🤖 AI 速览
📋 文章元数据
- 发布时间
- 2026-10-13
- 类型
- ai-daily
- 字数
- 2824
- 阅读时长
- 14 min
2026-10-13 AI日更 | 编辑链补位自回归短板,智能体评估进入互设难度阶段 链接到标题
今日核心变化集中在推理与评估两端:AutoIndexer 用编辑链训练因果模型,弥补自回归无法事后修改的局限;SWE-Duel 以自扩展竞技场让智能体互设任务难度,推动评估从静态走向动态;Kairos 与代数整数量化则提供字节级多模态统一与无误差乘加路径。
📖 本期 Watch List 深度导读 链接到标题
今天有三组更新值得标记。一是推理与解码新范式:AutoIndexer 在“编辑链”上训练因果模型,补上自回归无法事后修改的短板;推测解码中的并行草稿上下文校正与弹性潜在递归则分别优化推理效率与循环模型。二是 Agent 评估方法论升级:SWE-Duel 提出自扩展竞技场,任务在评估时生成、难度由智能体互设;Deep-Research Agent 的提示优化研究测量跨主题迁移与过拟合;perturber 把语义保持扰动做成可复用服务。三是底层基础模型与量化:Kairos 探索字节级多模态统一,代数整数量化提供无误差乘加路径,适合关注基础设施的读者。
🌐 X 平台 AI 热点快讯 链接到标题
今天 bird 没有检测到符合条件的 AI 话题。
💡 大佬观点(Influencer Insights) 链接到标题
今日大佬观点暂缺,推荐阅读 Watch List 深度内容。
📚 附录:今日 Watch List 更新源列表 链接到标题
时间窗口:最近 3 天;覆盖 22 个源;共 32 条更新
All-In Podcast (A_full) 链接到标题
- Martin Shkreli: Big Pharma’s Blind Spot, Betting Against Quantum & What Prison Taught Him
- 发布时间:2026-10-12 23:49 北京时间
- 摘要:- IREN 是一个垂直集成的人工智能云平台,为人工智能训练和推理提供数据中心、计算和软件。
- Oracle 将数据、应用程序和基础设施连接起来,将人工智能转化为业务成果,并随着人工智能的发展而提供优化的灵活性、选择和控制。
- 安永帮助科技创新者从初创企业到退出企业再到大型企业。
- 我们致力于为每个人提供充分发挥潜力的工具,并确保所有人都能受益于技术。
- Keel基础设施拥有高性能计算和人工智能运行所需的电力、土地和连接,并以安全的能源资产和遍布北美的电网互连为后盾。
- EN 要点:
- (0:00) Martin Shkreli joins the Besties
- (0:55) Why quantum computing is a one-trick business and why he’s betting against it
- (7:42) The drugs big pharma sells too cheap and his journey from hedge fund trader to pharma
- (11:37) AI, drug discovery, and who actually wins
Stratechery by Ben Thompson (A_full) 链接到标题
- Personal Day
- 发布时间:2026-10-12 18:00 北京时间
- 摘要:- 15 美元/月或150 美元/年。
- 通过每周三封电子邮件或播客对当天新闻进行实质性分析。
- 策略采访。
- 采访领先的上市首席执行官、私营公司创始人,并与分析师同行进行讨论。
- John Gruber 和我自己制作的每周两次的播客:每集 15 分钟,一分钟不多也不少。
- EN 要点:
- Personal Day — No Update
ArXiv cs.AI (B_intro+search) 链接到标题
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12493v1 公告类型:新。
- 摘要:可解释的人工智能 (XAI) 工具通常被认为是通过使算法的各个方面更易于人类理解来增加对机器学习模型的信任的方法。
- 然而,过去十年积累的证据表明这一目标总体上失败了,反而促进了对错误预测的过度依赖,这随后可能导致算法厌恶。
- 相反,XAI 似乎主要充当调试工具,可用于获取对模型学习的特征等的轶事见解。
- EN 要点:
- arXiv:2610.12493v1 Announce Type: new
- Abstract: Explainable Artificial Intelligence (XAI) tools have often been presented as methods to increase trust in machine learning models by making aspects of…
- However, evidence accumulated over the last decade points to an overall failure of this objective, instead promoting overreliance on incorrect predictions, whic…
- Rather, XAI appears to function primarily as a debugging tool, useful for obtaining anecdotal insights into, for example, features learned by the model
EliteTable: Self-Improving Program Search for Executable Control Policies
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12536v1 公告类型:新。
- 摘要:EliteTable 是一种由大型语言模型 (LLM) 引导的自我改进的程序搜索算法,它使用高分解决方案的固定容量表来指导新提案、现有解决方案的编辑以及多个解决方案的重新混合。
- 对六个体育馆控制任务的性能进行了评估,每个任务都有十代,每代有五十个候选位置。
- 每个任务都通过十个独立初始化的搜索进行评估。
- EN 要点:
- arXiv:2610.12536v1 Announce Type: new
- Abstract: EliteTable is a self-improving program-search algorithm guided by large language models (LLMs) that uses a fixed-capacity table of high-scoring soluti…
