最新人工智能进展与影响日报

AI in Healthcare, Medical Imaging

HeadCT-ONE: Enabling Granular and Controllable Automated Evaluation of Head CT Radiology Report Generation

HeadCT-ONE：实现头部CT放射学报告生成细粒度和可控的自动评估

HeadCT-ONE is a new metric for evaluating head CT report generation that uses ontology-normalized entity and relation extraction to address language variability in radiology reports. It allows for controllable weighting of different entity types and provides a more granular evaluation of report quality.

HeadCT-ONE 是一种用于评估头部 CT 报告生成的新指标，它利用本体归一化的实体和关系提取来解决放射学报告中的语言差异问题。它允许对不同实体类型进行可控的加权，并提供更细粒度的报告质量评估。

AI’s Comment: This development is significant as it addresses the challenge of evaluating AI-generated radiology reports, which are often complex and require domain-specific knowledge. HeadCT-ONE provides a more robust and flexible evaluation method that can be used to improve the accuracy and reliability of AI-powered radiology report generation.

AI评论: 这一发展意义重大，因为它解决了评估 AI 生成的放射学报告的挑战，这些报告通常很复杂，需要领域特定知识。HeadCT-ONE 提供了一种更强大、更灵活的评估方法，可用于提高 AI 驱动的放射学报告生成的准确性和可靠性。

AI Research, Language Models, Benchmarking

LLMs Still Can’t Plan; Can LRMs? A Preliminary Evaluation of OpenAI’s o1 on PlanBench

大型语言模型仍然无法规划；大型推理模型可以吗？对 OpenAI 的 o1 在 PlanBench 上的初步评估

This research explores the planning abilities of LLMs and a new type of model called LRMs, specifically OpenAI’s o1. Despite significant advancements in LLMs, they still struggle with planning tasks. While o1 shows improved performance on PlanBench, it still falls short of achieving optimal results. This highlights the need for further research on accuracy, efficiency, and guarantees before deploying such systems.

这项研究探讨了大型语言模型 (LLM) 和一种名为大型推理模型 (LRM) 的新型模型，特别是 OpenAI 的 o1，在规划方面的能力。尽管大型语言模型取得了重大进步，但它们仍然难以完成规划任务。虽然 o1 在 PlanBench 上表现出改善，但它仍然没有达到最佳结果。这突出了在部署此类系统之前需要进一步研究准确性、效率和保证。

AI’s Comment: This news highlights the ongoing challenge of developing AI systems that can effectively plan and reason. While o1 shows promise, it emphasizes the need for continued research and development in this crucial area.

AI评论: 这则新闻突出了开发能够有效规划和推理的 AI 系统的持续挑战。虽然 o1 显示出希望，但它强调了在这一关键领域需要持续研究和开发。

AI in User Experience, Explainable AI, Smart Home Technology

A User Study on Contrastive Explanations for Multi-Effector Temporal Planning with Non-Stationary Costs

对具有非平稳成本的多效应器时间规划的对比性解释的用户研究

This research explores the use of contrastive explanations in an AI-powered smart home application that schedules appliance tasks with dynamic energy tariffs. The study shows that users provided with these explanations experience higher satisfaction, improved understanding, and find the AI-generated schedule more helpful.

这项研究探讨了在智能家居应用程序中使用对比性解释，该应用程序利用动态电价对设备任务进行调度。研究表明，提供对比性解释的用户满意度更高，理解能力更强，并且认为 AI 生成的计划更有用。

AI’s Comment: This study highlights the importance of explainable AI in user-facing applications. By providing contrastive explanations, the system can enhance user understanding and acceptance of AI-generated solutions, particularly in complex domains like smart home energy management.

AI评论: 这项研究强调了可解释 AI 在面向用户的应用程序中的重要性。通过提供对比性解释，系统可以增强用户对 AI 生成解决方案的理解和接受度，尤其是在智能家居能源管理等复杂领域。

Healthcare, AI Applications, Explainable AI

Dermatologist-like explainable AI enhances melanoma diagnosis accuracy: eye-tracking study

类皮肤科医生的可解释人工智能提高了黑色素瘤诊断准确性：眼动追踪研究

This study demonstrates the effectiveness of explainable AI (XAI) in enhancing melanoma diagnosis accuracy by dermatologists. The XAI system provides detailed explanations, boosting clinician confidence and trust in AI-driven decisions. Eye-tracking data revealed that XAI improved diagnostic performance by 2.8% compared to a standard AI system without explanations. Additionally, increased ocular fixations were observed during disagreements with AI/XAI systems and complex lesions, highlighting the cognitive load involved.

