欢迎收看"进口AI"，关于AI研究的通讯。 进口AI运行于arXiv，卡布奇诺斯，以及读者的回馈。 请各位同意。
LLMs在低纬度任务中非常糟糕：
直到AI系统可以玩游戏或为你进行在线拍卖
研究人员建立了Realtime-Bench,“一套精心设计的动态任务，其中延迟行动直接影响成功”。 测试有助于评价现代，通用AI系统在实时需要做决定的地方如何很好地执行任务。
基准测试由50个交互式游戏和网络任务组成，分为4个类别(旅行、电子商务、服务和安全)和20个游戏任务(三角、谜题、战略和模拟)，分为30个网络任务。
谁做了研究：基准测试由人们在德克萨斯大学奥斯汀分校，俄亥俄州立大学，伊利诺伊大学厄巴纳-尚帕因分校，NVIDIA分校，西北大学和"稳定工作"分校建造。
其衡量标准：测试旨在探究三种特性：时间状态意识(“RT-Perceive”)，模型可以利用观测来跟踪世界的变化；适应性考虑(“RT-Deliberate”)，模型根据可用时间改变其推理长度和深度；以及时间意识行动选择(“RT-Act”)，模型需要适当时间对情况采取行动。
“例如，购买机票的agent必须跟踪销售状况并评估所观察到的可用性是否仍然有效(RT-Perceive)，限制购买窗口关闭时的进一步考虑(RT-Deliberate)，在等待售出打开和购买时选择(RT-Act)”。
它们有三个不同的条件：墙钟 (Wall Clock)，其设计目的是在现实世界的后期测试出性能。 被暂停的时钟，它允许你任意地暂停一些事情，以便你得到一种基线能力感，而Token时钟，它“以每代指示符的固定速度提高环境。 一项行动是在实施时所达到的模拟时间内实施的，因此，不管实际推论的延迟程度如何，消耗更多的象征物都会拖延执行”。
模特儿做得怎么样？作者从OpenAI(GPT-6-Astra,GPT-5.6 Sol,GPT-Realtime-2,OpenAI决定),Anthropic(Claude有线5.1,Claude Opus 5),Google(Gemini 3.8 Flash,Gemini 3.1 Flash Live,Gemini 3.8 Live),Jev 1.13以及从Deepseek(V4.1)和Moonshot AI(Z10Z-K3)测试出型号。
GPT-6 Astra(高)在 " 墙上钟 " 评价设置上获得最高的通过率，为36%。
相比之下，与一个被暂停的时钟相比，具有高推理的模型会更接近70%甚至80%的成功率。
为什么这很重要：这个基准测试是AI agents会打破的有效代名词——这里的低分点指向了AI agents尚不能预期表现良好的任务。
请检查access-date=中的日期值 (帮助) 领导板：Realtime-Bench (官方网站).
阅读论文：Realtime-Bench：边疆模型能否解决实时互动问题？(arXiv).
***
太空战争的未来，国防情报局的礼遇：
...反卫星技术太空核弹还有更多。..
在AI工作会很紧张，所以有时为了放松自我，我喜欢阅读其他话题，以了解别处的恐怖。 如果我感到悲伤，我会尝试听比我更悲伤的人的音乐，以为自己高兴(例如，Radiohead, Neil Young
因此，考虑到这一点，我最近读了国防情报局的“对空间安全的挑战”。
太空中发生了什么：与地球上的冲突一样，空间是由美国与中国、俄罗斯、伊朗和朝鲜之间的冲突所定义的，中国的技术最先进，而俄罗斯似乎是最强悍的。
“中国和俄罗斯正在发展从动能破坏到复杂的干扰和网络操作，“过去十年来，中国每年的空间发布增加了两倍多，2016年为22个发布，2025年为93个发布。 北京评委空间是美国及其盟国行动的核心。 北京也认为空间是一个进攻性的军事领域，它塑造了解放军军事规划和战役的组成部分。”
太空战争：阅读这篇100+页面报告，你得到的一般感觉是，大量的丑陋，可怕的事情发生在太空中，或者被地面上的东西对太空中的东西做。 一些例子引述：
- “到2020年代后期，中国可能开发空基定向能系统和能够威胁高轨道卫星的高能地定向能系统”。
- “中国正在研制天基技术，包括能使空间系统处于危险的在轨反空间武器”。
- " 俄罗斯的 " Peresvet " 是一种地基激光系统，旨在使敌方卫星的光学系统眩晕以掩盖俄罗斯战略导弹系统的移动 "。
- " 在过去几年中，俄罗斯有几颗发布卫星，在轨道上进行试验活动，以获取能够攻击其他国家的卫星并造成暂时或永久损害的技术 "。
中国有几架太空飞机，类似于美国的超机密X-37B飞行器，它们似乎越来越先进了。 2023年，中国前述车辆之一"发布并用物体相接".
