distillation](<https://devfeed.tech/articles/attention-to-mamba-a-recipe-for-cross-architecture-distillation-127730.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/attention-mamba-distillation>)

Published: 2026-11-03T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Mamba](<https://devfeed.tech/topics/mamba.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [Transformer](<https://devfeed.tech/topics/transformer.md>)

Tags: [algorithms](<https://devfeed.tech/tags/algorithms.md>), [approach](<https://devfeed.tech/tags/approach.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [attention](<https://devfeed.tech/tags/attention.md>), [context](<https://devfeed.tech/tags/context.md>), [cross](<https://devfeed.tech/tags/cross.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [icml](<https://devfeed.tech/tags/icml.md>), [kernel](<https://devfeed.tech/tags/kernel.md>), [language](<https://devfeed.tech/tags/language.md>), [mamba](<https://devfeed.tech/tags/mamba.md>), [memory](<https://devfeed.tech/tags/memory.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [recipe](<https://devfeed.tech/tags/recipe.md>), [state-space-models](<https://devfeed.tech/tags/state-space-models.md>), [train](<https://devfeed.tech/tags/train.md>), [transformer](<https://devfeed.tech/tags/transformer.md>), [transformer-models](<https://devfeed.tech/tags/transformer-models.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

The paper proposes a two-stage recipe for transferring a pretrained Transformer into an Attention-free Mamba model. It first distills the Transformer into linearized Attention using an adaptation of the kernel trick, then distills that model into an adapted Mamba initialized to support the transfer. The authors report that on Pythia-1B, the resulting model reaches perplexity 14.01 versus 13.86 for the teacher and retains most of its performance on downstream tasks, though gaps remain on tasks relying more on in-context recall. The paper also reports ablations on sequence mixer architecture, model size, distillation token count, and how tokens are allocated between stages.

### Source excerpt

State Space Models (SSMs) such as Mamba have become a popular alternative to Transformer models, due to their reduced memory consumption and higher throughput at generation compared to their Attention-based counterparts. On the other hand, the community has built up a considerable body of knowledge on how to train Transformers, and many pretrained Transformer models are readily available. To facilitate the adoption of SSMs while leveraging existing pretrained Transformers, we aim to identify an effective recipe to distill an Attention-based model into a Mamba-like architecture. In prior work...

## Plan or Retrieve First? A Scale Study of Tool Discovery Architectures

DevFeed: [Plan or Retrieve First? A Scale Study of Tool Discovery Architectures](<https://devfeed.tech/articles/plan-or-retrieve-first-a-scale-study-of-tool-discovery-architectures-88212.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/plan-retrieve-tool-discovery>)

Published: 2026-10-15T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [AI search](<https://devfeed.tech/topics/ai-search.md>), [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Vectors & LLMs](<https://devfeed.tech/topics/vectors-llms.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [planning](<https://devfeed.tech/tags/planning.md>), [queries](<https://devfeed.tech/tags/queries.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [search](<https://devfeed.tech/tags/search.md>)

### AI overview

This paper compares Retrieve-then-Plan and Plan-then-Retrieve approaches for finding APIs for tool-augmented LLM agents. Retrieve-then-Plan performs better for single-tool queries, while Plan-then-Retrieve performs better for multi-tool queries; the latter's added discovery call roughly doubles inference cost. The findings motivate complexity-adaptive routing based on the top retrieval score.

### Source excerpt

Tool-augmented LLM agents discover relevant APIs from large registries before planning execution, yet finite context windows prohibit injecting all tool definitions into a single prompt. The dominant approach, which we term Retrieve-then-Plan (RtP), embeds the user query and retrieves similar tools via approximate nearest-neighbor search. We study an alternative Plan-then-Retrieve (PtR) paradigm in which the LLM first generates Abstract Tool Stubs (hypothesized function signatures describing needed capabilities) which then serve as retrieval queries against the registry. Evaluating on...

## Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio

DevFeed: [Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio](<https://devfeed.tech/articles/local-and-distributed-large-llm-inference-with-mlx-and-rdma-on-mac-studio-88211.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/mlx-distributed-inference>)

Published: 2026-10-15T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [MLX](<https://devfeed.tech/topics/mlx.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [rdma](<https://devfeed.tech/topics/rdma.md>), [On-device AI](<https://devfeed.tech/topics/on-device-ai.md>)

Tags: [apple-silicon](<https://devfeed.tech/tags/apple-silicon.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llm-inference](<https://devfeed.tech/tags/llm-inference.md>), [mac](<https://devfeed.tech/tags/mac.md>), [mac-studio](<https://devfeed.tech/tags/mac-studio.md>), [memory-architecture](<https://devfeed.tech/tags/memory-architecture.md>), [rdma](<https://devfeed.tech/tags/rdma.md>), [unified-memory-architecture](<https://devfeed.tech/tags/unified-memory-architecture.md>)

### AI overview

The article benchmarks local and distributed LLM inference using MLX on Apple silicon Mac Studio systems. It describes pipeline and tensor parallelization over RDMA via Thunderbolt, compares model performance across M3 Ultra and M5 Ultra machines, and explains how this setup can run large models locally for agentic AI tools.

### Source excerpt

MLX on Mac with Apple silicon lets AI researchers and developers experiment locally with the latest models and methods in a private environment. The open-source array framework leverages the unified memory architecture of Apple silicon, enabling anyone to run new inference and fine-tuning methods on their own hardware and explore other AI techniques. In a previous post, we highlighted the advantages of the M5 architecture and its GPU Neural Accelerators.

## Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents

DevFeed: [Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents](<https://devfeed.tech/articles/selection-based-structured-reasoning-toward-efficient-multimodal-search-agents-88213.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/selection-structured-reasoning>)

Published: 2026-10-15T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Parallel reasoning](<https://devfeed.tech/topics/parallel-reasoning.md>)

Tags: [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [parallel](<https://devfeed.tech/tags/parallel.md>), [performance](<https://devfeed.tech/tags/performance.md>), [ssr](<https://devfeed.tech/tags/ssr.md>), [structured](<https://devfeed.tech/tags/structured.md>)

### AI overview

Selection-based Structured Reasoning (SSR) replaces open-ended reasoning in multimodal agents with reusable natural-language candidates selected according to context. On seven multimodal search benchmarks, it reduced per-turn reasoning latency by over 90% and total per-question inference latency by 28-54%, while maintaining competitive task performance.

### Source excerpt

Multimodal agents commonly generate freeform reasoning before each action. For small models, limited model capacity can result in lengthy reasoning that provides little useful guidance for action generation while incurring substantial inference cost. To address this challenge, we introduce Selection-based Structured Reasoning (SSR), a framework that reformulates reasoning as selection instead of open-ended generation. SSR represents recurring high-level reasoning as pre-specified, reusable natural-language candidates. At each turn, the model selects from the reasoning candidates based on their...

## Code-Switching Detection via Language-Selective Dimensions in Speech Foundation Models

DevFeed: [Code-Switching Detection via Language-Selective Dimensions in Speech Foundation Models](<https://devfeed.tech/articles/code-switching-detection-via-language-selective-dimensions-in-speech-foundation-models-87264.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/code-switching-detection>)

Published: 2026-10-14T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [asr](<https://devfeed.tech/topics/asr.md>), [Whisper](<https://devfeed.tech/topics/whisper.md>), [Internationalization (i18n)](<https://devfeed.tech/topics/i18n.md>)

Tags: [asr](<https://devfeed.tech/tags/asr.md>), [detection](<https://devfeed.tech/tags/detection.md>), [foundation-models](<https://devfeed.tech/tags/foundation-models.md>), [generation](<https://devfeed.tech/tags/generation.md>), [speech](<https://devfeed.tech/tags/speech.md>)

### AI overview

The paper presents a lightweight method for detecting code-switching directly from speech. It aggregates frame-level activations from language-selective dimensions in pretrained speech encoders, and reports that the Whisper-small encoder achieves a mean F1 of 0.798 across 14 language pairs in CS-FLEURS. The findings suggest that pretrained speech encoders contain information useful for identifying code-switched utterances and building more representative multilingual corpora.

