最新外刊素材改编关于AI的阅读理解专项练习（D篇，说明文）3篇-2026届高三英语二轮复习专项

阅读理解专项练习：脑机接口与AI 阅读下列短文，从每题所给的A、B、C、D四个选项中选出最佳选项。 For years, the complex electrical signals inside the human brain were too difficult to decode. However, artificial intelligence (AI) is now turning the dream of "mind reading" into reality. At Stanford University, a 52-year-old woman, paralyzed by a stroke for 19 years, recently saw her internal thoughts appear as text on a screen. This was made possible by a tiny array of electrodes inserted into her brain, which worked with AI to translate her imagined speech into real-time words. This breakthrough is part of a larger movement in neuroscience. In 2025, researchers in Japan introduced "mind captioning," a technique that uses AI and non-invasive brain scans to describe what a person is seeing or picturing. While these technologies are currently focused on helping patients with communication disabilities, such as those with ALS, they could eventually transform how all humans interact with each other and the world. The journey to this point has been long. Scientists have studied brain-computer interfaces (BCIs) since the 1960s, initially focusing on physical movements. Early BCIs allowed users to control prosthetic limbs or computer cursors, but decoding speech proved much harder. As neuroengineer Maitreyee Wairagkar explains, much of the early research was conducted on monkeys, who, despite being able to learn to move objects with their brain signals, simply cannot speak. Recent progress, however, has been rapid. In 2021, a study showed that a paralyzed man could "write" 18 words per minute by picturing himself drawing letters in the air. By 2024, Wairagkar’s lab trialled a new technique that translated attempted speech directly into text at 32 words per minute with 97.5% accuracy. Although this is still slower than natural human speech—which averages 150 words per minute—it marks a significant step toward everyday communication for those who have lost their voice. With companies like Neuralink seeking to bring these "brain chips" to the mass market, the era of commercialized BCIs may be just a few years away. 1. What is the function of the AI system mentioned in the first paragraph? A. To repair the damaged neurons of stroke patients. B. To stimulate the brain to produce clearer speech. C. To turn imagined speech into text on a screen. D. To predict the physical movements of paralyzed people. 2. Why was the development of speech BCIs slower than movement BCIs? A. Early animal subjects were unable to provide speech data. B. Movement signals are more complex for AI to recognize. C. Previous researchers lacked interest in communication tools. D. Non-invasive brain scans were not available in the 1960s. 3. What can we learn from the data in the last paragraph? A. Drawing letters in the air is the most efficient way to communicate. B. The 2024 technique doubled the speed of the 2021 method. C. AI can now help humans speak faster than their natural speed. D. The accuracy of brain-to-text translation has reached 100%. 4. What is the author’s attitude toward the future of BCI technology? A. Doubtful. B. Critical. C. Objective. D. Optimistic. 答案与解析 1. C 细节理解题。根据第一段 "...decoding the signals produced by her neurons as she imagined saying words, with the system translating them into text on a screen" 可知，该AI系统的功能是将想象的语言转化为屏幕上的文字。 2. A 细节理解题。根据第三段 "A lot of early work was done on non-human primates… and obviously, with monkeys you cannot study speech" 可知，早期研究多使用猴子等灵长类动物，由于它们不会说话，导致语音解码技术进展缓慢。 3. B 数据对比题。根据最后一段，2021年的研究速度是每分钟18个单词，而2024年的技术达到了每分钟32个单词。32接近18的两倍（且明显比18快得多），是该段落强调的技术进步体现。 4. D 观点态度题。通读全文，作者列举了从2021年到2025年的多项突破，并引用专家的话称其“令人兴奋（very exciting）”且即将“大规模应用（deployed at scale）”，体现了对该技术前景的乐观态度。 出题点评： 1. 文本改编：保留了原文的核心科学事实（如T16实验、ALS病人、Wairagkar的研究数据），同时简化了复杂的学术表达，使其符合高考300-350词的篇幅要求。 2. 考点分布：涵盖了细节事实、逻辑原因、数据分析和作者态度，全面考查学生的阅读素养。 3. 语言特色：使用了高考常考的科技类词汇（decode, electrode, non-invasive, prosthetic, commercialized），语篇逻辑清晰。 