AI 论文与研究

Google 与 NASA JPL 在 PNAS 发表 MAPL-EMIT，一个基于 Swin-S vision transformer 的深度学习框架，用 EMIT 高光谱卫星数据自动完成甲烷羽流的检测、定量与源头定位。

Google 的 WikiSkill 论文提出一个框架，让 AI 智能体将运行经验自动持久化到知识库（wiki）中，并据此持续调优技能以供复用。该方法与模型无关，进化后的技能可迁移至更小模型，有时甚至超越更大模型的表现。论文强调构建跨公司与项目的持久知识库以升级技能。

一篇论文提出FAR框架，将AI数学从"解题"转向"选题"，在文献中搜索开放问题并大规模尝试，再筛选出最值得专家关注的结果。在组合学试点中，该流程将51，110篇论文缩减至4，717个开放猜想，产出1，050个新解法，598个通过自动评判，77个推荐专家评审。人工抽查15项成果全部正确。

字节跳动新论文提出 Chain-of-Experience 方法，在上下文中保留先前尝试与反馈而非整理成简洁摘要。在 6 个数学、编程和知识基准上，自我反馈平均得分 71.0%，高于无反馈迭代求解的 66.8%；使用正确性或执行器反馈可达 79.3%。

清华NLP团队（OpenBMB成员）推出九歌推敲（Jiuge-Tuiqiao），入选EMNLP 2026系统演示。该工具将古诗创作建模为字符矩阵与锁定矩阵，用户锁定字词后模型在剩余位置硬约束生成，并附28，739条搭配、PPL排序名句等四类证据供用户决策。在50行完形任务中，结合名句证据将Hit@1从12%提升至60%；80次生成请求中格律与押韵合规率达100%。

亚马逊新论文表明，KV缓存策略不仅是推理优化，还能改变模型所需的训练方式。若模型推理时会遗忘部分上下文，训练时也应模拟这种遗忘--将微调与KV缓存策略匹配，可防止长上下文失效。在128k token测试中，全注意力微调模型在稀疏推理下常产生冗长无意义回答，而匹配缓存策略微调的模型能正常作答并停止；该方法支持任意缓存策略，可在40 GB A100上计算4B模型梯度。

This paper presents Murano, an open source framework for designing, running, and reproducing mechanistic interpretability studies of large language models, intended for researchers across disciplines. These studies often combine loading, r…

Large language model-based multi-agent systems are evolving from fixed interaction topologies toward dynamically orchestrated Agent Swarms. However, existing benchmarks are still largely based on single-agent or general-purpose agent tasks…

Antarctic sea ice concentration (SIC) forecasting is an important yet challenging task due to the coexistence of complex spatial structure, long-range temporal dependencies, and strong seasonal variability. Conventional convolution-based m…

Multimodal Large Language Models (MLLMs) are increasingly deployed in multi-image scenarios requiring complex reasoning across visual contexts. However, current MLLMs remain fundamentally limited by object hallucination-generating plausibl…

Embedding methods have become popular due to their scalability on link prediction and/or triple classification tasks on Knowledge Graphs (KGs). Embedding models are trained relying on both positive and negative samples of triples. However,…

LLM agents need to sustain goal-consistent reasoning across long multi-turn interactions under strict resource constraints. However, as the multi-turn context accumulates, it can destabilize the underlying LLM's internal representation of …

Removing a specific individual's information from multimodal large language models (MLLMs) is often needed after deployment, but existing methods rely on a retain set, which is hardest to obtain at that point, and rebuilding it recreates t…

Language models can answer from precomputed memory, a model's saved reading of a body of material, reused across requests instead of read again at each. This paper maps where that practice preserves correctness and the conditions under whi…

Reliable uncertainty estimation is a crucial requirement for deploying large language models (LLMs) and vision-language models (VLMs) in safety-critical settings, especially when the model parameters are not accessible (black-box). We prop…

While self-supervised approaches to reinforcement learning have achieved strong results by learning representations of states and actions, a key open question is the time scale over which actions should be modeled. Departing from the stand…

Optimizing absorption, distribution, metabolism, and excretion (ADME) is an important part of small molecule drug discovery. Many machine learning models have been built to predict ADME properties to facilitate this optimization process, b…

Timely post-disaster building damage assessment from satellite imagery is a critical engineering decision support task, yet it remains constrained by class imbalance, ambiguous intermediate damage states, and limited cross-event transferab…

Reinforcement learning (RL), particularly RL with Verifiable Rewards (RLVR), has recently emerged as a central paradigm for enhancing large language models' (LLMs) reasoning abilities, demonstrating remarkable effectiveness across reasonin…

As language-model compute continues to scale, high-quality training data is becoming an increasingly important bottleneck. Conventional next-token prediction supervises what follows a context but leaves the intermediate reasoning behind th…

Collecting natural-language referring expressions along with region annotations, such as masks or boxes, is a major bottleneck in visual grounding (VG), as annotators must write descriptions that distinguish target regions from visually si…

Synthetic data is increasingly used to train large language models (LLMs), yet its security implications remain poorly understood. Prior work on subliminal learning suggests that models can inherit behavioral traits from seemingly unrelate…