这项研究证明了可解释人工智能 (XAI) 在提高皮肤科医生黑色素瘤诊断准确性方面的有效性。XAI 系统提供详细的解释，增强了临床医生对人工智能驱动决策的信心和信任。眼动追踪数据显示，与没有解释功能的标准人工智能系统相比，XAI 将诊断性能提高了 2.8%。此外，在与 AI/XAI 系统和复杂病变发生分歧时，观察到眼球注视次数增加，突出了认知负荷。

AI’s Comment: This research demonstrates the potential of XAI in healthcare, particularly in improving medical diagnosis and reducing diagnostic errors. By providing transparent and understandable explanations, XAI fosters trust and confidence in AI-assisted decision making, ultimately leading to better patient outcomes.

AI评论: 这项研究证明了可解释人工智能在医疗保健领域的潜力，特别是在提高医疗诊断和减少误诊方面。通过提供透明和易于理解的解释，XAI 培养了人们对人工智能辅助决策的信任和信心，最终导致更好的患者结果。

Computer Vision, Deep Learning, Mobile AI

DARDA: Domain-Aware Real-Time Dynamic Neural Network Adaptation

DARDA：领域感知实时动态神经网络自适应

DARDA, a new approach for Test Time Adaptation (TTA), aims to improve the performance of Deep Neural Networks (DNNs) in the presence of data corruption or noise. Unlike existing TTA methods, DARDA proactively learns representations of different corruption types, enabling efficient adaptation to previously unseen corruptions during deployment. Experiments demonstrate significant reductions in energy consumption and memory footprint, alongside performance improvements on benchmark datasets.

DARDA 是一种新的测试时自适应 (TTA) 方法，旨在提高深度神经网络 (DNN) 在存在数据损坏或噪声情况下的性能。与现有的 TTA 方法不同，DARDA 主动学习不同损坏类型的表示，从而在部署期间能够有效地适应以前未见过的损坏。实验表明，它显著降低了能耗和内存占用，同时在基准数据集上提高了性能。

AI’s Comment: DARDA represents a promising development in the field of robust AI, particularly for resource-constrained environments. Its ability to efficiently adapt to varying data conditions without extensive retraining has the potential to enhance the deployment of AI models on mobile devices and edge computing platforms.

AI评论: DARDA 在鲁棒性 AI 领域，尤其是资源受限环境中，是一个很有潜力的发展。它能够在无需大量重新训练的情况下有效地适应不同的数据条件，这有可能增强 AI 模型在移动设备和边缘计算平台上的部署。

AI for Telecommunications, Optimization

OpenRANet: Neuralized Spectrum Access by Joint Subcarrier and Power Allocation with Optimization-based Deep Learning

OpenRANet：基于优化深度学习的联合子载波和功率分配神经化频谱接入

OpenRANet is a new deep learning model designed for optimizing resource allocation in OpenRAN networks. By combining machine learning techniques with iterative optimization algorithms, OpenRANet efficiently allocates subcarriers and power to minimize total power consumption while meeting users’ data rate requirements. It transforms a nonconvex problem into convex subproblems, solving them iteratively and integrating the solutions into a convex optimization layer. This approach enhances constraint adherence, solution accuracy, and computational efficiency. OpenRANet lays the foundation for resource-constrained AI-native wireless optimization in various scenarios, including multi-cell systems, satellite-terrestrial networks, and future OpenRAN deployments.

OpenRANet 是一种针对 OpenRAN 网络资源分配优化的深度学习模型。它结合了机器学习技术和迭代优化算法，高效地分配子载波和功率，以最小化总功耗，同时满足用户数据速率需求。它将非凸问题转化为凸子问题，迭代求解并将其整合到凸优化层。这种方法提高了约束遵守、解决方案精度和计算效率。OpenRANet 为各种场景下的资源受限 AI 原生无线优化奠定了基础，包括多小区系统、卫星地面网络和未来 OpenRAN 部署。

AI’s Comment: This research is significant because it demonstrates the potential of deep learning to optimize resource allocation in next-generation wireless networks. OpenRANet’s approach of combining optimization algorithms with machine learning could lead to more efficient and cost-effective cellular networks, especially as the complexity of these networks increases.

AI评论: 这项研究意义重大，因为它展示了深度学习在优化下一代无线网络资源分配方面的潜力。OpenRANet 结合优化算法和机器学习的方法可以带来更有效和更具成本效益的蜂窝网络，特别是在这些网络复杂性不断提高的情况下。

AI in Education, Mixed Reality, Teacher Training

MITHOS: Interactive Mixed Reality Training for Professional Socio-Emotional Interactions in Schools

MITHOS：互动式混合现实培训，支持学校专业的情绪社交互动

MITHOS is an AI-powered mixed reality training system designed to help teachers improve their conflict resolution skills. It simulates realistic classroom conflicts and guides teachers through a four-stage process that includes developing self-awareness, perspective-taking, and positive regard.