卫星对卫星的战斗：在整个报告中，我们显然处在一个敌对国家拥有卫星的世界中，这些卫星试图与其他人的其他卫星相撞。
- “2019年，俄罗斯发布卫星Cosmos 2542，这是发布的第二颗卫星Cosmos 2543。 Cosmos 2543随后开始追随美国国家安全卫星，并在相距足够相近的距离上多次接近，从而造成潜在的危险操作条件",DIA写道。 “俄罗斯正在发展动能轨道系统，使卫星永久失效或“致命”卫星。 莫斯科的 " Burevestnik " 是一种用于在低地轨道上杀死航天器的同轨航空-发布反卫星武器 "。
- “2022年2月，俄罗斯发布一颗名为宇宙的卫星，莫斯科当时声称正在测试空间辐射对电子系统的影响。 这颗卫星是发布，进入了任何其他航天器没有使用的不寻常和孤立的轨道”。
空间核武器：俄罗斯正在空间试验核武器。 DIA写道，“在空间发生核爆炸将产生电磁脉冲，这立即会影响到射程内的卫星并会影响到民用基础设施”。 " 空间核爆炸可针对大多数轨道的撞击卫星——用于商业和侦察卫星的低地轨道、用于关键通信的地球同步轨道和用于导航和定时卫星的中地轨道 "。
为什么这很重要——AI权力斗争下一阶段的基础设施：随着人们将更多资源花在建设以空间为基础的数据中心(以及最终也许以月球为基础的中心)上，空间将变得更具战略性和商业价值并更具争议性。 这样的研究让我们感觉到已经存在太空中的硬件并且将会在未来到达也许在遥远的将来，大赦国际也可能设法控制这些空间资产。
更多信息：DIA 发布对空间安全的挑战(国防情报局，新闻稿)。
阅读报告：对空间安全的挑战(DIA,PDF)。
***
Metis：一个试图将符号和神经推理融为一体的启动者：
...向着类似Vulcan的机器思维。..
Metis是一家初创公司，希望围绕开发推理更好的机器建立小说AI架构，他发表了一份研究论文，阐述了其研究方法。 高級簡介 – Metis想为现代开放域问题做DeepMind通过在AlphaGo和AlphaGo Zero上的工作为封闭域问题所做的。
他们想做什么：这里的关键想法是在现代AI架构中添加一个推理核心，它远比我们今天拥有的更可靠。 他们写道：“在现代AI架构中，使AlphaGo成为成功推理系统的其他许多周边机制要么是隐含的，要么是不存在的。” “第一个缺失的元素是一个明确而持久的感知状态。.....第二个缺失的元素是一个标准，允许系统选择在实现目标方面取得进展的推理动作。”
这更像是一份立场文件，主要思想是在现实世界数据收集的基础上，训练推理政策和价值函数，再用这个来指导系统进行能减少长期不确定性的推理，并将它们坐落在由声称或关于世界的假设所构成的基本数据核心上。 他们写道，" 对索赔证据地位的任何改变必须经证据许可或通过明确的计算加以核实并得到独立判决系统的批准 "。 “模型生成的内容不能仅仅通过合理性进入公认的认识状态；它必须通过合理评估，从而降低幻觉。 证据来源可以直接附在其支持的命题上，提供明确的出处链。 推理程序显然可以审计。”
之所以如此重要，即建立可靠且也极非人道的基底：像这里描述的建筑理念令人着迷，这样，当代的LLMs比这里描述的系统更像人类。
Metis写道，“尽管AI的历史可以被理解为象征性方法和神经方法之间的竞争，“分类系统使推理过程的某些方面变得明确，但取决于主要通过手工提供的表述、操作者、弹性和知识。 当代的LLMs学习了非同寻常的语义操作和大量隐含的知识，但将组织、评价和推理的积累留作隐含。”
本文描述的人工智能系统将“思考”方式与人类的想法大相径庭
更多信息：理性时代(Metis,PDF)。
***
Google召唤Genie 4为更好的世界建模：
...Google最有趣的项目之一仍在继续！
Google DeepMind发布了Genie 4的细节，这是其Genie世界模式的新迭代。 该公司正在推出Genie 4的进入权，目前作为一些值得信任的测试者的研究预览，最终将来到Google AI Ultra的订阅者。 Genie系列的模型很迷人，展示了一个世界，你可以拍下任何东西的照片，把它变成一个可互动的环境，或者一个你可以把2D画转换到舞台上进行视频游戏，等等。
主要特征：Google召唤出四个特性，使Genie 4与前作不同。 其中包括：
- 实时修改："你可以使用自然语言来修改实时的模拟——比如改变天气，移动相机，或者引入一个新人或新物体——而Gemini会根据背景调整你的提示来配合周围的场景".