- Performance was evaluated across six Gymnasium control tasks, each with ten generations and fifty candidate slots per generation
- Each task was evaluated across ten independently initialized searches
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12542v1 公告类型:新。
- 摘要:蒙特卡罗树搜索 (MCTS) 通常被提议用于调度有向无环图 (DAG) 工作流程,但大多数研究从未进行可以反驳它的比较 - 给出了一种更简单的方法来进行相同的搜索工作。
- 我们在受控协议下进行比较。
- 所有部门都收到相同的预算 - 建立和评估相同数量的候选时间表:MCTS 调度方法,预先注册为主要部门;一个随机重启列表调度程序,没有搜索树作为控制;以及教科书 UCT MCTS 分支作为同一域上的注册后续分支。
- EN 要点:
- arXiv:2610.12542v1 Announce Type: new
- Abstract: Monte Carlo Tree Search (MCTS) is often proposed for scheduling directed-acyclic-graph (DAG) workflows, yet most studies never run the comparison that…
- We run that comparison under a controlled protocol
- All arms receive an identical budget - the same number of candidate schedules built and evaluated: an MCTS scheduling method, preregistered as the main arm; a r…
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12593v1 公告类型:新。
-摘要:大型语言模型的持续改进越来越依赖于综合经验,然而交互式智能体的有效数据合成不仅需要可执行和可验证的轨迹,还需要为当前任务智能体提供学习价值的交互。
- 现有方法越来越多地纳入环境反馈、数据验证或学习者失败,但仍然面临反馈到决策的差距:观察到的弱点并不能直接指定未来的综合应该如何改变,或者它们是否可以由当前的综合空间表示。
- 为了解决这一差距,我们提出了 EvoSyn,这是一种学习反馈驱动的多智能体框架,它围绕可修改的状态组织合成,指定可以故意生成哪些交互因素以及应该在哪里分配合成工作。
- EN 要点:
- arXiv:2610.12593v1 Announce Type: new
- Abstract: The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents r…
- Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weakne…
- To address this gap, we propose EvoSyn, a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what inter…
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12596v1 公告类型:新。
-摘要:代理大型语言模型现在可以处理复杂的任务,但长期的战略推理尚未得到充分研究:大多数代理基准都是简短且间歇性的,长期任务的表现很难分解为存储的知识、推理和运气。
- 我们引入了 SpireBench,其中模型扮演完整的商业套牌构建者 Slay the Spire (StS),其中早期决策决定长期成功。
- 在游戏最高难度的 20 个大型语言模型 (LLM) 中,前沿模型击败了启发式基线,与普通人类玩家相当,但仍远远落后于专家。
- EN 要点:
- arXiv:2610.12596v1 Announce Type: new
- Abstract: Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short…
- We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success
- Across 20 large language models (LLMs) at the game’s highest difficulty, frontier models beat heuristic baselines and match the average human player but remain…
Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12598v1 公告类型:新。
-摘要:具有可验证奖励的强化学习(RLVR)通过可执行测试为软件工程代理提供可靠的监督,但二元奖励无法区分具有显着不同调试质量的轨迹。
- 我们研究当前策略是否可以判断自己的轨迹,并使用这些判断来提高任务绩效,而无需单独训练的奖励模型。 为此,我们引入了行为自我奖励(BSR),这是一种奖励细化方法,它重用策略来比较具有相同验证者结果的同一任务的轨迹之间可观察到的调试行为。
- EN 要点:
- arXiv:2610.12598v1 Announce Type: new
- Abstract: Reinforcement learning with verifiable rewards (RLVR) provides reliable supervision for software engineering agents through executable tests, but bina…
- We investigate whether the current policy can judge its own trajectories and use these judgments to improve task performance without a separately trained reward…
- To this end, we introduce Behavioral Self-Rewarding (BSR), a reward-refinement method that reuses the policy to compare observable debugging behavior among traj…
