Apple· 1 OctResearch paper19h ago

Apple researchers found that, under equal time and using the same frontier LLM, a minimal-harness coding agent baseline matches or outperforms open-source state-of-the-art harnesses on current MLE benchmarks, suggesting the backbone model is the primary driver of performance.

Apple· 30 Sep1 day ago

Authors Youngmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang, Manjot Bilkhu Continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons. Evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. Yet existing benchmarks and training frameworks schedule only the benchmark’s own events, leaving each benchmark and agent pair to build a custom scheduling loop. We present SCLATE, an execution substrate where benchmarks and unmodified agents each add their events to one open event scheduler through an adapter. A hybrid simulated clock runs these events on a shared timeline, flowing in real time while the agent works and skipping idle gaps, which compresses a month-long scenario into hours. SCLATE also serves as a rollout engine that runs any agent’s harness and memory unmodified, recording the tokens and log probabilities of every model call through an in-container proxy. We port seven benchmarks to SCLATE and compare ten unmodified harness and memory configurations head to head on ten models. …

Apple· 30 SepResearch paper1 day ago

Apple researchers systematically studied conditioning methods for LLMs, finding that efficient steering often degrades fluency and that activation steering is less effective on instruction-tuned models than on base models. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics correlate well with costly LLM-as-judge scores.

Apple· 29 SepResearch paper2 days ago

Apple researchers propose a round-trip protocol to measure how much tree-structured compositional content survives when language models serialize expressions into natural language. Evaluating 16 models, they find the channel is lossy and asymmetric, with at least 73.6% of failures originating at generation, and that fine-tuning on ∼3600 examples lifts open-weight models above an untrained frontier model.

Apple· 28 SepResearch paper3 days ago

Apple researchers present a paper studying federated optimization for stochastic variational inequalities. They provide improved convergence rates for the Local Extra SGD algorithm and propose a new algorithm, LIPPAX, which reduces client drift and achieves better guarantees in bounded Hessian, bounded operator, and low-variance settings. The work also extends results to federated composite variational inequalities.

Apple· 24 SepResearch paper7 days ago

Apple researchers present a method for semi-supervised federated learning in automatic speech recognition that uses per-client online pseudo-label generation stabilized by server-side updates on labeled data. The approach improves over prior methods on 9 of 11 evaluation pairs, narrowing the gap to fully-supervised federated learning.

Apple· 24 SepResearch paper7 days ago

Apple researchers present a method for compressing streaming neural audio encoders used in on-device dictation through latent-space distillation. The technique trains a student encoder to regress the teacher's pre-quantizer latent representation, achieving 2.8× compression while maintaining within 1.9% relative word error rate of the teacher on most model pairs without fine-tuning.

Apple· 23 SepResearch paper8 days ago

Apple researchers introduce probe guidance, a method to guide flow matching models that uses frozen internal states from existing diffusion models to construct guidance signals. Applied to a 1.7B diffusion language model, probe guidance achieves state-of-the-art performance on unconditional generation and improves multiple choice question answering benchmarks while eliminating the need for additional forward passes at inference time.

Apple· 18 SepResearch paper13 days ago

Apple researchers introduced Dynamically Scaled Activation Steering (DSAS), a method-agnostic framework that adaptively modulates the strength of steering interventions across layers and inputs to guide generative model behavior. The approach decouples when to steer from how to steer, intervening strongly only when undesired behavior is detected, and demonstrates improved trade-offs between toxicity mitigation and utility preservation compared to existing steering methods alone.

Apple· 17 SepResearch paper14 days ago

Apple introduces REVERSAL-BENCH, a benchmark that controls environmental reversibility via a continuous parameter and provides a reset oracle to test state recoverability across eight manipulation settings in five physics engines. Evaluating various policy architectures reveals a sharp reversibility cliff: reset-free agents are consistently absorbed into irrecoverable states as reversibility increases, whereas episodic agents maintain steady learning. The benchmark suite, a large multi-simulator dataset labeled with recoverability, and a reset oracle are released.

Apple· 16 SepResearch paper15 days ago

Apple researchers investigate unintended effects of value induction in conversational LLMs, finding that inducing values leads to expression of related or contrastive values, increases safety for positive values, and universally increases anthropomorphic language making models more validating and sycophantic.

Apple· 16 SepResearch paper15 days ago

Apple researchers propose Trajectory-Shaped Discrete Flow Matching (TS-DFM), a method that improves few-step discrete flow matching by using a lightweight energy compass to guide token selection during training. On a 170M-parameter language model, the 8-step student achieves 32% lower perplexity than the 1,024-step teacher while being 128× faster.