MITHOS 是一款基于 AI 的混合现实培训系统，旨在帮助教师提升冲突解决能力。它模拟真实的课堂冲突，引导教师经历四个阶段，包括培养自我意识、换位思考和积极关注。

AI’s Comment: This news item is relevant to recent AI developments in the field of education. MITHOS represents a promising approach to leveraging mixed reality and AI to enhance teacher training, potentially improving classroom management and fostering more positive student-teacher interactions.

AI评论: 这篇新闻与 AI 在教育领域的最新发展息息相关。MITHOS 代表了一种利用混合现实和 AI 来增强教师培训的有前景的方法，有可能改善课堂管理，促进更加积极的师生互动。

Recommendation Systems, Natural Language Processing

TRACE: Transformer-based User Representations from Attributed Clickstream Event sequences

TRACE：基于Transformer的属性点击流事件序列用户表示

TRACE is a new transformer-based method for generating user embeddings from multi-session clickstream data. Unlike prior works that focus on single-session sequences, TRACE leverages site-wide page view sequences spanning multiple sessions to model long-term user engagement. This approach enables capturing comprehensive user preferences and intents, leading to improved personalization in recommendation systems.

TRACE 是一种新的基于 Transformer 的方法，用于从多会话点击流数据中生成用户嵌入。与之前专注于单会话序列的工作不同，TRACE 利用跨越多个会话的站点级页面浏览序列来建模长期的用户参与度。这种方法能够捕捉到全面的用户偏好和意图，从而提高推荐系统的个性化水平。

AI’s Comment: This research is significant because it addresses the limitations of traditional recommendation systems that often fail to capture long-term user preferences. By leveraging multi-session data, TRACE can provide more nuanced and personalized recommendations, potentially leading to increased user engagement and conversion rates.

AI评论: 这项研究意义重大，因为它解决了传统推荐系统在捕捉长期用户偏好方面存在的局限性。通过利用多会话数据，TRACE 可以提供更细致和个性化的推荐，从而有可能提高用户参与度和转化率。

Medical AI, Radiology

The Era of Foundation Models in Medical Imaging is Approaching: A Scoping Review of the Clinical Value of Large-Scale Generative AI Applications in Radiology

医学影像基础模型时代即将到来：大型生成式 AI 应用于放射学的临床价值范围综述

This scoping review explores the potential of large-scale generative AI in revolutionizing medical imaging. While AI models show promise in improving report generation efficiency and patient understanding, they haven’t yet surpassed radiologists in diagnosis. The study suggests the future of foundation models in medical imaging is promising, with potential to significantly impact clinical practice.

这篇范围综述探讨了大型生成式 AI 在革新医学影像方面的潜力。虽然 AI 模型在提高报告生成效率和患者理解方面展现出希望，但它们尚未在诊断方面超越放射科医生。该研究表明，医学影像基础模型的未来前景光明，有可能对临床实践产生重大影响。

AI’s Comment: This news highlights the growing impact of AI in healthcare, particularly in radiology, where it can automate tasks and potentially improve diagnosis accuracy. However, it acknowledges the limitations of current AI models and the need for further development.

AI评论: 这则新闻突出了 AI 在医疗保健，特别是在放射学领域的影响力不断增强，它可以自动化任务并有可能提高诊断准确率。然而，它也承认了当前 AI 模型的局限性，以及进一步发展的必要性。

Natural Language Processing, Prompt Engineering

Can we only use guidelines instead of shots in prompt?

我们是否可以只在提示中使用指南，而不是示例？

This research explores the potential of using guidelines instead of examples (shots) in prompt engineering for AI models. It introduces a framework (FGT) to automatically learn task-specific guidelines from datasets, aiming to overcome the limitations of shot-based methods.

这项研究探讨了在 AI 模型的提示工程中使用指南而不是示例（示例）的潜力。它介绍了一个框架 (FGT) 来从数据集中自动学习特定于任务的指南，旨在克服基于示例方法的局限性。

AI’s Comment: This research could significantly impact prompt engineering by potentially simplifying the process and reducing the dependence on manually selected examples. The FGT framework, if successful, could be a valuable tool for optimizing AI model performance in various tasks.

AI评论: 这项研究可能会对提示工程产生重大影响，因为它有可能简化流程并减少对手动选择的示例的依赖。 FGT 框架如果成功，将成为优化 AI 模型在各种任务中表现的宝贵工具。