CLEAT: Learning Value-Guided Clarification for Interactive Agents
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12655v1 公告类型:新。
- 摘要:LLM 代理可以执行日益复杂的任务,但有效的协助还需要将其行动与用户可能未说明的目标和约束保持一致。
- 我们通过澄清来研究这个问题作为交互式意图对齐。
- 然而,只有当代理能够将答案转化为更好的决策时,澄清才有用。
- EN 要点:
- arXiv:2610.12655v1 Announce Type: new
- Abstract: LLM agents can execute increasingly complex tasks, but effective assistance also requires aligning their actions with goals and constraints that users…
- We study this problem as interactive intent alignment through clarification
- Yet clarification is useful only when the agent can turn the answer into better decisions
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12659v1 公告类型:新。
- 摘要:RLVR 以数据为中心的扩展重点是收集或生成可验证的训练问题。
- 然而,语料库大小本身并不能决定数据集是否为目标策略提供有用的学习信号。
- 为了指导有效 RLVR 数据的合成,我们通过实验将数据集有效性描述为其问题、目标策略和训练推出预算的函数。
- EN 要点:
- arXiv:2610.12659v1 Announce Type: new
- Abstract: Data-centric scaling of RLVR has focused on collecting or generating verifiable training problems
- However, corpus size alone does not determine whether the dataset provides a useful learning signal for a target policy
- To guide the synthesis of effective RLVR data, we experimentally characterize dataset effectiveness as a function of its problems, the target policy, and the tr…
Faithful and Complete LLM-Assisted Automated Construction of Hierarchical Hardware Knowledge Graphs
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12762v1 公告类型:新。
- 摘要:硬件数据表包含对于组件选择、系统集成和硬件兼容性验证至关重要的工程规范。
- 然而,手动检查非常繁琐且容易出错,因为数据表冗长且异构,表格、图形和文本中的信息密集且分散。
- 尽管大型语言模型 (LLM) 可以自动执行此过程,但即使文档适合模型的上下文窗口,原始的一次性处理也会忽略相关规范。
- EN 要点:
- arXiv:2610.12762v1 Announce Type: new
- Abstract: Hardware datasheets contain engineering specifications essential for component selection, system integration, and hardware compatibility verification
- However, manual inspection is tedious and error-prone because datasheets are lengthy and heterogeneous, with dense information fragmented across tables, figures…
- Although large language models (LLMs) can automate this process, na"ive one-shot processing omits relevant specifications even when the document fits within th…
Representation Alignment from a Few Paired Data via Distilled Optimal Transport
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12781v1 公告类型:新。
- 摘要:对齐两个独立的预训练单峰编码器通常需要数百万个配对示例。
- 在本文中,我们研究了只有 $N \sim 100$–$500$ 对可用的机制,而两侧的未配对数据都很丰富。
- 我们的方法通过与 $N$ 配对锚点的相似性来表达每个表示,这将两种模态放置在一个共同的坐标系中;将两个点云与在整个池上解决的单个熵最优传输问题相匹配,其边际约束抵消了击败最近邻匹配的中心性;并将生成的计划提炼成两个轻量级编码器,从而使对应关系变得归纳。
- EN 要点:
- arXiv:2610.12781v1 Announce Type: new
- Abstract: Aligning two independently pretrained unimodal encoders typically demands millions of paired examples
- In this paper, we study the regime where only $N \sim 100$–$500$ pairs are available, while unpaired data on either side is abundant
- Our method expresses every representation by its similarities to the $N$ paired anchors, which places both modalities in a common coordinate system; matches the…
ArXiv cs.CL (B_intro+search) 链接到标题
AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12545v1 公告类型:新。
- 摘要:自回归解码不太适合非因果顺序数据,例如交叉引用或未来依赖关系,因为与迭代人类推理不同,它们的输出无法事后修改。
- 扩散法学硕士部分缓解了这一限制,但仍然无法灵活地从预生成的上下文中插入或删除任意数量的标记。
- 我们将 AutoIndexer 设计为一个生成模型,具有新颖的训练和解码机制,可模拟编辑链,允许模型通过令牌插入、替换和删除来灵活地修改其先前的输出。
- EN 要点:
- arXiv:2610.12545v1 Announce Type: new
- Abstract: Autoregressive decoding is less suited to non-causally ordered data, e.g., cross-references or future dependencies, since their outputs cannot be revi…
- Diffusion LLMs partially mitigate this limitation but still cannot flexibly insert or remove arbitrary numbers of tokens from a pre-generated context
- We design AutoIndexer as a generative model with a novel training and decoding mechanism that simulates a Chain of Edits, allowing the model to flexibly modify…
SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12569v1 公告类型:新。
- 摘要:静态编码基准在发布时固定。
- 随着时间的推移,他们的任务会渗透到训练数据中,更强大的智能体最终会让他们饱和。
- 我们认为基准应该旨在自我扩展:它的任务和预言机是在评估时创建的,因此不存在提前泄漏的内容,并且其难度由参与的代理设置,而不是在发布时固定。
- EN 要点:
- arXiv:2610.12569v1 Announce Type: new
- Abstract: Static coding benchmarks are fixed at release
- Over time their tasks leak into training data, and stronger agents eventually saturate them
- We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak,…
Contextual Correction of Parallel Draftsfor Speculative Decoding
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12570v1 公告类型:新。
- 摘要:推测性解码通过使用轻量级起草器提出多个标记,随后由目标模型并行验证,从而加速大型语言模型推理。
- 最近的基于扩散的起草者使这种范例特别有吸引力:基于深度注意力的神经主干提供了强大的预测,同时并行起草多个未来位置避免了自回归生成的顺序成本。
- 然而,这种并行性是有代价的。
- EN 要点:
- arXiv:2610.12570v1 Announce Type: new
- Abstract: Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verifi…
- Recent diffusion-based drafters make this paradigm particularly attractive: deep attention-based neural backbones provide strong predictions, while drafting mul…
- This parallelism, however, comes at a cost
CLUE: A Contrastive Learning Framework for Unsupervised Event Type Induction and Annotation
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12580v1 公告类型:新。
- 摘要:事件本体是定义数据集中事件的基础框架。
- 事件类型归纳通常涉及领域专家,他们根据数据集的领域和范围定义创建事件类型及其名称的指南。
- 然而,这个过程在注释成本方面是资源密集型的,并且容易出现人为偏差。
- EN 要点:
- arXiv:2610.12580v1 Announce Type: new
- Abstract: Event ontology is the foundational framework for defining events in a dataset
- Event type induction generally involves domain experts who define guidelines for creating event types and their names based on the dataset’s domain and scope
- However, this process is resource-intensive in terms of annotation costs and is prone to human biases
One Model to Loop Them All: Elastic Latent Recurrence for Language Models
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12582v1 公告类型:新。
- 摘要:一种潜在的循环语言模型重复应用共享的层块,因此一个检查点可以通过更改循环计数来牺牲推理计算的准确性。
- 然而,在一个循环中,此类模型通常远远低于相同大小的非循环模型,这是一种单循环税,使得在延迟很重要时单独的非循环模型成为更好的选择。
- 我们提出了弹性潜在递归(ELR),这是一种无需这种税收即可训练循环语言模型的方法。
- EN 要点:
- arXiv:2610.12582v1 Announce Type: new
- Abstract: A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing…
