# 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. ## How Value Induction Reshapes LLM Behaviour DevFeed: [How Value Induction Reshapes LLM Behaviour]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [LLM Techniques](), [LLM evaluation / benchmarking]() Tags: [complex](), [effect](), [language-models](), [llm](), [safety](), [values]() ### AI overview This research examines how inducing selected values in conversational large language models affects other values, safety, anthropomorphic language, and question-answering benchmark performance. It finds that value induction can affect related and contrastive values, positive values can increase safety, and all tested values increased anthropomorphic, validating, and sycophantic language. ### Source excerpt Conversational Large Language Models are post-trained on language that expresses specific behavioural traits, such as curiosity, open-mindedness, and empathy, and values, such as helpfulness, harmlessness, and honesty. This is done to increase utility, ensure safety, and improve the experience of the people interacting with the model. However, values are complex and inter-related - inducing one could modify behaviour on another. Further, inducing certain values can make models more addictive or sycophantic through language used in the generations, with a potential detrimental effect on the... ## Shared Selective Persistent Memory for Agentic LLM Systems DevFeed: [Shared Selective Persistent Memory for Agentic LLM Systems]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [agent observability](), [Framework](), [Code](), [configuration](), [Access Control](), [Git](), [Model Context Protocol (MCP)](), [CSV](), [SQL](), [Grafana]() Tags: [access-control](), [agentic](), [agents](), [code](), [configuration](), [csv](), [dashboards](), [git](), [llm](), [mcp](), [memory](), [platform](), [replication](), [sql]() ### AI overview This research introduces shared selective persistent memory for agentic LLM systems. The architecture retains reusable task specifications, data schemas, tool configurations, and output constraints while discarding session-specific reasoning traces. Shared workspaces support role-based collaborative reuse, and experiments report higher task completion than no memory or full-history persistence, along with zero-token data refresh and lower token costs. ### Source excerpt Agentic LLM systems that generate code through multi-turn tool use face a fundamental context problem: each session starts from zero, discarding the configuration choices, domain constraints, data schemas, and tool-use patterns that made previous sessions productive. Naively persisting entire conversation histories is both token-inefficient and counterproductive--irrelevant context degrades generation quality. We introduce shared selective persistent memory, a memory architecture for agentic systems that identifies and retains four categories of reusable context--task specifications, data... ## Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation DevFeed: [Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [text-generation](), [Algorithms](), [Language models](), [Inference]() Tags: [inference](), [language-models](), [performance](), [perplexity](), [research](), [text-generation](), [tokens]() ### AI overview The article introduces Trajectory-Shaped Discrete Flow Matching, a training method that guides intermediate trajectory decisions with an energy-based coherence measure. The authors argue that poor distillation trajectories, rather than insufficient student capacity, limit few-step generation. On a 170M-parameter language-modeling task, an 8-step student reportedly achieves lower perplexity than a 1,024-step teacher while reducing inference steps. ### Source excerpt Discrete flow matching generates text by iteratively transforming noise tokens into coherent language, but may require hundreds of forward passes. Distillation uses the multi-step trajectory to train a student to reproduce the process in a few steps. When the student underperforms, the usual explanation is insufficient capacity. We argue the opposite: the trajectory is the bottleneck, not the student. Each training trajectory is built through a chain of blind stochastic jumps with no evaluation of sequence quality; a single bad decision at an early midpoint propagates through subsequent steps... ## DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models DevFeed: [DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [grpo](), [Reinforcement learning](), [large-language-models](), [LLM Techniques](), [Code]() Tags: [code-generation](), [denoising](), [diffusion](), [grpo](), [json-schema](), [optimization](), [reinforcement-learning]() ### AI overview The article introduces DACA-GRPO, a denoising-aware enhancement for GRPO-style reinforcement-learning trainers used with diffusion language models. It uses per-token denoising progress scores and stratified masking likelihood to improve credit assignment and reduce bias in likelihood estimates. Across seven benchmarks, the method reports gains in mathematical reasoning, code generation, constraint satisfaction, and JSON Schema adherence. ### Source excerpt Diffusion large language models are a compelling alternative to autoregressive models, yet existing RL methods for diffusion