Neural intel Pod

This document introduces CodePDE, a new framework for using large language models (LLMs) to generate code that solves partial differential equations (PDEs). The authors frame PDE solving as a code generation problem and demonstrate that, with techniques like debugging and refinement, LLMs can create solvers that are competitive with, and sometimes surpass, human-written solvers on various PDE families like Burgers, Advection, and Darcy flow. They highlight the ability of LLMs to reason, debug, and improve code through feedback, suggesting a promising future for LLMs in scientific computing, despite challenges with certain PDE types like Reaction-Diffusion.

This research investigates online learning for feedforward neural networks utilizing the sign activation function. The paper identifies a margin condition in the first hidden layer as crucial for learnability, demonstrating that the optimal error bound is closely tied to the totally-separable packing number of the input space, showing an exponential dependence on dimension in some cases. To address this dimensionality issue, the authors examine two scenarios: a multi-index model where the function depends on a lower-dimensional projection, achieving better bounds, and a setting with a large margin throughout all layers, yielding bounds dependent on network depth and the number of output labels. The study also provides a method for adaptive learning when these margin parameters are unknown and extends its analysis to the agnostic learning setting.

This document introduces CityAVOS, a new benchmark dataset designed for Aerial Visual Object Search (AVOS)tasks using Unmanned Aerial Vehicles (UAVs) in realistic urban environments. The text describes the unique challenges of urban AVOS, such as complex semantics and difficult identification of similar objects, which differentiate it from previous navigation and object search tasks. It also presents PRPSearcher, a novel Multi-modal Large Language Model (MLLM)-powered agentic method that employs spatial perception, target reasoning, and action planningusing semantic, cognitive, and uncertainty maps to improve UAV search capabilities. Experimental results demonstrate that PRPSearcher outperforms existing baseline methods, though human performance remains the ultimate benchmark, highlighting areas for future research in autonomous urban search.

This document introduces BAT (Benchmark for Auto-bidding Task), a new resource for researching autobidding algorithms in online advertising auctions. It provides a large-scale dataset from the Avito platform, including information on two common auction types, VCG and FP, along with campaign details, auction outcomes, and traffic patterns. The paper also presents several baseline autobidding algorithms tested on this dataset and evaluates their performance on key challenges like budget pacing uniformity and Cost Per Click (CPC) constraint optimization. The resource aims to offer a standardized and user-friendly framework for researchers to develop and assess new automated bidding strategies.

This collection of texts from Amazon Science highlights the company's extensive research and development efforts across various scientific and technical domains, including machine learning, artificial intelligence, and robotics. A significant portion focuses on improving the training of large language models, particularly through a novel method called SeRAwhich aims to reduce spurious correlations in data used for reinforcement learning with human feedback. Several job descriptions are also included, showcasing the types of applied science roles Amazon is actively recruiting for in areas like personalized recommendations, video content analysis, and supply chain optimization. The content collectively demonstrates Amazon's commitment to advancing technology and applying scientific principles to real-world problems.

This document introduces T2I-R1, a novel text-to-image generation model that uses Reinforcement Learning (RL) and a bi-level Chain-of-Thought (CoT) process to improve image generation. Unlike traditional methods, T2I-R1 leverages semantic-level CoT for high-level planning based on the text prompt and token-level CoT for detailed, patch-by-patch image generation. A key component is BiCoT-GRPO, an RL method that optimizes both levels of CoT simultaneously, utilizing an ensemble of vision experts to provide diverse and robust generation rewards. By applying this approach to a Unified Large Multimodal Model (ULM), T2I-R1 achieves superior performance on established benchmarks, outperforming baselines and state-of-the-art models.

This document proposes pretraining strategies for estimating heterogeneous treatment effects (HTE), particularly the conditional average treatment effect (CATE), which varies based on individual characteristics. The core idea is to leverage the shared support between factors that are prognostic of the outcome (influence baseline risk) and those that are predictive of treatment effect heterogeneity (interact with treatment). The authors demonstrate this approach primarily within the R-learner framework, showing that using information from the prediction of the mean outcome can improve the accuracy, power, and support recovery for estimating the CATE, especially in challenging, high-dimensional settings where data may be limited. They also discuss extending this concept to the DR-learner and non-linear models like generalized random forests, suggesting it could be beneficial when such overlap between prognostic and predictive factors exists.

This academic paper introduces a novel mathematical formula that precisely predicts when a large language model (LLM) might suddenly shift from producing beneficial output to generating incorrect or harmful content, referred to as a "Jekyll-and-Hyde" tipping point. The authors attribute this change to the AI's attention mechanism, specifically how thinly its attention spreads across a growing response. They argue that this tipping point is predetermined by the AI's initial training and the user's prompt, and can be influenced by altering these factors. Notably, the study concludes that politeness in user prompts has no significant impact on whether or when this behavioral shift occurs. The research provides a foundation for potentially predicting and mitigating such undesirable AI behavior.

This document introduces Yo’Chameleon, a novel method for personalizing Large Multimodal Models (LMMs) like Chameleon. Recognizing that current LMMs lack user-specific knowledge, the paper proposes using only 3-5 images of a novel concept to train the model for both personalized language and vision generation. Yo’Chameleon employs soft-prompt tuning to integrate subject-specific information and utilizes a dual soft prompt architecture with a self-prompting mechanism to handle both understanding and generation tasks effectively. Furthermore, a unique soft-positive training strategy leverages similar negative images to enhance generation quality, demonstrating a significant step towards making LMMs more personalized for real-world applications while avoiding catastrophic forgetting of general abilities.

Three texts published in April 2025 discuss the burgeoning integration and implications of Artificial Intelligence (AI) in different sectors. One source focuses on the financial system, outlining the potential benefits and significant risks AI presents for financial stability, including operational, market, and cyber concerns, emphasizing the need for monitoring and regulatory adaptation. Another article examines the use and ethical considerations of AI, alongside collaboration and audience participation, within local Ibero-American newsrooms, highlighting limited AI knowledge and widespread ethical worries despite openness to human-centered collaborative practices. The final source, an news roundup on AI in the construction industry, illustrates the practical applications of AI for efficiency and safety, noting continued investment alongside instances of project cancellation and predictions for a revolutionary year for AI in the sector.