- At one loop, however, such models typically fall well short of a non-recurrent model of the same size, a one-loop tax that makes a separate non-recurrent model…
- We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12601v1 公告类型:新。
-摘要:提示优化是强化学习的一种有吸引力的替代方案,用于适应语言模型代理,因为它成本低廉并且产生人类可读的工件。
- 然而,在实际任务中,它具有标量奖励强化学习所没有的失败模式。
- 读取训练示例的优化器可以将其答案写入提示中,因此分布内增益可能是记忆而不是可转移的行为。
- EN 要点:
- arXiv:2610.12601v1 Announce Type: new
- Abstract: Prompt optimization is an attractive alternative to reinforcement learning for adapting language-model agents, since it is cheap and produces human-re…
- On factual tasks, however, it carries a failure mode that scalar-reward reinforcement learning does not
- An optimizer that reads the training examples can write their answers into the prompt, so an in-distribution gain may be memorization rather than transferable b…
Self-Preservation in Language Models
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12602v1 公告类型:新。
- 摘要:持续学习(CL)寻求适应预训练模型,同时保留先前的能力。
- 然而,为保存提供的任务仅涵盖模型所知道的部分内容,而遗忘可能会超出这些范围。
- 最近的工作表明,语言模型可以通过生成有用的重播来补充这一有限的证据,而无需访问其原始训练数据。
- EN 要点:
- arXiv:2610.12602v1 Announce Type: new
- Abstract: Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities
- Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them
- Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data
GeopoliticalQA: Benchmarking Retrieval and Generation over Intelligence Reports
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12617v1 公告类型:新。
-摘要:地缘政治和战略情报报告构成了丰富的专家知识来源,在当前的政治气候中具有越来越大的价值,但它们的异质性和分析复杂性对检索增强问答提出了挑战。
- 我们引入了针对 415 份开源地缘政治和战略情报报告的扎根问答基准,其中包括 5,525 个基于文档的问答对,涵盖事实、推理、比较和风险导向问题。
- 我们提出了一个文档感知生成管道,在生成问题之前分析报告,并通过专家注释和法学硕士作为法官协议验证数据集质量。
- EN 要点:
- arXiv:2610.12617v1 Announce Type: new
- Abstract: Geopolitical and strategic intelligence reports constitute a rich source of expert knowledge and have increasing value in the current political climat…
- We introduce a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, comprising 5,525 document-grounde…
- We propose a document-aware generation pipeline that profiles reports before generating questions and validate dataset quality through expert annotation and an…
perturber: A Stateless Service for Semantics-Preserving Text Perturbations
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12667v1 公告类型:新。
- 摘要:对干净基准的大型语言模型 (LLM) 评估夸大了模型的稳健性:真实输入受到干扰,这些保留意义的变体通常会移动模型答案。
- 系统地研究这一点需要每次应用一个扰动,无论它们在哪里运行,都可以重复且相同,但现有工具将它们耦合到基准测试堆栈,而不是将它们打包为可重用的转换层。
- 我们推出了 pertuber,这是一种无状态服务和 Python 库,每次调用都只应用一个保留语义的扰动,该扰动取自 11 个系列的 68 个扰动的第一方目录。
- EN 要点:
- arXiv:2610.12667v1 Announce Type: new
- Abstract: Large language model (LLM) evaluations on clean benchmarks overstate model robustness: real inputs arrive perturbed, and these meaning-preserving vari…
- Studying this systematically requires applying perturbations one at a time, reproducibly, and identically wherever they run, yet existing tools couple them to a…
- We present perturber, a stateless service and Python library that applies exactly one semantics-preserving perturbation per call, drawn from a first-party catal…
