AI 论文与研究

Meta AI 提出 MIRA，将长周期研究智能体拆成外层 meta-reasoner 与执行器：前者读取持久研究记录并生成下一项调查的工作单，后者负责执行，决策只发生在工作单边界。

用 Jev-as-a-Judge 提升智能体评估的可靠性。 https://x.com/omarsar0/status/2107472222398886011?s=20

上海人工智能实验室推出 SpatialASM 框架，通过两阶段训练将主动空间心理化（ASM）内化进 LVLM，使其能主动规划视角并想象对应观测。该框架包含 Allocentric Novel-View Generator、ASM 轨迹合成与视角想象强化学习三部分，在 Qwen3-VL 和 Qwen3.5 四个骨干模型上完成训练。

Meta 论文《RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models》（arxiv.org/abs/2609.39551）发现，让 2 个不同编码智能体互相审查补丁，比给单个智能体更大预算更能发现静默缺陷。

PhAI Labs 联合 Stanford、Oxford 和 Princeton 的团队发布 JEPA-Anything，将 Yann LeCun 提出的 JEPA 架构扩展为可覆盖物理、机器人、天气、生物等七个领域的统一世界模型，方法是把预测状态拆成四个正交因子分别预测再重组。

Google 新论文提出 AIM 方法，让研究智能体维护按主题排序的想法地图，并核查代码与想法是否匹配，避免代码偏离想法时学到错误经验。AIM 在两组任务上分别比此前最佳智能体 ScientistOne 高出 1.6 和 4.9 分，并以最高 3.1 倍的速度追平其最佳成绩。该方法在可能方案多、优质方案少的任务上收益最大。

Zero-shot coordination in embodied settings requires acting while the partner is intermittently out of view, leaving existing methods with ambiguous partner representations and uncertainty over hidden partner states. We propose Predicting …

Legal intelligence aims to support reliable decision-making across long-horizon legal processes involving evolving case states and multiple roles. However, real-world legal deployment exhibits substantial case heterogeneity in facts, evide…

Concept drift threatens production machine learning, yet the empirical behavior of multivariate two-sample drift detectors at scale remains under-characterized. Existing benchmarks rarely address the hundreds of millions of rows and high-c…

Achieving compositional concept generalization (CoCoGen), the ability to understand novel situations by recombining learned primitives, remains a fundamental challenge in artificial intelligence. Compositional semantic models such as Compo…

Cooperation with unfamiliar partners requires adapting to communication conventions that are not known in advance. We study this problem in a controlled Hanabi-derived environment with scripted hint generation, LLM-controlled receiving dec…

Product Identification has sprung up to become one of the most challenging problems in the automation of the retail industry. With the new industry 5.0 standards, automated inventory management, and catalog creation tasks are vitally impor…

Full-duplex spoken dialogue models listen and speak at the same time, enabling voice agents to have natural, low-latency interactions that turn-based systems cannot offer. However, they are commonly evaluated against single-sided interlocu…

End-to-end planners based on waypoint regression achieve strong open-loop accuracy, but they primarily learn to mimic expert geometry and remain difficult to adapt to deployment-time safety constraints. We propose a query-based cost-learni…

Covariate-aware time-series foundation models (TSFMs) promise training-free what-if answers for instrumented plants: the change in output that a different future input would cause. We test this on forced engineering systems with exact coun…

Path-following control strategies typically follow the bi-objective optimization dilemma: minimizing deviations from a reference path while maintaining smooth speed profiles. The latter objective is especially relevant for Electric Vehicle…

Diffusion language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) language models, offering flexible token-update orders and parallel decoding. Recent dLLMs are often initialized from pretrained AR models bef…

Attacker ASV systems for voice anonymization have been studied primarily in English, leaving their behavior in multilingual settings largely unexplored. Conventional ASV has shown that both acoustic and contextual information are important…

Multimodal large language models (MLLMs) advance rapidly, while conventional benchmark development lags behind, delaying investigation of newly observed capability gaps. Such investigation requires an expressive task format and an on-deman…

Conversational MLLM agents are increasingly expected to assist in professional workflows, from AI research and engineering design to product management and business operations. Yet this capability remains underexplored: existing benchmarks…

Shared memory coordinates agents' actions, but correct records do not establish that those actions satisfy task requirements. Memory governance and failure diagnosis regulate or inspect recorded information; they do not by themselves estab…

AI agents increasingly rely on modular third-party "skills" that are dynamically selected by skill routers to execute complex tasks. While recent studies highlight the threat of prompt injections embedded in these skills, existing evaluati…

Mathematical problem solving often requires deterministic computational steps that agents delegate to tools and implicitly trust. Yet tools can fail silently, returning plausible but incorrect results. How well can agents detect and correc…