学科网（北京）股份有限公司 $ 英语阅读理解专项练习：AI与物理世界 AI tools are being prepared for the physical world Can a machine truly understand the world it inhabits, or is it merely "a wordsmith in the dark"? This question lies at the heart of a transformative shift in artificial intelligence: the quest for "world models". First proposed by psychologist Kenneth Craik in 1943, a world model is an internal representation of reality that allows an organism—or an AI—to test hypotheses before acting. Without such a grasp of physical laws, robots or self-driving cars would be forced into a purely reactive existence, unable to navigate the unpredictable spaces of daily life. Current efforts to build these models often begin with video generators like Google’s Project Genie. By simulating coherent interactive environments from a single prompt, Genie allows AI systems to learn through billions of hours of simulated experience, which is far more accessible than real-world data. However, these 2D simulations have clear boundaries: they often "fray at the edges" after a short duration and fail to capture hidden details, such as the smell of rotting food behind a broken freezer. To overcome these limitations, Stanford scientist Dr. Fei-Fei Li advocates for "spatial intelligence," focusing on 3D environments that are internally consistent and multimodal. Her approach creates entire digital worlds from the start, enabling multiple users to interact within the same space—a tool already proving invaluable for architects. In contrast, Yann LeCun, a pioneer in the field, argues that focusing solely on physical spaces is a distraction. His Joint-Embedding Predictive Architecture (JEPA) seeks to help AI model abstract environments, such as legal documents or HR systems, allowing machines to "think ahead" and factor in risks without needing to visualize every second of a task. A more radical view, held by OpenAI cofounder Ilya Sutskever, suggests that existing Large Language Models (LLMs) are already "learning a world model" by compressing vast amounts of internet data into underlying principles. Evidence for this surfaced in 2023, when a model trained only on text moves of the game Othello was found to have developed a detailed internal map of the game board. Similarly, researchers have identified "artificial neurons" in models like Claude that correspond to specific physical features or even complex emotions. Yet, the debate remains unsettled. Dr. Li maintains that reading about a country is not the same as experiencing it; language alone may not provide a "grounded understanding". As AI prepares to "visit" the physical world, the race to define and build its internal compass continues. 32. According to the text, why are "world models" essential for future AI systems? A. To help AI replace human decision-making in legal systems. B. To provide a platform for AI to test actions before execution. C. To reduce the need for AI to process internet-based data. D. D. To allow machines to react faster to immediate physical pain. 33. What is a major limitation of video-based world models like Project Genie? A. They require manual input for every interactive movement. B. They cannot be used to train other robotic systems. C. They struggle to maintain consistency over extended periods. D. They are unable to generate realistic pointillist paintings. 34. How does Yann LeCun’s JEPA approach differ from Dr. Fei-Fei Li’s "spatial intelligence"? A. It focuses on 2D simulations rather than 3D environments. B. It emphasizes predicting abstract outcomes over literal physical mapping. C. It relies entirely on LLMs to visualize every second of a sequence. D. It prioritizes architectural design over health-tech applications. 35. What is the significance of the Othello example mentioned in the text? A. It proves that AI can master any game without knowing the rules. B. It demonstrates that LLMs can form internal maps of reality through data. C. It shows that artificial neurons are identical to human neurons. D. It warns of the ethical risks of AI controlling board game outcomes. 答案与解析 32. 答案：B 解析：本题考查细节理解。根据文章第一段，世界模型被定义为现实的内部表征，它允许生物或人工智能在采取行动之前测试假设（test hypotheses before acting）。如果没有这种能力，机器人或自动驾驶汽车将只能进行纯反应式的存在，无法处理复杂的现实生活。选项B中的“test actions before execution”是对文中“test hypotheses before acting”的同义转述。 干扰项分析：A项（法律决策）是第三段LeCun的观点，而非世界模型通用的必要性；C项与D项在文中均无依据。 33. 答案：C 解析：本题考查细节理解与词义推断。文章第二段指出，像Project Genie这样的视频生成器作为2D模拟，具有明显的局限性：它们在短时间后往往会“边缘磨损”（fray at the edges），且无法捕捉隐藏的细节。这里的“fray at the edges”暗示了模拟在逻辑或视觉上失去了连贯性。因此，选项C“难以长时间保持连贯性”符合文意。 干扰项分析：A项（手动输入）、B项（不能训练机器人）和D项（点彩画）均不符合文中对视频模型局限性的描述。 34. 答案：B 解析：本题考查对比分析能力。根据第三段，李飞飞博士提倡“空间智能”，侧重于3D环境的构建；而Yann LeCun认为仅仅关注物理空间是一种干扰，他的JEPA架构旨在帮助AI建模抽象环境（如法律文件或人力资源系统），从而在无需可视化每一秒任务的情况下预测风险。因此，LeCun的方法更强调对抽象结果的预测，而非字面上的物理映射。 干扰项分析：A项（2D与3D）是对李飞飞与Genie模型的对比，而非与LeCun的对比；C项与LeCun“无需可视化每一秒”的描述截然相反。 35. 答案：B 解析：本题考查事实论据的意义推断。第四段引用了Othello（奥赛罗棋）的例子来支持Ilya Sutskever的观点，即现有的大语言模型（LLMs）正通过压缩互联网数据来学习世界模型。证据显示，一个仅接受过文本棋谱训练的模型，竟然在内部形成了棋盘的详细地图。这有力地证明了LLMs可以通过数据形成对现实（或特定规则世界）的内部表征或地图。 干扰项分析：A项（无需规则）与文中“训练文本步数”不符；C项（神经元相同）表述过于绝对且非该例子的核心逻辑；D项（道德风险）在该段中并未提及。 --- 命题亮点总结 本篇D篇试题高度契合2026年高考命题趋势： 1. 语料前沿化：引入了人工智能领域最新的“世界模型”与“空间智能”争议，素材具有极强的时效性和科学深度。 