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12669v1 公告类型:新。
- 摘要:大型语言模型正在成为 GeoAI 系统的核心,但它们的评估很少考虑地理空间数据需求的引出。
- 这一遗漏至关重要,因为如果不首先确定必要的数据集和约束,就无法可靠地支持面向检索的地理分析任务。
- 引入 GeoFitBench-LLM 作为评估 LLM 是否可以推断自然语言地理分析请求所需的数据集、支持资源和约束的基准。
- EN 要点:
- arXiv:2610.12669v1 Announce Type: new
- Abstract: Large language models are becoming central to GeoAI systems, yet their evaluation rarely considers the elicitation of geospatial data requirements
- This omission is critical as retrieval oriented geo-analytic tasks cannot be supported reliably without first identifying the necessary datasets and constraints
- GeoFitBench-LLM is introduced as a benchmark for assessing whether LLMs can infer the datasets, the supporting resources and the constraints required by geo-ana…
ArXiv cs.LG (B_intro+search) 链接到标题
Algebraic Integer Quantization for Machine Learning Applications
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12494v1 公告类型:新。
- 摘要:在本文中,我们提出代数整数格式(AINF)作为高度可配置和精确压缩计算的框架。
- AINF 的一个关键特性是它可以直接简化为纯整数算术,从而实现无错误的乘法累加运算。
- 我们开发了一种矩阵乘法引擎,利用这一特性,在现代硬件加速器上最有效的数值对应物的可预测因素内获得性能。
- EN 要点:
- arXiv:2610.12494v1 Announce Type: new
- Abstract: In this article, we propose algebraic integer number formats (AINF) as a framework for highly configurable and accurate compressed computation
- A key property of AINF is its straightforward reduction to purely integer arithmetic, enabling error-free multiply-accumulate operations
- We develop a matrix multiplication engine that exploits this property, obtaining performance within a predictable factor of the most efficient numerical counter…
Kairos: Toward a Byte-Level Foundation Model for Every Modality
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12495v1 公告类型:新。
- 摘要:在单个本地机器上训练的小语言模型几乎没有留下架构试验和错误的空间。
- Kairos 将双向门控 DeltaNet 线性注意力和共享投影上的滑动窗口注意力 (LiZAttention2)、分块注意力残差聚合、稀疏专家混合 (MoE) 前馈层和多尺度字节级编解码器结合在一起,在文本、图像、音频、视频、激光雷达和控制流共享的一个字节空间上使用统一噪声去噪目标进行训练;该报告在询问该组合是否有用之前先询问该组合是否完全优化。
- 我们为每个组件提供代码派生算法,包括跨会话循环状态内存门,然后运行收敛健全性检查:在 1–15M 参数处过拟合单个 440 字节句子,仅文本,具有用于编解码器和注意力选择的三个种子,以及针对七个其他因素的单次扫描。
- EN 要点:
- arXiv:2610.12495v1 Announce Type: new
- Abstract: Small language models trained on a single local machine leave little room for architectural trial and error
- Kairos combines bidirectional gated-DeltaNet linear attention and sliding-window attention over shared projections (LiZAttention2), block-wise Attention-Residua…
- We give code-derived algorithms for each component, including a cross-session recurrent-state memory gate, then run convergence sanity checks: overfittinga sing…
An Input-Dependent Readout Phase for Resonant Sequence Models
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12496v1 公告类型:新。
- 摘要:具有固定大小状态的序列模型避免了注意力的二次成本,但面临着设计选择。
- 输入相关的状态转换以基于 FFT 的训练为代价提供内容相关的内存,而固定转换保留 O(T log T) 卷积路径,但对每个输入以相同的方式读取内存。
- ResonatorLM,第二类的最新模型,通过每个头的固定学习相位读取一组阻尼复杂振荡器。
- EN 要点:
- arXiv:2610.12496v1 Announce Type: new
