Amazon (United States)

Taxonomies organize knowledge into hierarchical structures that support effective information seeking behaviors. However, developing taxonomies in fast-evolving domains like e-commerce remains a labor-intensive process. In this paper, we present an interactive system that assists users in expanding taxonomies through automated knowledge discovery from large text corpora. On the back end, our hybrid methods combine topic modeling and large language models (LLMs) to uncover emerging concepts, generate concise summaries, and suggest mappings to taxonomy nodes. On the front end, we develop an interactive web-based interface that supports iterative, human-in-the-loop taxonomy expansion. We demonstrate the system’s versatility through two scenarios using publicly available datasets: amplifying a preliminary taxonomy in the e-commerce domain and refining a mature taxonomy in the medical domain.

The presentation of a robot's capability and identity directly influences a human collaborator's perception and implicit trust in the robot. Unlike humans, a physical robot can simultaneously present different identities and have them reside and control different parts of the robot. This paper presents a novel study that investigates how users perceive a robot where different robot control domains (head and gripper) are presented as independent robots. We conducted a mixed design study where participants experienced one of three presentations: a single robot, two agents with shared full control (co-embodiment), or two agents with split control across robot control domains (split-embodiment). Participants underwent three distinct tasks -- a mundane data entry task where the robot provides motivational support, an individual sorting task with isolated robot failures, and a collaborative arrangement task where the robot causes a failure that directly affects the human participant. Participants perceived the robot as residing in the different control domains and were able to associate robot failure with different identities. This work signals how future robots can leverage different embodiment configurations to obtain the benefit of multiple robots within a single body.

The emergence of Multimodal Large Language Models (MLLMs) has propelled the development of autonomous agents that operate on Graphical User Interfaces (GUIs) using pure visual input. A fundamental challenge is robustly grounding natural language instructions. This requires a precise spatial alignment, which accurately locates the coordinates of each element, and, more critically, a correct semantic alignment, which matches the instructions to the functionally appropriate UI element. Although Reinforcement Learning with Verifiable Rewards (RLVR) has proven to be effective at improving spatial alignment for these MLLMs, we find that inefficient exploration bottlenecks semantic alignment, which prevents models from learning difficult semantic associations. To address this exploration problem, we present Adaptive Exploration Policy Optimization (AEPO), a new policy optimization framework. AEPO employs a multi-answer generation strategy to enforce broader exploration, which is then guided by a theoretically grounded Adaptive Exploration Reward (AER) function derived from first principles of efficiency η=U/C. Our AEPO-trained models, InfiGUI-G1-3B and InfiGUI-G1-7B, establish new state-of-the-art results across multiple challenging GUI grounding benchmarks, achieving significant relative improvements of up to 9.0% against the naive RLVR baseline on benchmarks designed to test generalization and semantic understanding.

Abstract Latent Gaussian processes are widely used in latent variable models because they offer a flexible nonlinear mapping from a low-dimensional latent space to a high-dimensional space via Gaussian processes. However, these models are computationally intensive and do not scale well. In this paper, we propose a temporal categorical model that utilizes latent Gaussian processes, inducing-input approximation, and a regularization framework to model multivariate categorical processes, with and without priors on hyperparameters. We analyze the underlying properties and introduce two variational inference approaches: one based on a Monte Carlo method and the other on the delta method. We also found that latent dynamics tend to collapse into a constant zero, which hinders the reflection of dynamic information. To address this, we propose two strategies to regularize the latent dynamics for better alignment with observation dynamics: introducing a regularization term based on inducing variables and incorporating a scale prior for latent Gaussian processes. Additionally, we propose an efficient and effective stochastic variational inference technique. Finally, we demonstrate our model and inference methods using both synthetic data and real financial data.

Este estudo teve como objetivo realizar uma aula prática sobre moluscos bivalves, utilizando a espécie nativa (Crassostrea gasar) como estratégia pedagógica para potencializar a aprendizagem. A pesquisa foi realizada em 2024, com 18 estudantes da Escola Municipal de Ensino Fundamental no município de Santo Antônio do Tauá (Estado do Pará, Brasil). A metodologia incluiu o levantamento do perfil dos estudantes, planejamento e a organização de aulas teóricas e práticas com biometrias de ostras. Durante as atividades, foram abordados conceitos relacionados à introdução, morfologia, moluscos em geral, criação e biologia da ostra. Um questionário padronizado foi aplicado para medir e comparar os conhecimentos prévios e adquiridos pelos alunos. Os resultados mostraram um aumento significativo no rendimento dos estudantes em relação aos conceitos trabalhados, bem como a prática. Esses achados corroboram a hipótese de que aulas práticas são eficazes para melhorar o aprendizado dos discentes.