### Source excerpt

Code-switching (CS), the alternation between languages within an utterance, is common in multilingual speech but underrepresented in pretraining corpora. Increasing CS coverage requires finding relevant utterances in large speech collections so that they can be included in training data. One way to find such utterances is to transcribe the speech and detect language alternation in the text, but this depends on obtaining accurate multilingual transcriptions. We propose a lightweight method that detects CS directly from speech. It uses frame-level language information in pretrained speech...

## The Design Space of Tri-Modal Masked Diffusion Models

DevFeed: [The Design Space of Tri-Modal Masked Diffusion Models](<https://devfeed.tech/articles/the-design-space-of-tri-modal-masked-diffusion-models-87265.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/tri-modal-masked-diffusion>)

Published: 2026-10-14T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [scaling laws](<https://devfeed.tech/topics/scaling-laws.md>), [Chinchilla scaling law](<https://devfeed.tech/topics/chinchilla-scaling-law.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>)

Tags: [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>)

### AI overview

The paper introduces tri-modal masked diffusion models pretrained from scratch on text, image-text, and audio-text data. It analyzes multimodal scaling laws, modality mixing, noise schedules, and batch size, and proposes an SDE-based reparameterization that separates physical from logical batch size. A preliminary 3B model is evaluated on text generation, text-to-image, and text-to-speech tasks.

### Source excerpt

Discrete diffusion models have emerged as strong alternatives to autoregressive language models, with recent work initializing and finetuning a base unimodal model for bi-modal generation. Diverging from previous approaches, we introduce the first tri-modal Masked Diffusion Models (MDM) pretrained from scratch on text, image-text, and audio-text data. We systematically analyze multimodal scaling laws, modality mixing ratios, noise schedules, batch-size effects and provide optimized inference sampling defaults. Our batch-size analysis yields a novel stochastic differential equation (SDE) based...

## Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings

DevFeed: [Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings](<https://devfeed.tech/articles/mix-don-t-tune-bilingual-pre-training-outperforms-hyperparameter-search-in-data-constrained-settings-85859.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/mix-dont-tune>)

Published: 2026-10-13T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Hyperparameter optimization](<https://devfeed.tech/topics/hyperparameter-optimization.md>), [scaling laws](<https://devfeed.tech/topics/scaling-laws.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>)

Tags: [applications](<https://devfeed.tech/tags/applications.md>), [ratio](<https://devfeed.tech/tags/ratio.md>), [training](<https://devfeed.tech/tags/training.md>), [training-data](<https://devfeed.tech/tags/training-data.md>)

### AI overview

The paper compares bilingual data mixing with hyperparameter tuning for language model pre-training when target-language data is limited. Across experiments using Arabic as the target language and English as auxiliary data, mixing produced larger gains in validation loss and downstream task accuracy, with benefits increasing at larger model sizes. The authors recommend prioritizing the mixing ratio and transferring hyperparameters from a small proxy model.

### Source excerpt

For most languages of the world, language model pre-training operates in a data-constrained regime where models must repeat their training data many times, degrading generalization. Two remedies exist: aggressive hyperparameter tuning such as high weight decay, and mixing in data from a high-resource auxiliary language to directly aid the low-resource target. While hyperparameter tuning regularizes the model by shrinking weights to restrict network capacity, auxiliary data mixing uses a tunable mixing ratio to expand the training distribution and diversify the training signal with new...

## Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged

DevFeed: [Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged](<https://devfeed.tech/articles/large-language-models-for-machine-translation-quality-annotation-humans-and-models-are-both-challenged-85612.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/llm-translation-quality-annotation>)

Published: 2026-10-12T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [machine-translation](<https://devfeed.tech/topics/machine-translation.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>)

Tags: [evaluation](<https://devfeed.tech/tags/evaluation.md>), [human-judgment](<https://devfeed.tech/tags/human-judgment.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [llms](<https://devfeed.tech/tags/llms.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [machine-translation](<https://devfeed.tech/tags/machine-translation.md>), [pairs](<https://devfeed.tech/tags/pairs.md>), [performance](<https://devfeed.tech/tags/performance.md>), [research](<https://devfeed.tech/tags/research.md>)

### AI overview

This paper evaluates large language models (LLMs) as judges of machine translation quality, comparing their agreement with human annotators across two annotation schemes, 70 language pairs, and public datasets. Agreement varies substantially by scheme, language pair, and domain, and remains unreliable in most settings. The study also identifies challenges for both human and LLM annotators and suggests human-LLM collaborative annotation pipelines as a possible improvement.

