arXiv:2512.17131v3 Announce Type: replace-cross Abstract: We propose Generalized Primal Averaging (GPA), an extension of Nesterov’s method that unifies and generalizes recent averaging-based optimizers like single-worker DiLoCo and Schedule-Free, within a non-distributed setting. While DiLoCo relies on a memory-intensive two-loop structure to periodically aggregate pseudo-gradients using Nesterov momentum, GPA eliminates this complexity by decoupling Nesterov’s interpolation constants to enable smooth iterate averaging at every step. Structurally, GPA resembles Schedule-Free but replaces uniform averaging with exponential moving averaging. Empirically, GPA […]

Ver mais

Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and centralized learning. Distributing rollout execution offers opportunities to leverage more cost-efficient inference resources, but introduces challenges in wide-area coordination and policy dissemination. We present ECHO-2, a distributed RL framework for post-training with remote inference workers and non-negligible dissemination latency. ECHO-2 combines centralized learning with distributed rollouts and treats bounded policy staleness as a user-controlled parameter, enabling […]

Ver mais

Background: Extracting the stages that structure Machine Learning (ML) pipelines from source code is key for gaining a deeper understanding of data science practices. However, the diversity caused by the constant evolution of the ML ecosystem (e.g., algorithms, libraries, datasets) makes this task challenging. Existing approaches either depend on non-scalable, manual labeling, or on ML classifiers that do not properly support the diversity of the domain. These limitations highlight the need for more flexible and reliable solutions. Objective: […]

Ver mais

Engaging the AI community through building, research, and shared learning Machine learning Staff writer February 02, 02:51 PM February 02, 03:34 PM Advancing AI requires more than breakthrough models. It depends on communities of builders and researchers who experiment, test assumptions, and share what they learn. That belief is guiding how Amazon engages developers and academics around Amazon Nova, Amazons portfolio of AI offerings including the Nova models, Nova Forge and Nova Act. Today, two Nova initiatives launch […]

Ver mais

MP4 video was built for broadcast, not for the dynamic, data-driven web. As video becomes the primary interface online, static files can’t personalize, adapt, or respond in real time. Blings proposes a runtime video model (MP5) where video is assembled on the fly like software—interactive, composable, privacy-first, and personalized for every viewer.

Ver mais

This study develops an adaptive workflow allocation mechanism for anti-money laundering (AML) operations, aiming to improve the accuracy and efficiency of suspicious-transaction review. A multi-agent simulation platform was constructed to model transaction flows, alert generation, and analyst decision behaviors. The system integrates model-confidence estimation, analyst-fatigue prediction, and real-time workload signals to dynamically route alerts. Experiments were conducted using 27.3 million historical transactions and 186,000 alerts from a large commercial financial dataset. Compared with fixed allocation rules, the adaptive […]

Ver mais

Large Language Models (LLMs) are increasingly applied to critical domains such as medical report generation, where accuracy and trust are essential. Evaluating the quality of generated text is non-trivial: surface word matches may miss key semantic errors, while semantic metrics may overlook domain-specific mistakes. This review goes through five categories of evaluation metrics, using a consistent medical example to illustrate their differences: Reference report: “The chest X-ray shows evidence of pneumonia. No pleural effusion is present.” Generated report: […]

Ver mais