Apple· 16 SepResearch paper15 days ago

Apple researchers propose DACA-GRPO, a plug-and-play enhancement for GRPO-style trainers that introduces Denoising Progress Scores and Stratified Masking Likelihood to improve credit assignment in reinforcement learning for diffusion language models. The method achieves consistent gains across seven benchmarks, including up to 5.6pp on math reasoning, 7.4pp on code generation, 36.3pp on constraint satisfaction, and 5.9pp on JSON schema adherence.

Apple· 16 SepResearch paper15 days ago

Apple researchers present Glyph, a production system using cooperating LLM agents to automatically generate column descriptions and assign data classification labels to enterprise data catalog tables. The system combines multiple strategies including retrieval-augmented generation from source code, description-based tagging, regex matching, and a fine-tuned metadata encoder, fusing results with Reciprocal Rank Fusion to improve classification accuracy.

Apple· 16 SepResearch paper15 days ago

Apple researchers introduce shared selective persistent memory, a memory architecture for agentic LLM systems that identifies and retains four categories of reusable context—task specifications, data schemas, tool configurations, and output constraints—while discarding session-specific reasoning traces. The architecture is implemented in a deployed collaborative workspace platform where LLM agents produce, edit, and maintain git-versioned artifacts from heterogeneous data sources. Across three enterprise deployment scenarios, shared selective persistent memory achieves 96% task completion (vs. 79% without memory and 71% with full history), and a complementary zero-token data refresh mechanism eliminates LLM re-invocation for recurring data updates (14× task time reduction).

Apple· 11 SepResearch paper21 days ago

Apple researchers present SimpleDesign, a multi-modal protein design model that jointly generates protein sequences and structures using a single-stage end-to-end training objective, trained on over 2 million sequence-structure pairs. The model achieves competitive performance on co-design and unconditional generation benchmarks.

Apple· 11 SepResearch paper21 days ago

Apple researchers introduced CapQuiz, a reference-free benchmark for evaluating video caption quality in Visual Large Language Models, along with the CapF1 composite metric. The benchmark uses human-verified multiple-choice questions across 10 question types and 24 video domains, and experiments show it correlates better with human judgments than existing metrics.

Apple· 11 SepResearch paper21 days ago

Apple researchers introduce DiscoSign, a computational framework for discourse-aware text to sign language gloss translation that addresses spatial coreference resolution, question-answer clauses, and concept-gloss consistency using a modular large language model approach. The work establishes the first systematic framework for discourse-level text to sign language gloss translation with novel evaluation metrics designed to assess discourse coherence.

Apple· 2 SepResearch paper2 Sep

Apple researchers introduce REFACTOR-VLA, a system that learns reusable motor skills using a wake/sleep architecture with a Behavioral-Equivalence Kernel and typed lambda terms. Evaluated on the LIBERO benchmark, the study finds that increasing world model size from 188M to 430M parameters worsened performance on all four suites, while adding an auxiliary InfoNCE contrastive loss during world-model warmup improved skill clustering quality.

Apple· 28 AugResearch paper28 Aug

Apple researchers present Agent Seer, a pipeline that synthesizes realistic evaluation scenarios for tool-using AI agents from a single Model Context Protocol (MCP) specification without examples, live tool access, or domain-specific tuning. The pipeline generates graded scenarios and multi-turn dialogues, achieving strong tool-calling correctness and conversational coherence across seven MCP specifications. The analysis finds that parameter schema complexity is the strongest correlate of quality variation, with argument value accuracy as the dominant failure mode.

Apple· 28 AugResearch paper28 Aug

Apple researchers introduce a technique to study LLMs as information processing rules, using the deviation from Bayesian updates to measure internal inconsistencies in how LLMs update probabilistic beliefs from evidence. Their experiments show that non-Bayesian heuristic updates often outperform exact Bayesian updates in downstream tasks, suggesting LLMs' probabilistic world models are misspecified. The measure can provide diagnostics for issues in LLM-powered inferential systems.

Apple· 27 AugResearch paper27 Aug

Apple researchers introduce a rubric-based reward framework for open-domain question answering that generates query-specific rubrics grounded in retrieved evidence and decomposed into multiple quality dimensions. The approach improves over the instruction-tuned baseline by 6.5% and over flat rubric variants by 4% across three evaluation axes (composition, grounding, and instruction-following). Conditioning rubrics on retrieved evidence improves factual support, while decomposing rubrics into quality-specific dimensions further improves coherence, organization, and adherence to query requirements.

Apple· 26 AugResearch paper26 Aug

Apple researchers propose Luce, a 3D representation that unifies geometry and physically based rendering (PBR) materials within a voxelized multimodal Gaussian cloud. A rectified-flow transformer generates this representation from a single image, and the latent decodes into relightable PBR Gaussians and an optional textured mesh. On the Toys4K benchmark, Luce achieves state-of-the-art single-image-to-3D generation, improving FID by 28% over the strongest baseline.