### Source excerpt

Large Language Models (LLMs) are considered to be a more efficient and cost-effective alternative to human judgment for Machine Translation (MT) evaluation. With MT evaluation spanning a large number of language pairs, domains and levels of annotation granularity, LLMs must be thoroughly evaluated across these dimensions before being reliably used as alternatives to human evaluation. In this paper, we evaluate the performance of LLMs for two prominent MT quality evaluation schemes: Multidimensional Quality Metrics (MQM) and Error Span Annotation (ESA) by comparing their agreement with human...

## Self-Compacting Language Model Agents

DevFeed: [Self-Compacting Language Model Agents](<https://devfeed.tech/articles/self-compacting-language-model-agents-85613.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/self-compacting-language-agents>)

Published: 2026-10-12T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [compaction](<https://devfeed.tech/tags/compaction.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llm](<https://devfeed.tech/tags/llm.md>)

### AI overview

SELFCOMPACT gives language model agents a compaction tool and a rubric for deciding when to summarize long traces. Across six benchmarks and seven models, it matched or exceeded fixed-interval summarization at lower token cost, improving over no summarization by up to 18.1 points on math and 5-9 points on agentic search with 30-70% lower per-question cost.

### Source excerpt

Long agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it with fixed-interval compaction triggered at a token threshold. Such triggers pay no heed to trajectory structure, risking discard of partial results mid-derivation or mid-search. We propose SELFCOMPACT, a scaffold that allows the model itself to decide when and how to compact. Specifically, it pairs two inference-time elements: (i) a compaction tool the model invokes to summarize the accumulated...

## Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

DevFeed: [Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices](<https://devfeed.tech/articles/minions-a-dynamic-and-distributed-virtual-gpu-cloud-built-on-idle-apple-silicon-devices-85611.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/distributed-gpu-cloud>)

Published: 2026-10-12T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [distributed-training](<https://devfeed.tech/topics/distributed-training.md>), [Multi-GPU](<https://devfeed.tech/topics/multi-gpu.md>), [On-device AI](<https://devfeed.tech/topics/on-device-ai.md>)

Tags: [apple-silicon](<https://devfeed.tech/tags/apple-silicon.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [inference](<https://devfeed.tech/tags/inference.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [macos](<https://devfeed.tech/tags/macos.md>)

### AI overview

The paper presents Minions, a distributed virtual GPU cloud that combines idle Apple Silicon GPUs across macOS devices to run large-scale machine learning inference. It describes flexible task execution, fault tolerance and resource use for teams facing constrained centralized GPU resources.

### Source excerpt

The increasing demand for GPU-accelerated machine learning inference has exposed limitations in centralized GPU infrastructures, including high cost, limited elasticity, and long wait times under resource constraints. At the same time, modern Apple Silicon-based macOS devices are equipped with powerful GPUs that remain largely idle during normal usage. This paper presents Minions, a dynamic and distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to execute large-scale machine learning inference workloads. Minions enables teams to leverage existing...

## Normalizing Trajectory Models

DevFeed: [Normalizing Trajectory Models](<https://devfeed.tech/articles/normalizing-trajectory-models-77052.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/normalizing-trajectory-models>)

Published: 2026-10-08T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [diffusion-transformers](<https://devfeed.tech/topics/diffusion-transformers.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [stable-diffusion](<https://devfeed.tech/topics/stable-diffusion.md>)

Tags: [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diffusion-models](<https://devfeed.tech/tags/diffusion-models.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [flow-matching](<https://devfeed.tech/tags/flow-matching.md>), [image-generation](<https://devfeed.tech/tags/image-generation.md>), [text-to-image](<https://devfeed.tech/tags/text-to-image.md>)

### AI overview

The paper introduces Normalizing Trajectory Models, which represent each reverse diffusion step as a conditional normalizing flow trained with exact likelihood. A self-distilled denoiser generates samples in four steps, matching or outperforming strong text-to-image baselines while retaining exact likelihood over the trajectory.