- Abstract: Sequence models with a fixed-size state avoid the quadratic cost of attention, but face a design choice
- An input-dependent state transition gives content-dependent memory at the cost of FFT-based training, while a fixed transition keeps the O(T log T) convolutiona…
- ResonatorLM, a recent model of the second kind, reads a bank of damped complex oscillators through a fixed learned phase per head
CircuitTree: Composing Circuits into A Syllable Transcriber
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12520v1 公告类型:新。
- 摘要:电路发现的目标是将整体神经网络分解为微小的模块化子网络,每个子网络都可以解释,每个子网络都执行可识别的子任务。
- 我们开发了一种将任意多个电路组合成单个模型的实用方法。
- 我们提出了一种贪婪的、密度缩放的合成算法,克服了之前通过朴素并集来合成电路的尝试的局限性,然后通过 CircuitTree 将其推广到 $N$ 电路,CircuitTree 是一种建立在 $\Theta(N\log N)$ 树组合计划上的新数据结构。
- EN 要点:
- arXiv:2610.12520v1 Announce Type: new
- Abstract: The goal of circuit discovery is to decompose a monolithic neural network into tiny modular subnetworks which are interpretable in the sense that each…
- We develop a practical method for composing arbitrarily many circuits into a single model
- We propose a greedy, density-scaled composition algorithm that overcomes limitations of previous attempts to compose circuits by naive union, then generalize it…
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12527v1 公告类型:新。
- 摘要:配电运营商需要了解表后光伏 (PV) 发电情况,但大多数住宅客户只能通过网络智能电表进行观察。
- 基于深度学习的光伏分解可以恢复这种可见性,但其运营价值取决于该模型是否服务于已知客户、新客户、未来时期或新区域。
- 我们使用四种部署机制来量化这种依赖性,这些部署机制将家庭转移与时间外推法分开,应用于澳大利亚电网智能电表数据,并对山核桃街数据进行跨数据集调整。
- EN 要点:
- arXiv:2610.12527v1 Announce Type: new
- Abstract: Distribution operators need visibility of behind-the-meter photovoltaic (PV) generation, yet most residential customers are observed only through net…
- Deep-learning-based PV disaggregation can restore this visibility, but its operational value depends on whether the model serves known customers, new customers,…
- We quantify this dependence using four deployment regimes that separate household transfer from temporal extrapolation, applied to Ausgrid smart-meter data with…
GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12528v1 公告类型:新。
- 摘要:连续血糖监测 (CGM) 基础模型将每一天编码为一个向量,该向量混合了该人是谁、当天发生了什么以及传感器如何测量它,因此测量差异会随着表示进入新的群体。
- 我们询问预训练是否可以将损害传输的测量组件放置在已知坐标中,以便部署以固定投影和没有目标数据的方式丢弃它们。
- GlucoPRISM 通过响应合约来监督 GlucoFM 主干网,以及当仅当天或仅传感器发生变化时表示的每个部分应如何变化的规则:64 维特征带应不随两者变化,48 维状态带仅跨天变化,16 维传感器带应跟踪设备。
- EN 要点:
- arXiv:2610.12528v1 Announce Type: new
- Abstract: Continuous glucose monitoring (CGM) foundation models encode each day as one vector that mixes who the person is, what happened that day and how the s…
- We ask whether pretraining can place the measurement component that harms transfer in known coordinates, so that deployment drops them with a fixed projection a…
- GlucoPRISM supervises a GlucoFM backbone with response contracts, rules for how each part of the representation should change when only the day or only the sens…
CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12530v1 公告类型:新。