- 物理和人类动作模拟：“Genie 4模拟现实世界的物理情景，如空气动力学、相撞
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.
LLMs are very bad at low-latency tasks:
…Realtime-Bench says it’ll be a while till AI systems can play games or do online auctions for you…
Researchers have built Realtime-Bench, “a suite of carefully curated dynamic tasks where delayed actions directly affect success”. The test helps evaluate how well modern, general-purpose AI systems can perform tasks where decisions need to be made in real-time.
The benchmark consists of 50 interactive game and web tasks, split across 30 web tasks in four categories (travel, e-commerce, service, and security), and 20 game tasks (arcade, puzzle, strategy, and simulation).
Who did the research: The benchmark was built by people at the University of Texas at Austin, Ohio State University, University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University, and Steadyworks.
What it measures: The test is designed to probe for three qualities: temporal state awareness (”RT-Perceive”), where models can use observations to track changes in the world; adaptive deliberation (”RT-Deliberate”), where models alter their reasoning length and depth according to available time; and time-aware action selection (”RT-Act”) where models need to appropriately time their actions for the situation.
“For example, an agent purchasing a ticket must track the sale status and assess whether observed availability is still current (RT-Perceive), limit further deliberation when the purchase window is closing (RT-Deliberate), and choose between waiting for sales to open and purchasing while tickets remain available (RT-Act).”
They have three different conditions: Wall clock, which is designed to test out performance under real-world latency. Paused clock, which lets you arbitrarily pause things so you can get a sense of baseline capability, and Token clock, which “advances the environment at a fixed rate per generation token. An action is applied at the simulated time reached when it is committed, so consuming more tokens delays execution regardless of actual inference latency”.
How well do models do? The authors tested out models from OpenAI (GPT-6-Astra, GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions), Anthropic (Claude Fable 5.1, Claude Opus 5), Google (Gemini 3.8 Flash, Gemini 3.1 Flash Live, Gemini 3.8 Live), Jev 1.13, as well as Chinese models from Deepseek (V4.1) and Moonshot AI (Kimi-K3).
GPT-6 Astra (high) got the highest pass rate on the ‘wall clock’ evaluation setting, with 36%.
By comparison, with a paused clock, models with high reasoning get closer to 70% or even 80% success.
Why this matters: This benchmark is an effective proxy for where AI agents will break - the low scores here point to tasks where AI agents cannot yet be expected to perform well.
Check out the leaderboard: Realtime-Bench (official website).
Read the paper: Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems? (arXiv).
***
The future of spacewar, courtesy of the Defense Intelligence Agency:
…Anti-satellite tech, space nukes, and more…
Working in AI can be stressful so sometimes to relax myself I like to read about other topics to get a sense for what horrors lie elsewhere. I take the same approach with depression, where if I feel sad I will try and listen to music by people sadder than me in an attempt to cheer myself up (e.g., Radiohead, Neil Young, and in very dark moments Bill Callahan / Smog, etc).
So with that in mind I recently read “Challenges to Security in Space”, by the Defense Intelligence Agency.
What’s going on in space: As with conflict on Earth, space is defined by conflict between the US versus China, Russia, Iran, and North Korea, with China the most technologically sophisticated, and Russia the most seemingly gung ho.
“China and Russia are developing a formidable spectrum of offensive capabilities to degrade or destroy space services of other countries, from kinetic destruction to sophisticated jamming and cyber operations,” the DIA writes. “China has more than tripled its space launches per year in the past decade, with 22 launches in 2016 and 93 launches in 2025… Beijing judges space is central to the United States and its allies’ operations. Beijing also considers space an offensive military domain, which has shaped integral components of PLA military planning and campaigns.”
Spacewar: The general vibe you get from reading this 100+ page report is there is a huge amount of ugly, frightening stuff happening in space, or being done to things in space by things on the ground. Some example quotes:
- “By the late 2020s, China may field space-based directed energy systems and higher-powered ground-based directed energy systems capable of threatening satellites in higher orbits”
- “China is developing space-based technology, including on-orbit counterspace weapons, capable of holding space systems at risk”
- “Russia’s Peresvet, a ground-based laser system, is designed to dazzle enemy satellites’ optical systems to mask the movement of Russia’s strategic missile systems”
- “Russia has launched several satellites over the last few years which conducted test activities on orbit for technologies that could enable attacks on other countries’ satellites and result in temporary or permanent damage”
- China has a couple of spaceplanes, akin to the US’s hyper-secret X-37B vehicle, and they appear to be getting more sophisticated. In 2023, one of China’s aforementioned vehicles, “released and interacted with an object.”
Satellite-on-satellite fights: It’s obvious throughout the report that we’re in a world where rival nations have satellites that try to go up and mess with other satellites from other people.