Organizing unstructured feedback text into hierarchical taxonomy is a fundamental challenge in NLP, particularly in domains where feedback arrives at massive scale in varied forms such as reviews, transcripts, and surveys. Existing approac…

Large language models are increasingly capable in general, but their utility can remain modest in niche or understudied areas. One approach to address this limitation is to specialise existing models through additional training on target-d…

With the development of large language models (LLMs), generative retrieval is becoming increasingly important in e-commerce scenarios. Current mainstream approaches typically use a two-stage training strategy: first train a product embeddi…

Robust estimation is a core computer vision task frequently tackled using sample consensus. However, traditional methods suffer from inefficient sampling as they struggle to identify effective minimum sets before hypothesis evaluation. To …

研究发现，恶意检索文档可提升模型token置信度与输出一致性，导致基于不确定性的检测器失效。作者将攻击下注意力集中于投毒文档而非分散于检索证据的现象称为"注意力崩溃"（Attention Collapse）。因此，仅检查最终答案或置信度可能漏报，监控检索文档间的注意力分布或可提前暴露投毒。

一项实验让AI智能体在6天、3000美元预算内产出两篇研究论文，结果均遭拒稿。执行层面毫无问题--智能体运行了数百次实验、调试GPU集群并完成LaTeX排版，但判断力缺失导致其面对负面评审时只会收窄论点、添加免责声明，而非重新设计实验。两轮实验结束时预算均剩余过半，说明问题在于想法匮乏而非资源不足。

Code as Worlds 面向物理推理的可执行世界表征的智能体发现 论文：https://huggingface.co/papers/2608.27549

新论文《Safety Does Not Compose》指出，自主智能体若跨迭代记忆，其安全系统就不能在迭代间遗忘，长期任务安全无法简化为重复运行安全的短期轨迹。攻击者可将恶意证据分散在多个看似良性的步骤中，仅基于轨迹的监控器无法积累足够上下文识别攻击。论文地址：arxiv.org/abs/2608.27141。

新论文提出 PILOT 框架，让智能体在任务运行中实时自我改进，而非等运行结束后才学习。监督者实时观察工作进程、可引导或中止，并将经验写入持久技能与记忆。在 6 项模型基准配置中 5 项排名第一，GLM-5.1 与 Kimi-K2.6 在 Terminal-Bench 2.0 上最佳通过率分别提升 14.6 和 12.4 个百分点，输出 token 量下降 42.9% 和 47.4%。

论文提出 Verification-Aware Training（VAT），一个即插即用框架，在训练每一步模拟投机解码的验证过程，把接受与拒绝模式转化为监督信号。

论文提出 Qwen-Drive-1.0，是迈向自动驾驶视觉语言基础模型的初步尝试，保留预训练 VLM 架构，将 3D 感知、视觉问答与运动规划统一在同一框架内。

研究者提出 MNIST-PRO 基准，把 MNIST 数字识别改造成带回看约束的序列化瞥视搜索任务，以隔离评测智能体的感知状态构建能力。评测覆盖十个多模态模型和四种记忆表征，发现模型在全可观测下表现良好，但部分可观测下暴露三类瓶颈：难以整合碎片化瞥视、过早停止探索、面对矛盾证据时难以修正早期错误信念。

论文提出 CAST，一种批判感知训练框架，将稀疏任务结果转化为动作级监督，用于批判学习与策略优化，通过分析智能体轨迹合成解释动作有效性的结构化理由。在动态工具调用基准上微调 Qwen3 系列模型，Retail 任务 pass^4 超过 GPT-OSS-120B 达 10% 以上，域外 Telehealth 场景额外提升 9%。

arXiv 论文提出面向持久智能体的感知中心架构 Pera，将智能体组织为感知与控制组件，持续从情景任务执行、内部上下文和环境变化中感知服务相关信号并构造生命周期任务。作者用 Pera 回顾性梳理近期工作并给出案例研究，将语言智能体的演进类比软件工程从小程序到大型程序的架构转变。

论文提出 Normalized Low-Rank Adaptation（NoRA），在训练中对 LoRA 的下投影矩阵做归一化，并发现仅在初始化时归一化也能改进标准 LoRA。在预训练、监督微调和强化学习中，NoRA 加速收敛、提升性能与训练稳定性并缓解灾难性遗忘，且不增加可训练参数和推理时计算。

论文提出 NavMCP，一个将 VLM 推理智能体与导航基础模型（NFM）执行器耦合的智能体脚手架框架，通过意图、观察、记忆三条通道协作，无需重训任一模型即可实现长时程探索。

研究团队发布 LightNav-0，一个紧凑的通用具身导航模型，通过双通道指向和残差向量量化动作 tokenizer 激发预训练 VLM 的空间智能，无需任务专用预测头。训练语料覆盖 2K+ 场景和 4K+ 小时具身导航数据，LightNav-0 在全部 10 个公开导航仿真环境中取得单目成功率 SOTA，真实世界评测显示跨机器人本体、场景及静态动态目标的零样本泛化能力。

论文提出 MineAmongUs，一个 3D 多模态 Among Us 沙盒，让 imposter 智能体通过言语与非言语联合行动欺骗船员，并配套可配置的 VLM 智能体框架 ARIA 和基于欺骗分类学的原子与弧级标注方案。