Social insects represent a major component of tropical forest biomass, yet the ecological fate of their necromass remains poorly understood. Stingless bees (Apidae: Meliponini), particularly species in the genus Trigona, frequently engage in aggressive, sometimes lethal conflicts while competing for nutrient-rich resources such as carrion. Here we report a previously undescribed trophic interaction in which rove beetles of the genus Macrogerodonia (Staphylinidae) prey exclusively on wounded or recently killed Trigona workers. Field observations and cafeteria choice experiments showed that beetles consistently preyed on injured Trigona bees, ignoring carrion and alternative invertebrates, including flies, grasshoppers, and other stingless bee genera. Same behavior was observed at additional sites in Peru and Costa Rica, suggesting this interaction is geographically widespread. Our observations indicate a previously undescribed, highly specialized trophic interaction linking social insect warfare to predator specialization, emphasizing the ecological significance of social insect necromass in structuring tropical forest food webs.

The rapid development of one of the biggest advances in the sphere of Generative AI has led to natural language generation still being limited to large-scale implementation due to latency, reproducibility, integration complexity, and scalability challenges. This project offers a simple and repeatable deployment pipeline which is based on the Kaggle Fitness Exercises dataset. The final corpus was then fine-tuned with GPT-2 on Hugging Face and PyTorch using extensive preprocessing, that is, data cleaning, instruction merging, normalization, tokenization, and outlier filtering. The model was evaluated on five configurations with a loss of evaluation of approximately 0.4083, 0.4094, and perplexity of 1.5043, 1.5058, and sensitivity and ablation studies indicate the most sensitive hyperparameter is the learning rate. The fine-tuning of the model was performed through a FastAPI service, and it was then containerized using Docker, which made it suitable for deployment on any cloud provider. The outcome was an even training process, low perplexity, and excellent performance reproducibility, while the tests in the production environment showed moderate in inference and throughput, which in turn led to the decision to carry out further optimization on the runtime. The proposed method offers excellent reproducibility, minimum operational cost, and easy deployment compared to the more complex orchestration and full MLOps pipelines.

For e-commerce retailers, high-quality product catalogs are vital to customer experience. Yet, despite lots of data cleaning efforts, catalog quality, especially in large catalogs, remains suboptimal. This paper shows how to use unstructured brand knowledge base data as a reference and a large language model agent to automatically enhance an e-commerce retailer's catalog quality. Unlike prior methods that usually repair and match product entries separately, our method does both concurrently. Our evaluation results show its effectiveness.

The internal combustion engine has served as the foundation of modern transportation and power generation for more than a century, with the four-stroke engine being the prevalent design in both automotive and industrial sectors. However, due to growing concerns about fuel efficiency and environmental impact, alternative engine designs have developed. One such innovation is the six-stroke engine, which expands upon the traditional four-stroke design by adding two extra strokes to the cycle. This modification seeks to improve thermal efficiency, lower emissions, and enhance overall fuel economy. In a typical six-stroke engine, the additional strokes involve injecting water or air to cool the combustion chamber and recover more power. Consequently, the working principle of the six-stroke engine features a distinct combination of increased power output and reduced operating temperatures, potentially resulting in lower emissions and improved engine longevity. This paper examines the operational principles of six-stroke engines, highlighting the mechanisms that set them apart from traditional four-stroke designs. We investigate various six-stroke engine architectures, particularly those that incorporate water vapor or air for additional cooling, and assess their performance in terms of fuel efficiency, emissions, and mechanical complexity. Furthermore, the paper compares the six-stroke engine with the four-stroke engine, emphasizing key metrics such as thermal efficiency, fuel consumption, and pollutant emissions. We discuss the potential advantages of six-stroke engines, along with the engineering challenges related to their development and mass production. Additionally, we evaluate practical considerations, including cost, scalability, and the applicability of six-stroke engines across different industries such as automotive, marine, and power generation. These results indicate that although six-stroke design may be a potentially viable technological solution for enhanced efficiency and emissions performance, there are substantial technical–economic–manufacturing barriers to implementation on a widespread basis. That said, the six-stroke engine has Ericsson on the right path at least in the search for more ecological and optimized internal combustion technologies. Finally, this paper ends with an outlook on the future of six-stroke engines and their place in the development path toward cleaner and more resource-efficient powertrains.

The European Broadcasting Union (EBU) Dynamic Media Facility (DMF) provides a layered model for software-defined production infrastructures running on commonly available IT infrastructure. The DMF seeks to enable customizable and dynamic workflows with cost-effective scalability and enhanced operational efficiency, security, and reliability to meet the challenges of the evolving media business landscape. A key element of the DMF is the Media eXchange Layer (MXL), an open source SDK for high-performance live media interchange between containerized software media functions. MXL provides an interoperable vendor-neutral platform with clear observability and uses the lowest latency asynchronous data transfer methods to avoid delays in the live signal chain. Shared memory is used for same-host functions, and remote direct memory access (RDMA) networking will be used for cross-host communication. The Linux Foundation hosts the MXL Project in collaboration with the EBU and the North American Broadcasting Association (NABA). The MXL SDK code is available on a public GitHub repository.