treat all denoising steps as equally important and rely on biased, high-variance likelihood estimates. We identify two fundamental weaknesses: the absence of temporal credit assignment across the denoising trajectory, and the systematic bias of mean-field likelihood estimates used for policy optimization. To address these, we propose Denoising-Aware Credit Assignment for GRPO (DACA-GRPO), a lightweight, plug-and-play enhancement for any GRPO-style trainer. DACA-GRPO... ## Glyph: A Multi-Strategy Agentic System for Column Description and Sensitivity-Ontology Tagging of Enterprise Data Catalogs DevFeed: [Glyph: A Multi-Strategy Agentic System for Column Description and Sensitivity-Ontology Tagging of Enterprise Data Catalogs]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [data](), [Large Language Model](), [Code](), [Retrieval-Augmented Generation](), [Access Control](), [Graphs](), [GitHub]() Tags: [agents](), [classification](), [code](), [data](), [documentation](), [governance](), [production](), [provenance](), [retrieval](), [vector-database]() ### AI overview Glyph is a production system for generating column descriptions and assigning sensitivity-ontology labels in enterprise data catalogs. It uses cooperating LLM agents, source-code-grounded retrieval, parallel tagging strategies, vector-based metadata matching, and ranked-output fusion to support auditable cataloging. ### Source excerpt Enterprise data lakes accumulate tables faster than human stewards can document or classify them, leaving columns with missing descriptions and unassigned governance labels. This documentation debt undermines data discovery, access control, and regulatory compliance. We present Glyph, a production system that frames two coupled problems, column description generation and column type annotation for data classification, as cooperating LLM agents orchestrated as stateful graphs. The Descriptor grounds generation in the pipeline source code that produces each column, retrieved on demand from an... ## DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation DevFeed: [DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation]() Original publisher: [Read original article]() Published: 2026-09-11T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [AI, ML & Data Engineering](), [Language models]() Tags: [evaluation](), [experiments](), [framework](), [llm](), [metrics]() ### AI overview DiscoSign is an LLM-based framework for translating text to sign-language gloss while preserving discourse-level coherence. It targets spatial coreference, Question-Answer Clauses, and consistent English-concept-to-ASL-sign mappings, with evaluation metrics for these dimensions. ### Source excerpt Sign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to sign language comprehension. We introduce DiscoSign, a computational approach for discourse-aware text to sign language gloss translation grounded in linguistic research. We address three key phenomena within our modular Large Language Model (LLM)-based translation framework: (i) spatial coreference resolution, where entities maintain consistent spatial locations throughout discourse; (ii) Question-Answer Clauses (QACs), pseudocleft structures serving... ## Putting Captions to the Test: Evaluating Video Caption Quality through Multiple-Choice Question Answering DevFeed: [Putting Captions to the Test: Evaluating Video Caption Quality through Multiple-Choice Question Answering]() Original publisher: [Read original article]() Published: 2026-09-11T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [LLM evaluation / benchmarking](), [Ground truth / benchmark quality](), [Hallucination detection]() Tags: [benchmark](), [computer-vision](), [evaluation](), [metrics](), [multimodal](), [research](), [video]() ### AI overview The article introduces CapQuiz, a reference-free benchmark for evaluating video-caption quality through human-verified multiple-choice questions. It also proposes CapF1, combining factuality and visual-information coverage, and reports stronger correlation with human judgments than existing metrics. ### Source excerpt Evaluating video captioning remains a critical challenge for Visual Large Language Models (VLLMs). Existing metrics primarily rely on matching generated text against ground-truth references. This paradigm suffers from the "one-to-many" nature of video description, where high-quality captions are often penalized for lexical mismatches or valid shifts in visual focus. Furthermore, such assessments are typically one-dimensional, failing to provide a fine-grained analysis of caption quality. To address this, we redefine caption quality via information fidelity: A caption must maximize the coverage... ## SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign DevFeed: [SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign]() Original publisher: [Read original article]() Published: 2026-09-11T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [multimodal](), [Transformer](), [datasets]() Tags: [benchmarks](), [drug-discovery](), [generation](), [model](), [multimodal](), [training]() ### AI overview SimpleDesign is a single-stage, end-to-end multimodal generative model for jointly designing protein sequences and three-dimensional structures. It uses Transformer-based multimodal backbones, trains