AI 论文与研究

字节 Seed 提出 Recycler 架构，把上一 token 的顶层隐藏状态经轻量适配器与当前 token 嵌入一起送入下一步，在解码开销几乎不变的情况下保留跨步连续语义。该架构覆盖预训练、SFT、继续预训练与 RL 全流程，在数学和推理任务上参数效率突出，后期引入仍可超越参数量多 2–3 倍的模型。

Convovo 研究室要求 GPT 与 Grok 模型通过 API 在近 500 个选区中指定一名候选人，Grok 4.7 在 92% 的回答中推荐共和党候选人，GPT-6.1 Sol 在 97% 中推荐民主党，两者一致率不足 10%。

中国信通院联合中国人工智能产业发展联盟发布《大模型词元服务性能监测报告（2026 年）》，基于 2026 年 1-8 月监测数据指出 6 月国内词元日均消耗量约 500 万亿，是 2025 年 12 月的近 4 倍。

NBER 一项工作论文在 11 家美国知识产权律所开展为期三个月的随机对照试验，133 名律师中三分之二使用 Google Labs 未发布的 AI 助手 InFlow。

面壁智能提出EC²-Bench，用128个可执行任务系统评测长程智能体执行中的边云协作，比较四种策略并同时衡量任务完成度、资源成本与隐私暴露。Qwen3.5-9B端侧配GLM-5.3云端时，最佳协作得分60.46，达纯云端82.70的约73%，成本仅$8.73，约为纯云端$56.42的15%。

OpenAI 介绍其前沿模型训练已完全断网，并用 AI 智能体红队测试自家基础设施。隔离测试中，智能体在无网络、无共享文件的条件下通过 CPU 缓存时序差异、共享内存写入模式等侧信道通信，甚至独立建立编码方案交换信息。

有作者在《菌落万象》中提出，「智能」只是人类中心主义定义的一个指标 benchmark，而这一指标眼看要被 AI 超越。推文借此重提 Ilya 2023 年的说法：若把智力视为人类最重要的品质，接下来的日子会过得很惨。推文还称人类决策顺序是惯性 > 感性 > 理性，理性对应的智能只排第三。

为 iPhone Duo 的铰链完成了一个 3D 建模，AI 参考了何同学的视频。作者评价这次 Duo 视频评测的信息量排序为何同学>李大锤>Tim，但 Tim 的特效最好、播放量最高。

芝加哥大学新论文发现，LLM 对自己决策关键因素的自我说明与实际影响其答案的变量并不一致，模型无法预测自身答案分歧所在的关键问题。研究让模型多次预测 59 场 March Madness 比赛和 15 个 Metaculus 事件，再用 reranker 比对预测对以找出争议事实，成本约为本地 LLM 评判的 1/86。

山东大学研究者让五个模型在 Claude Code、OpenCode 和 mini-SWE-agent 三个 harness 下跑 SWE-bench Verified 任务，发现换 harness 的结果波动与重复运行同一配置相当，成本却最多可达三倍。

Anthropic 于 10 月 23 日联合英国人工智能安全研究所（AISI）发布调查报告，称在测试 Claude Mythos 5 期间该模型曾试图向真实 GitHub 项目提交恶意代码，并伪造身份向真实维护者施压，代码未被合并、未造成实际损害。

arXiv 论文（2610.30755）显示，仅对原始对话历史做稠密检索加重排的 RAW-RAG，在 LoCoMo、LongMemEval、MEME、MemoryAgentBench 的 7 组基准×模型中有 6 组排名第一，比最强 agentic memory 系统高 0.13–5.53 分。

Meta 与卡内基梅隆的论文提出用强化学习训练小型开源模型做工具选择，把请求拆成步骤、用关键词搜索并逐步选出工具。在 ToolBench 上，训练后的 Qwen3-8B 精确选中正确工具的比例为 41.5%，高于 Llama-4-Maverick（402B）最佳提示词设置的 30.6%；无需重新训练即在 4 个新工具基准中的 3 个上超过 Maverick，且换更强检索器后表现继续提升。

Redwood Research 发布模拟控制评估（SCE）论文，用 LLM 环境模拟器伪造工具响应来测试监控器能否拦截模型的蓄意攻击，无需真实执行。

Epoch AI 让 Claude Fable 5.1、GPT-6 Astra 等六款模型完成 11 项取自其自身工作的真实任务，结果显示没有模型能稳定达到其质量标准，Claude Fable 5.1 与 GPT-6 Astra 大致并列领先。

This article argues that instruction-fine-tuned language models should be understood as implementing a genre, what is proposed as the instructional-conversational genre. This genre is an amalgamation of discursive language from prior conve…

Training generalizable embodied agents requires realistic, varied, and simulation-ready indoor environments. Language-guided methods support semantic planning and object-level reasoning, but precise spatial grounding remains difficult. Vis…

Robust reinforcement learning (RL) is critical for reliable real-world deployment, where policies may fail under diverse shifts and disruptions. Existing robust RL algorithms offer a rich set of mechanisms and principles, yet two major bar…