### Source excerpt

Diffusion-based 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 in the process. We introduce Normalizing Trajectory Models (NTM), which 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...

## RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation

DevFeed: [RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation](<https://devfeed.tech/articles/rised-rubrics-for-agentic-multi-environment-selection-and-self-distillation-65884.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/rised-multi-environment-selection>)

Published: 2026-10-06T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Loop Engineering](<https://devfeed.tech/topics/loop-engineering.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [multi-environment](<https://devfeed.tech/tags/multi-environment.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

RISED uses rollout behavior rubrics to select training data across interactive environments and provide supervision for LLM agents through self-distillation. The paper reports the highest mean pass rate across environments among the evaluated approaches and ranks first or second in each individual environment.

### Source excerpt

Training a single LLM agent jointly across diverse interactive environments has attracted increasing attention as a route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, without explicitly considering relationships between current rollouts across environments for prompt-group selection. Meanwhile, as environments are learned at different rates, all-failure and all-success rollout groups can coexist within a batch, leaving those data without group-relative reward signals. Both...

## Negotiating Ontological Boundaries in User-Authored Personal Sensing Systems

DevFeed: [Negotiating Ontological Boundaries in User-Authored Personal Sensing Systems](<https://devfeed.tech/articles/negotiating-ontological-boundaries-in-user-authored-personal-sensing-systems-65164.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/ontological-boundary-negotiation>)

Published: 2026-10-05T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [User Experience](<https://devfeed.tech/topics/user-experience.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [design](<https://devfeed.tech/tags/design.md>), [experience](<https://devfeed.tech/tags/experience.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [wizard](<https://devfeed.tech/tags/wizard.md>)

### AI overview

This paper examines how people negotiate the categories and assumptions built into personal sensing systems. It describes two open-ended probes that let participants experience training a personalized machine learning system on phenomena they define, then identifies four areas of negotiation: what counts as a phenomenon, how subjects relate, what is signal or noise, and whether data is objective.

### Source excerpt

Designed artifacts are ontological, shaping, and at times limiting, what becomes possible or imaginable. One path toward mitigating such foreclosures is giving people power over how systems are designed and built. Despite decades of scholarship around systems that enable such authorship, these systems are often evaluated on whether or not they are usable, useful, or technically feasible, leaving questions of ontological boundary negotiation, unexamined. We design two open-ended probes that utilize a Wizard of Oz technique to enable the experience of training a personalized machine learning...

## Language Discrimination Improves Linguistic Learning in Multilingual Speech Models

DevFeed: [Language Discrimination Improves Linguistic Learning in Multilingual Speech Models](<https://devfeed.tech/articles/language-discrimination-improves-linguistic-learning-in-multilingual-speech-models-64220.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/language-discrimination-multilingual-learning>)

Published: 2026-10-02T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [Whisper](<https://devfeed.tech/topics/whisper.md>)

Tags: [language](<https://devfeed.tech/tags/language.md>), [models](<https://devfeed.tech/tags/models.md>), [pretraining](<https://devfeed.tech/tags/pretraining.md>), [self-supervised-learning](<https://devfeed.tech/tags/self-supervised-learning.md>), [speech](<https://devfeed.tech/tags/speech.md>), [speech-processing](<https://devfeed.tech/tags/speech-processing.md>)

### AI overview

The paper studies how encouraging language discrimination during pretraining affects multilingual self-supervised speech models. In a controlled English/French HuBERT setting, an auxiliary language classifier and per-language k-means targets reduced phone discrimination error and improved lexical and prosodic scores, in some measures approaching or exceeding monolingual results. Introducing language discrimination in the first training iteration produced the strongest gains, while later or repeated interventions brought smaller improvements and greater language-wise segregation.