- “In 2019, Russia launched satellite Cosmos 2542, which released a second satellite, Cosmos 2543. Cosmos 2543 then began following a U.S. national security satellite and approached it multiple times at a distance close enough to create potentially dangerous operating conditions,” the DIA writes. “Russia is developing kinetic orbital systems to permanently disable, or “kill,” satellites. Moscow’s Burevestnik is a developmental, co-orbital air-launched ASAT weapon for killing spacecraft in LEO”.
- “In February 2022, Russia launched a satellite called Cosmos, which Moscow claimed at the time was testing the effects of space radiation on electronic systems. This satellite was launched into an unusual and isolated orbit not used by any other spacecraft”.
Space nukes: Russia is experimenting with nukes in space. “A nuclear detonation in space would create an electromagnetic pulse, which would immediately affect satellites within range and could impact civilian infrastructure,” the DIA writes. “A nuclear detonation in space could be tailored to impact satellites in most orbits—LEO for commercial and reconnaissance satellites, GEO for critical communications, and MEO for navigation and timing satellites”.
Why this matters - infrastructure for the next phase of the AI power struggle: As people spend more resources on building space-based data centers (and eventually, perhaps, lunar-based ones), space will become ever more strategically and commercially valuable and ever more contested. Research like this gives us a sense of the hardware which is already in space and will arrive in the future. Perhaps in some far-off future an AI might seek to take control of such space assets as well.
Read more: DIA Releases Challenges to Security in Space (Defense Intelligence Agency, press release).
Read the report: Challenges to Security in Space (DIA, PDF).
***
Metis: A startup trying to fuse symbolic and neural reasoning:
…Towards Vulcan-like machine minds…
Metis, a startup that wants to build novel AI architectures oriented around developing machines with better reasoning, has published a research paper laying out its research approach. The high-level summary is simple - Metis wants to do for modern open-domain problems what DeepMind did for closed-domain problems like board games via its work on AlphaGo and AlphaGo Zero.
What they want to do: The key idea here is to add a reasoning core to modern AI architectures which is far more reliable than what we have today. “In modern AI architectures much of the other surrounding machinery that made AlphaGo a successful reasoning system remains either implicit or absent,” they write. “The first missing element is an explicit and persistent epistemic state… the second missing element is a criterion that allows the system to choose reasoning moves that make progress towards the goal.”
This is more of a position paper than a technical paper, so they don’t have a ton of detail here. The main idea is to train a reasoning policy and value function based on real-world data collection, then use this to guide the system to perform reasoning that reduces long-term uncertainty, and have these sit on a basic core of data made up of claims or hypotheses about the world. “Any change to a claim’s evidential status must be licensed by evidence or verified through explicit calculation and approved by an independent judgment system,” they write. “Model-generated content cannot enter the accepted epistemic state merely by being plausible; it must pass warrant assessment, which lowers hallucinations. Sources of evidence can be attached directly to the propositions they support, providing a clear chain of provenance. The reasoning process becomes explicitly auditable.”
Why this matters - building reliable substrates that are also deeply inhuman: Ideas like the architecture described here are fascinating both because of the promise they bring for increasing the reliability of AI systems, but also for how deeply inhuman they are: humans do not reason from some pristine base of facts, rather it’s all a big messy blob of memory and intuitions and some rules (some of which are learned from experience and some of which seem more platonic, like forms of logical reasoning). In this way, contemporary LLMs are more like humans than the system described here.
“While the history of AI can be understood as a contest between symbolic and neural methods, it may be more constructive to view it as a story of two incomplete successes,” Metis writes. “Classical systems made aspects of the reasoning process explicit but depended on representations, operators, heuristics and knowledge supplied largely by hand. Contemporary LLMs learn extraordinary semantic operations and vast bodies of implicit knowledge but leave the organisation, evaluation and accumulation of reasoning implicit.”
The AI system described here would ‘think’ in a way that is very different to how humans think and such a system might show us more about the differences between our own intelligence and that of machines.
Read more: The Age of Reasoning (Metis, PDF).
***
Google summons Genie 4 for better world modeling:
…One of Google’s most interesting projects continues to live!...
Google DeepMind has published details on Genie 4, a new iteration of its Genie world model. The company is rolling out access to Genie 4 currently as a research preview for some trusted testers and will eventually come to Google AI Ultra subscribers. The Genie series of models are fascinating, showing a world where you can take a photograph of anything and turn it into an interactable environment, or one where you can convert a 2D drawing into the stage for a videogame, and more.
Key features: Google calls out four features that make Genie 4 different to previous ones. These include:
- Real-time modification: “You can use natural language to modify a simulation in real-time — like changing the weather, moving the camera, or introducing a new person or object — and Gemini will contextually adapt your prompt to match the surrounding scene.”
- Physics and human action simulation: “Genie 4 simulates real-world physics scenarios, such as aerodynamics, collisions
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.