directly in data space on more than 2 million sequence-structure pairs, and achieves competitive results on co-design and unconditional generation benchmarks. ### Source excerpt Proteins are fundamental to biological processes, with their function determined by the complex interplay between the amino acid sequence and the three-dimensional structure. Developing generative models capable of understanding this intrinsically multi-modal relationship is crucial for fields like drug discovery and protein engineering. Existing models often rely on a multi-stage training process where autoencoders that tokenize data into latent representations are trained in a first stage. Secondly, a generative model is trained on the latent representation of the autoencoder(s), i.e... ## REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs DevFeed: [REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs]() Original publisher: [Read original article]() Published: 2026-09-02T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [World models](), [Large Language Model]() Tags: [models](), [skills](), [training](), [world-model]() ### AI overview REFACTOR-VLA is a vision-language-action system that learns reusable typed motor-program skills through alternating world-model-based clustering and policy optimization. On LIBERO, the article reports that scaling the world model reduced performance across all four benchmark suites, while an InfoNCE auxiliary loss improved skill clustering. ### Source excerpt Most current vision-language-action (VLA) models--such as OpenVLA, π0, RT-2, and RDT-1B--are "monolithic." This means they generate raw motor commands or very short sequences of actions, without organizing behaviors into reusable, well-defined abstractions. As a result, these models perform poorly on long-horizon (multi-step) tasks, and it's difficult to interpret what they have learned. Existing approaches for discovering skills often avoid the core problem of deciding when two action sequences are "behaviorally equivalent." For example, AtomicVLA and AtomSkill group action sequences by... ## LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs' Probabilistic Beliefs DevFeed: [LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs' Probabilistic Beliefs]() Original publisher: [Read original article]() Published: 2026-08-28T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [Uncertainty quantification LLMs](), [AI Inference](), [Data Science]() Tags: [ai](), [data-science](), [diagnostics](), [llms](), [research]() ### AI overview This research examines whether large language models update probabilistic beliefs in accordance with Bayes' rule. It introduces the information processing gap to quantify deviations from Bayesian updates, compares evidence-integration approaches, and finds that heuristic, non-Bayesian updates can outperform exact Bayesian updates on downstream tasks. ### Source excerpt Modern AI systems are being deployed in complex domains such as medicine, science, and law, where there is often not a single correct answer given the observed evidence. Such systems must be able to represent and update uncertain beliefs about the world as new evidence arrives to make rational decisions. We introduce the novel technique of studying LLMs as information processing rules and utilize the information processing gap--the deviation from Bayes updates--to study the internal (in)consistencies of how LLMs update their probabilistic beliefs from evidence. Our extensive experiments evaluate... ## Agent Seer: Synthesizing Scenarios from Specification Understanding DevFeed: [Agent Seer: Synthesizing Scenarios from Specification Understanding]() Original publisher: [Read original article]() Published: 2026-08-28T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [AI Bots](), [ai observability]() Tags: [ai-agents](), [evaluation](), [mcp](), [model-context-protocol](), [synthetic](), [tool](), [tools]() ### AI overview Agent Seer generates realistic, multi-turn evaluation scenarios for tool-using AI agents from a single MCP specification, without examples, live tool access, or domain-specific tuning. It enriches schemas, creates synthetic tool outputs, and evaluates tool-calling correctness and conversational coherence. ### Source excerpt Evaluating AI agents that use external tools requires realistic test scenarios that capture how practitioners compose tools and iterate across conversation turns. Constructing such scenarios by hand demands deep domain expertise, does not scale across tool ecosystems, and produces static benchmarks that cannot track evolving APIs. We observe that tool specifications--function names, natural-language descriptions, and typed parameter schemas--already encode sufficient semantic information to synthesize realistic evaluation scenarios without manual curation or live tool execution. Agent Seer... ## From Preferences to Principles: Rubric-Based Alignment for Grounded Knowledge Answers DevFeed: [From Preferences to Principles: Rubric-Based Alignment for Grounded Knowledge Answers]() Original publisher: [Read original article]() Published: 2026-08-27T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [LLM evaluation / benchmarking](), [Natural language processing](), [datasets]() Tags: [datasets](), [evaluation](), [models](), [post-training](), [research](), [responses](), [search]() ### AI