2. 批判性思维考查：34题要求学生对比两位科学巨头的不同科研范式（物理映射 vs 抽象建模），考查了对复杂逻辑结构的深度解析能力。 3. 高阶素养测评：35题要求学生理解实验证据（Othello例子）与抽象理论（LLM是否拥有世界模型）之间的逻辑支撑关系，而非简单的信息检索。 学科网（北京）股份有限公司 $ 英语阅读理解专项练习：AI与科学测评 阅读下列短文，从每题所给的A、B、C、D四个选项中选出最佳选项。 For years, researchers have dreamed of developing artificial intelligence (AI) that could supercharge science by posing novel questions and designing experiments. Recent breakthroughs in large language models (LLMs) have inched us closer to that future. But a critical question remains: How do you test whether an AI model can truly "do" science? To find answers, researchers turn to benchmarks: standardized sets of tasks that help assess an AI’s capacities. One of the most popular is Humanity’s Last Exam (HLE), which uses 2,500 questions drawn from the frontier of human knowledge. However, some argue that many of HLE’s questions test for "arcane"—and even trivial—knowledge, rather than an ability to do meaningful research. As Chenru Duan, founder of an AI company, puts it: “How will knowing the number of colors of phosphorus allotropes help anyone do scientific discovery?” In response, new benchmarks like FrontierScience and the Scientific Discovery Evaluation (SDE) have emerged. FrontierScience focuses on “expert-level scientific reasoning,” using questions that resemble those in science Olympiads. Meanwhile, the SDE focuses on real-world research. Rather than asking disconnected questions, it presents AI with tasks from ongoing research projects. Models are evaluated on their ability to piece together entire projects—proposing, testing, and refining hypotheses. SDE scores show that an AI’s ability to answer individual questions doesn’t always mean it can handle full projects. Another approach, LABBench2, aims to test whether AI can manage a project from the first idea to a finished paper. It evaluates "agentic AI"—systems that act independently to complete multi-step tasks like searching literature and interpreting data. So far, results are mixed: while AI is good at searching through papers, it often struggles with more involved tasks, such as cross-referencing multiple databases. Despite these challenges, researchers believe benchmarks are essential. They don't just chart who is winning; they also drive innovation by providing "North Stars" for the field. Just as a 2012 image-recognition challenge spurred the foundation of modern AI, today’s science benchmarks may lead to future breakthroughs. As science requires a wide range of skills, experts suggest leaning on a portfolio of tests rather than a single measure. After all, in order to make progress, you must be able to measure it. 1. What is the main purpose of "benchmarks" in the field of AI? A. To provide AI with the latest scientific data. B. To evaluate and compare the capacities of AI models. C. To help AI models pose novel scientific questions. D. To replace human experts in conducting experiments. 2. Why does Chenru Duan criticize the Humanity’s Last Exam (HLE)? A. Its questions are too easy for expert-level AI. B. It focuses too much on Olympiad-style reasoning. C. Some of its tasks are not helpful for actual discovery. D. It fails to cover a diverse range of scientific fields. 3. What makes the Scientific Discovery Evaluation (SDE) different from FrontierScience? A. It uses tasks based on ongoing, real-world research. B. It awards points for intermediate reasoning steps. C. It focuses specifically on the field of biology. D. It is primarily used to test "agentic AI" models. 4. What can be inferred about the future of AI in science from the text? A. A single, perfect test will soon be developed to measure AI. B. AI will focus solely on literature search instead of data analysis. C. Diverse evaluation methods are needed to catalyze AI progress. D. Modern AI has already reached the level of a human PhD scientist. 答案与解析 1. B 细节理解题。根据第二段第一句 "benchmarks: standardized sets of tasks that help assess an AI’s capacities and compare it against other models" 可知，benchmark 的作用是评估和比较 AI 的能力。 2. C 细节理解题。根据第四段，Duan 认为 HLE 测试的是“隐晦且琐碎（arcane and trivial）”的知识，并质疑这些知识对“科学发现（scientific discovery）”是否有帮助。 3. A 比较理解题。根据第六段，FrontierScience 侧重于奥赛类的科学推理，而 SDE 的指导原则是测量“真实世界研究的能力（real-world research）”，任务源于“进行中的、真实世界的研究项目（ongoing, real-world research projects）”。 4. C 推理判断题。根据最后两段，专家认为科学需要多种技能，因此需要“一系列测试组合（portfolio of tests/diverse set of evaluations）”来驱动（catalyze）不同部分的发展，并强调“为了取得进步，必须能够衡量它”。 这篇文章结构清晰，遵循了“提出问题—现状分析（不同评测工具对比）—未来展望”的逻辑。语言表达上使用了高考常见的长难句（如非限定性定语从句、动名词作主语等），词汇涵盖了科技类常考词汇（benchmark, assess, innovative, reasoning），非常适合作为高考英语模拟练习。 学科网（北京）股份有限公司 $