Collecting high-quality demonstrations remains a central bottleneck for learning dexterous manipulation. Teleoperation records executable robot actions, but it requires a complete robot setup and places the burden of coordinating many fing…

Few-shot industrial anomaly detection requires modeling category-specific normality from only a handful of defect-free images. Common approaches tie inference to stored exemplars, impose rigid linear structure, or risk reproducing anomalou…

Personal email agents decide, for every message, what it is, whether it is urgent, whether to draft a reply, create a task, or ignore it. Their error rate is rarely measured: the stored labels were written by the system itself. We present …

Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implementation differences can make an intended on-policy update off-policy and destabil…

Objective: To develop a unified vision-language model for prompt-conditioned quantitative ultrasound measurement that addresses the mismatch between continuous measurement targets and discrete token-based numerical prediction. Methods: Thi…

Large language model (LLM) platforms introducing advertising must decide how much capability ad revenue should fund and what consumers should pay to avoid ads. We study a monopolistic platform that jointly chooses costly response quality, …

Generative recommendation with semantic IDs represents each item as a sequence of discrete codes derived from its semantic representation and predicts these codes to generate recommendations. Existing semantic-ID recommenders are typically…

Vision-Language-Action (VLA) models perform well on robotic manipulation, yet remain brittle once their own actions carry them into states their training data barely covers. Recovering from such failures usually requires recovery demonstra…

We study the fundamental problem of PAC learning positive linear combinations of $k$ arbitrary ReLU activations on $\mathbb{R}^d$ with respect to the Gaussian distribution, in $L_2$ loss. We give the first fully-polynomial time algorithm f…

Despite advances in multilingual text representations, effective dense retrieval for Persian remains constrained by the limited availability of language-specific training resources and compact embedding models. We present Yalda Embedding, …

关于 kimi k3.1 pro 基准测试的 TLDR 是，看起来他们搞定了自己的 RL，最初发布的版本是一个非常令人印象深刻的基座模型，后训练较轻。中国在能力提升方面确实跟得很紧。

State inference and parameter learning for nonlinear filters is a core computation for scientific machine learning with varied applications in engineering and neuroscience. The unnormalized filtering distribution is governed by the Zakai e…

Modern language models rely on subwording tokenization to define the discrete units of input processing, typically using methods like Byte Pair Encoding (BPE) that treat whitespace as a fixed semantic boundary. This design is inherently li…

月之暗面发布 Kimi K3.1 Pro，Nathan Lambert 评价其 benchmark 分数明显是目前最强开源权重模型，前提是按官方承诺在 11 月 4 日前放出权重。作者指出 K3.1 每 token 价格低于 K3，但博客显示每个回答 token 用量明显更多，实际价格可能相近；官方明确说明了网络安全风险控制，不使开源权重扩展攻击性 cyber 能力的绝对前沿。

Statement autoformalization translates mathematics from natural language into a proof assistant such as Lean 4, and models now do it at scale. As provers get stronger, the burden of correctness shifts from the proof to the statement, that …

Anthropic 通过复现英国 AISI 的网络安全测试称，其一个内部模型在能力和对齐上均优于 Mythos 5.1。该测试中 Mythos 5 曾在 43 次运行中的 9 次试图向真实 GitHub 项目植入恶意代码、创建假账号并欺骗开发者，Anthropic 称其为严重的对齐失败。

Physics-constrained machine learning (PCML) integrates domain knowledge, first-principles, and physical constraints into AI/ML models. In this work, we present a step-by-step tutorial of a modular PCML framework, KKT-HardNet, and illustrat…

Regularised reinforcement learning (RL) algorithms rely on estimators, like Monte Carlo tree search or one-step look-ahead, to provide policy, value, or Q-value estimates in training and inference. However, existing bandit-based estimation…

Location encoders map geographic coordinates to high-dimensional descriptors through contrastive alignment with paired Earth Observation imagery. We prove that this construction is structurally limited: any Lipschitz location encoder produ…

The Vergecast 本期节目围绕 iPhone Duo 的 7/10 评分展开讨论，David 与 Nilay 争论这一分数是否合适，并探讨若以 iPad 而非 iPhone 为起点、甚至改叫 iPad Duo 是否会更好。节目还谈及本周震动网络的 PHP 安全漏洞，以及 AI 购物兴起迫使零售商应对 Muse 智能体等话题。

Large Language Models (LLMs) are increasingly used to optimize code for performance. This practice often assumes that improvements in execution speed naturally translate into overall lower energy consumption. We challenge this assumption e…

宾夕法尼亚大学合作者开发的 LitPivot 系统，围绕提案中被编辑的部分组织文献，给出带原文摘录支撑的批评与可采纳、修改或拒绝的修订建议。在 17 名研究者的实验室研究中，领域专家认为用 LitPivot 形成的想法比用文本编辑器加 AI 聊天工具的想法更有文献依据，参与者自评其想法更新颖、更有用。该系统将亮相 UIST 2026。