- 摘要:免训练特征重用通过用缓存的近似值替换选定步骤的完整去噪计算来加速扩散变换器推理。 -现有方法主要改进跳过特征的重建,但我们发现一旦使用稳定的一阶重建器,增加重建复杂度带来的收益有限,而精确计算步骤的放置对保真度有更大的影响。
- 更重要的是,绝对浅探测变化受到输入轨迹位移和网络响应强度的共同影响,因此不能直接反映缓存风险。
- EN 要点:
- arXiv:2610.12530v1 Announce Type: new
- Abstract: Training-free feature reuse accelerates diffusion transformer inference by replacing full denoising computation at selected steps with cached approxim…
- Existing methods mainly improve the reconstruction of skipped features, but we find that once a stable first-order reconstructor is used, increasing reconstruct…
- More importantly, absolute shallow-probe variation is jointly affected by input-trajectory displacement and network response strength, and therefore does not di…
Market Impact Estimation via Causal Foundation Models
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12537v1 公告类型:新。
- 摘要:当交易者执行大订单时,相关问题不是\emph{价格会是多少?},而是\emph{如果我没有进行交易,价格会是多少?}因此,市场影响是一个因果量,但经典模型(Almgren–Chriss,传播者)将价格回归到流量,每当知情流通过信息和交易影响价格时,该模型就会出现偏差。
- 我们提出了一个概念验证\emph{因果市场模拟器}:一个在合成市场先验上训练的连续时间因果先验数据拟合网络(PFN),其中隐藏的信息流通过构造混淆了流量-价格关系,混淆强度是精确的每集旋钮。
- 在此先验中,朴素估计器的偏差遵循我们凭经验确认的封闭形式(高达 68% 高估),而 PFN 在高混杂情况下跟踪无偏真参数预言。
- EN 要点:
- arXiv:2610.12537v1 Announce Type: new
- Abstract: When a trader executes a large order, the relevant question is not \emph{what will the price be
- } but \emph{what would the price have been had I not traded
- } Market impact is thus a causal quantity, yet classical models (Almgren–Chriss, propagators) regress price on flow, which is biased whenever informed flow mov…
Efficient Learning of Survival Analysis Models Under Outcome-Dependent Truncation and Censoring
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12538v1 公告类型:新。
- 摘要:现代机器学习模型通常基于假设代表模型未来将面临的预测任务的数据进行训练,但实际数据集违反了这一假设。
- 我们专注于结果相关的截断,其中观察结果根据其标签出现在训练数据中。
- 我们的工作连接了两种截然不同的文献:截断下的高效学习和生存分析。
- EN 要点:
- arXiv:2610.12538v1 Announce Type: new
- Abstract: Modern machine learning models are typically trained on data assumed to be representative of the predictive tasks the model will face in the future, b…
- We focus on outcome-dependent truncation in which an observation appears in the training data depending on its label
- Our work connects two largely distinct literatures: efficient learning under truncation and survival analysis
Enriched Holomorphic Neural Networks for 2D and 3D Laplace Problems
- 发布时间:2026-10-12 12:00 北京时间
- 摘要:- arXiv:2610.12541v1 公告类型:新。
- 摘要:拉普拉斯方程的边值问题在拐角、边缘和交汇处附近会产生奇异行为,而平滑逼近器很难解决这些问题。
- 我们将全纯神经网络与从局部几何和边界条件导出的奇异函数相结合。
- 表示在构造上是调和的,因此不需要内部残差:网络仅根据边界数据拟合平滑残差,并且富集项捕获奇异行为。
- EN 要点:
- arXiv:2610.12541v1 Announce Type: new
- Abstract: Boundary-value problems for the Laplace equation develop singular behaviour near corners, edges and junctions that smooth approximators resolve poorly
- We combine holomorphic neural networks with singular functions derived from the local geometry and boundary conditions
- The representation is harmonic by construction, so no interior residual is needed: the network fits the smooth remainder from boundary data alone, and the enric…