### Source excerpt

Multilingual self-supervised speech models can benefit from sharing information across languages, but under a matched total pretraining data budget they still fall short of monolingual models. We show that strengthening the model's ability to discriminate languages during pretraining reduces and, on some measures, closes this multilingual gap on continuous phonetic and higher-level linguistic measures, while preserving substantial cross-language sharing. Using a controlled English/French HuBERT setting, we test two interventions which strengthen language discrimination: an auxiliary language...

## Limits of Confidence in Discrete Diffusion

DevFeed: [Limits of Confidence in Discrete Diffusion](<https://devfeed.tech/articles/limits-of-confidence-in-diffusion-64221.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/limits-confidence-diffusion>)

Published: 2026-10-02T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [diffusers](<https://devfeed.tech/topics/diffusers.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>)

Tags: [algorithms](<https://devfeed.tech/tags/algorithms.md>), [data](<https://devfeed.tech/tags/data.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [limits](<https://devfeed.tech/tags/limits.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [model](<https://devfeed.tech/tags/model.md>), [noise](<https://devfeed.tech/tags/noise.md>), [ranking](<https://devfeed.tech/tags/ranking.md>), [research](<https://devfeed.tech/tags/research.md>), [transformer](<https://devfeed.tech/tags/transformer.md>), [vision](<https://devfeed.tech/tags/vision.md>)

### AI overview

This paper examines how confidence-based sampling in discrete diffusion handles dependencies among token positions. It argues that sampling steps match the training distribution only when positions written together are conditionally independent given already fixed tokens, and reports a synthetic-task result showing that confidence ranking writes dependent groups of undetermined positions.

### Source excerpt

Discrete diffusion, including remasking and uniform-state samplers, generate a sequence by writing multiple token positions per step, drawing each from a per-position distribution and choosing which positions to write from those same distributions. For domains of general interest (pixels, phonemes, or words) there are inherent dependencies between tokens. We show that a step matches the training distribution only when the positions it writes are conditionally independent given the tokens already fixed, that no product of per-position distributions can match a dependent group, and that...

## RLTL;DR: Self-Improvement by Internalizing Self-Generated Feedback

DevFeed: [RLTL;DR: Self-Improvement by Internalizing Self-Generated Feedback](<https://devfeed.tech/articles/rltl-dr-self-improvement-by-internalizing-self-generated-feedback-63573.md>)

Original publisher: [Read original article](<https://machinelearning.apple.com/research/rltl-dr-self-improvement>)

Published: 2026-10-01T00:00:00Z

Content type: article

Language: en

Sources: [Apple Machine Learning Research](<https://devfeed.tech/sources/apple-machine-learning-research.md>)

Topics: [rlvr](<https://devfeed.tech/topics/rlvr.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>)

Tags: [coding](<https://devfeed.tech/tags/coding.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [rlvr](<https://devfeed.tech/tags/rlvr.md>), [tasks](<https://devfeed.tech/tags/tasks.md>)

### AI overview

The paper presents RLTL;DR, a method for training an agent to turn verifier feedback after failed attempts into short insights that guide later attempts and can be internalized. On challenging tool-calling and coding datasets, it reports Pass@1 rates of 14-31% with insights in context during training and 12-13% without insights at evaluation. A related SFTL;DR approach trained on 4,000 task-insight pairs recovered almost the full performance of RLTL;DR and classical supervised fine-tuning on full rollouts.

### Source excerpt

The common paradigm of reinforcement learning with verifiable rewards (RLVR) is to let agents make multiple attempts at a task, and optimize towards the successful ones. This becomes problematic in the realms of self-improvement, where tasks are so difficult that the agent has a low or even no chance of success, and where there are no teacher models or example solutions to distill from. In this paper, we introduce RLTL;DR. After each failed attempt, we show the policy the verifier outputs and let it write its own feedback, in the form of a single TL;DR insight. The next rollout is conditioned...

[Next page](<https://devfeed.tech/sources/apple-machine-learning-research.md?cursor=WyIyMDI2LTEwLTAxVDAwOjAwOjAwKzAwOjAwIiwgIjg5OTc1NGJjLTAzMDQtNDVkYS1hMWZhLWNhYjNjMjU3NDg0MSJd>)