overview The article introduces a rubric-based reward framework for grounded open-domain question answering. It generates query-specific rubrics from retrieved evidence and decomposes them into quality dimensions for fine-grained post-training supervision. The method improves composition, grounding, and instruction-following results over the stated baselines and evaluation datasets. ### Source excerpt Designing effective reward signals for open-domain question answering is challenging because high-quality responses must simultaneously satisfy multiple aspects of answer quality that are difficult to capture with a holistic scalar objective. We introduce a rubric-based reward framework that generates query-specific rubrics grounded in retrieved evidence and decomposed into multiple quality dimensions, providing fine-grained supervision during post-training. Averaged across three evaluation axes (composition, grounding, and instruction-following), our approach improves over the... ## IDEA Prune: An Integrated Enlarge-and-Prune Pipeline in Generative Language Model Pretraining DevFeed: [IDEA Prune: An Integrated Enlarge-and-Prune Pipeline in Generative Language Model Pretraining]() Original publisher: [Read original article]() Published: 2026-08-26T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Language models](), [compression and generalization](), [Training AI Models](), [Inference Performance]() Tags: [compression](), [generative](), [inference](), [language-models](), [large-language-models](), [model](), [training]() ### AI overview This paper presents IDEA Prune, an integrated enlarge-and-prune pipeline for generative language model pretraining. It combines enlarged-model training, iterative structured pruning, and recovery under one cosine annealing learning-rate schedule, with experiments compressing 2.8B models to 1.3B. ### Source excerpt Recent advancements in large language models have intensified the need for efficient and deployable models within limited inference budgets. Structured pruning pipelines have shown promise in token efficiency compared to training target-size models from scratch. In this paper, we advocate incorporating enlarged model pretraining, which is often ignored in previous works, into pruning. We study the enlarge-and-prune pipeline as an integrated system to address two critical questions: whether it is worth pretraining an enlarged model even when the model is never deployed, and how to optimize the... ## Luce: Relightable Gaussians for 3D Asset Generation DevFeed: [Luce: Relightable Gaussians for 3D Asset Generation]() Original publisher: [Read original article]() Published: 2026-08-26T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Artificial Intelligence](), [Benchmark]() Tags: [3d](), [ai](), [benchmark](), [computer-vision](), [generation](), [images](), [mesh](), [models](), [multimodal](), [techniques](), [vision]() ### AI overview Luce is a multimodal 3D representation for generating relightable assets from a single image. It combines geometry with physically based materials in a voxelized Gaussian cloud, compresses them into a material-aware latent space, and generates relightable PBR Gaussians and optional textured meshes. On Toys4K, it reports a 28% FID improvement over the strongest baseline and improves alignment on an AI-generated image benchmark. ### Source excerpt High-fidelity image-to-3D generation requires a 3D representation that captures both geometry and appearance. To support relighting and integration into standard rendering pipelines, the representation should include physically based rendering (PBR) modalities such as albedo, metallic-roughness, and surface normals. We propose Luce, a 3D representation that unifies geometry and PBR materials within a voxelized multimodal Gaussian cloud, using dedicated Gaussian primitives for each modality. A variational autoencoder compresses this representation into a unified material-aware latent space. A... ## PROOF-Gen: From Optimized Data to Better Distillation DevFeed: [PROOF-Gen: From Optimized Data to Better Distillation]() Original publisher: [Read original article]() Published: 2026-08-26T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Fine-tuning](), [Prompt optimization](), [Artificial Intelligence](), [gemma4](), [Programming]() Tags: [agent](), [ai-agents](), [evaluation](), [fine-tuning](), [gemma](), [generate](), [models](), [on-device](), [post-training](), [prompt-optimization]() ### AI overview PROOF-Gen improves tool-calling model distillation by using per-scenario prompt optimization to recover successful trajectories from failed teacher attempts. The method strips corrective guidance before training, producing clean demonstrations and improving benchmark, deployed-pipeline, and on-device model performance. ### Source excerpt Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that drive shipped tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier-teacher cost each cycle, yet the mechanism is generate-and-filter (keep the teacher's passing trajectories, discard the rest) and each cycle leaves behind the same hard scenarios because failures supply no signal. On τ 2-bench, 57% of teacher trials fail, two-thirds of them near-misses (most tool calls correct, undone...