Natural-language access to RDF knowledge graphs is a core Semantic Web ambition. Large language models (LLMs) have advanced Text-to-SPARQL, yet on unfamiliar graphs they often generate valid queries that misrepresent the populated data mod…

Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availability of high-quality, expert-annotated training data. This challenge is exacerbated by stringent privacy…

IPv6 extension headers (EHs), such as fragmentation, segment routing, and in-situ telemetry, are operationally important yetwidely dropped in transit, and characterising their behaviour from packet captures is a recurring measurement probl…

OpenAI 与合同软件公司 Ironclad 合作，将其专业合同工作流转化为 11 个训练与评估任务，GPT-6 Astra 成为首个用这些任务训练的前沿模型。在研究评估中，Astra 平均得分为 55.0%，GPT-5.6 Sol 为 41.6%，估算每次尝试平均用时从 37.0 分钟降至 19.2 分钟；训练使用基于 SEC EDGAR 公开合同的合成任务和强化学习，未使用客户数据。

PhAI Labs、CUHK、复旦、斯坦福、牛津与普林斯顿的研究者发布 JEPA-Anything，一个领域无关的世界模型框架，用共享学习配方覆盖视觉、生物、临床轨迹、控制、分子动力学、物理场和天气 7 个领域。

Q Labs Research 发布 Dust，一种零阶优化方法，通过在每个 token 上独立扰动激活（虚拟种群）预训练 GPT 式 Transformer，无需反向传播。

AI 写学术论文的一个微妙问题在于，AI 没有针对反馈进行校准。AI 往往直接就妥协了，但如果你把论文送审，同行评审的质疑应该促使你为自己的想法做一些辩护（当然如果确实错了还是要放弃）。

arXiv 论文提出 PAIR，通过从同一智能体状态分别带压缩和不带压缩地续跑来定位压缩造成的影响，而非比较整条随机的运行轨迹。研究发现有害压缩会丢掉尚未解决的任务条件或把已读的 API 规格缩成模糊描述，导致重开文档和重新登录、多出约 5 步；PAIR 据此重写压缩提示词的对应部分，不改智能体、压缩模型和工具。

一项在田纳西州 18 所中学开展两年期整群随机对照试验发现，在每日补救数学课中使用带 AI 家教 Khanmigo 的 Khan Academy，学生数学成绩每学期提升 1.3 个全国百分位，约合每学年 0.06 到 0.08 个标准差，与不用 AI 的 Khan Academy 练习效果相近。

Apple 研究团队提出 RISED，将评分量表（rubrics）从奖励信号扩展为多环境 LLM 智能体训练中的数据选择与策略监督工具。该方法由 LLM judge 用跨环境共享的评分量表词表标注每条 rollout，据此挑选与混合环境批次行为构成一致的数据，并用正向量表为正策略自蒸馏教师提供 token 级监督、负向量表引导后续生成避开重复失败模式。

论文提出 Personal-Agent Mediated Recommendation 范式：平台推荐系统用本地信息排序候选集，个人 LLM 智能体再借助用户授权的跨平台历史对排序进行中介，产出最终 top-K 列表。

UNREAL 框架从冻结 LLM 的内部表示直接导出检索查询并编码 chunk，仅新增不到 500K 可训练参数，统一覆盖语料检索与长上下文推理。

研究者提出面向端侧级联语音智能体的推测性工具执行方案，通过 Predictor 模块从部分 ASR 假设中预测工具调用，在用户说话时提前执行并缓存结果，再注入 LLM 提示词以加快响应。在完整实现的 Android 语音助手上，该方法将首音频响应中位时间从 5.79s 降至 4.60s，标准差从 3.49s 降至 2.81s。

CtrlCache 是一个免训练的控制感知缓存框架，通过动作感知调度与刷新策略，将视频块标记为初始、过渡、转向或稳态，仅在部分去噪步骤复用 Transformer 残差，并引入频率混合历史先验引导。

针对长时程具身任务中目标常不在视野内、状态连续变化的问题，研究者提出 Attacca，用与执行环境解耦的目标图像在完整"搜索到交互"轨迹上训练视觉目标条件策略。该方法在 Minecraft 多项短时程与长时程任务上取得 39.0-47.5% 的 clean success，较最强基线提升 1.7-2.4 倍；长时程任务完成率达 54%、30% 和 28%，最高提升 7 倍。

研究重新审视跨 Tokenizer 的 On-Policy Distillation（OPD），发现严格 1:1 对齐已覆盖大部分学生生成 token，将 reverse KL 限制在学生选出的共享词表 top-16 子集即可达到全共享词表 OPD 的精度，并优于所评估的跨 Tokenizer 基线。

HiPLEX 是一个将预训练全双工文本策略因子化为控制策略与条件内容策略的强化学习框架，前者决定何时输出内容（pad/epad/con），后者仅在 con 时选择 token。