论文

简化了 MoE 路由算法，设计出通信与计算成本更低的直观改进模型，并首次证明大型稀疏模型可以使用更低精度格式进行训练This work simplifies the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs and shows large sparse models may be trained, for the first time, with lower precision formats.

Pointer sentinel-LSTM 模型在 Penn Treebank 上以远少于标准 softmax LSTM 的参数量达到 SOTA 语言建模性能，并开源了 WikiText 语料库The pointer sentinel-LSTM model achieves state of the art language modeling performance on the Penn Treebank while using far fewer parameters than a standard softmax LSTM and the freely available WikiText corpus is introduced.

在视觉问答与视觉定位任务上对多模态紧凑双线性池化（MCB）进行了广泛评测，结果一致表明 MCB 优于去掉 MCB 的消融版本This work extensively evaluates Multimodal Compact Bilinear pooling (MCB) on the visual question answering and grounding tasks and consistently shows the benefit of MCB over ablations without MCB.

DASH (Drift Aware advantage SHaping) 根据每个推理片段是朝向还是偏离正确性来分配片段级 credit，在 overthinking 普遍的场景下取得最高准确率，同时减少 overthinking 行为，比基线实现更高效的自纠错。DASH (Drift Aware advantage SHaping), which assigns segment-level credit based on whether each reasoning segment leads toward or away from correctness, achieves the highest accuracy where overthinking is prevalent while reducing overthinking behaviors and achieving more productive self-correction than baselines.

基于欧盟法院判决构建的大规模法律 IR 数据集，支持法律信息检索方法的开发与严格评估，覆盖多种查询-文档粒度（案例对案例、段落对案例、段落对段落检索）。A large-scale legal IR dataset constructed from Court of Justice of the European Union judgments, which supports both the development and rigorous evaluation of legal IR methods, at multiple query-document levels (case-to-case, paragraph-to-case, and paragraph-to-paragraph retrieval).

提出 CALVER (Causal Axiom-Level VERification)，一种无需训练的对称验证器，根据 Pearl 的因果准则（包括 d-分离、backdoor 调整与干预）对结构化 trace 评分，并在不参考标准答案的情况下选择得分最高的候选。This work introduces CALVER (Causal Axiom-Level VERification), a training-free symbolic verifier that scores structured traces against Pearl's causal criteria, including -separation, backdoor adjustment, and intervention, and selects the highest-scoring candidate without consulting a reference answer.

提出 ReflectRL，一个轻量级即插即用框架，在 on-policy 训练中从 Golden Negative Trajectories 中学习：先利用这些 trajectory 引出 Reflective Reasoning，再通过 Reflective-to-Direct Policy Transition 将所学到的推理行为迁移回 Direct Reasoning。ReflectRL is proposed, a lightweight plug-and-play framework that learns from Golden Negative Trajectories during on-policy training, and first uses these trajectories to elicit Reflective Reasoning, then applies Reflective-to-Direct Policy Transition to transfer the acquired reasoning behavior back to Direct Reasoning.

提出 ARCHead，一种紧凑型 LM-head 压缩器，结合量化低秩核心、组级 INT4 残差以及基于激活导出度量拟合的低秩校正，在 block 量化器无法触及的大输出投影上实现互补压缩。ARCHead is presented, a packed LM-head compressor that combines a quantized low-rank core, group-wise INT4 residuals, and a low-rank correction fitted in an activation-derived metric to complements block quantizers by compressing the large output projection they can leave untouched.

提出 RestoreKV，在相同 KV 总预算下以可学习的恢复机制补充基于选择的 KV 压缩方案，显著降低压缩带来的性能损失。RestoreKV is introduced, which complements this selection-based formulation with learned restoration with learned restoration under the same total KV budget, and substantially reduces compression-induced degradation.

K-EXAONE 2.0 在 K-EXAONE 基础上进一步提升，并保持与开源权重模型的竞争力，其最大提升体现在 agentic coding 与长上下文理解上，在长上下文检索与安全性方面优势最为明显。K-EXAONE 2.0 improves over K-EXAONE and remains competitive with open-weight models, showing its largest gains in agentic coding and long-context understanding and its clearest strengths in long-context retrieval and safety.

提出 MultiPathFormer，一种自回归基础模型，将每条发射端–接收端链路表示为连续值路径 token 的有序序列，并通过下一路径预测进行预训练，证明路径级预训练可学习无线传播的可复用表征。MultiPathFormer, an autoregressive foundation model that represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and pretrains with next-path prediction, is presented, showing that path-level pretraining can learn reusable representations of wireless propagation.

提出 WorldCycle，一种自验证 RL 框架，从普通动作序列中构建闭合动作循环及其重复执行，并优化两个互补奖励：空间闭合奖励（强制镜像的前向与反向片段之间的对称性）以及时间一致性奖励（对齐多次循环执行间的状态）。WorldCycle is introduced, a self-verifiable RL framework that constructs closed action cycles and their repeated executions from ordinary action sequences, and optimizes two complementary rewards: a spatial closure reward enforcing symmetry between mirrored forward and reverse segments, and a temporal consistency reward aligning states across repeated cycle executions.

实验表明，在 Ego2Robot 合成数据与机器人数据上的联合预训练在多种扰动类型上一致提升分布外泛化能力，并在真实机器人部署中得到验证。Experiments show that joint pretraining on Ego2Robot-synthesized and robot data consistently improves out-of-distribution generalization across multiple perturbation types, with benefits validated on real-robot deployment.

提出 RSTG（Recovering Learning Signals via Adaptive Teacher Guidance），在最关键的位置选择性地、精确地施加蒸馏，显著优于朴素的 GRPO+OPD。RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), which applies distillation selectively and precisely where it matters most, and substantially outperforms naive GRPO+OPD.

提出 Answer-Backtracked Credit Assignment（ABC），一种面向长程搜索 agent 训练的细粒度信用分配框架，将稀疏的轨迹级结果转换为稠密的步骤级监督，对有用动作（即使在失败轨迹中）给予奖励，同时抑制错误或冗余动作。Answer-Backtracked Credit Assignment (ABC) is proposed, a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions.

提出 FocusMem，在紧凑的潜空间记忆接口中分离情景记忆与工作记忆，一致优于完全匹配的动作-only 固定记忆基线以及先前的潜空间记忆适配方法。FocusMem is introduced, which separates episodic memory and working memory within a compact latent-memory interface and consistently outperforms a fully matched action-only fixed-memory baseline and prior latent memory adaptations.

开发 PIMiner，一种用于 prompt injection 红队的 agentic 系统，可在训练阶段从零构建策略库，并在测试时无需额外训练直接迁移到未见过的目标 LLM。PIMiner is developed, an agentic system for prompt injection red-teaming that builds a strategy library from scratch during training and can be directly transferred to a previously unseen target LLM without additional training at test time.

提出 CoCoEvolve，通过定义显式的一一对应关系，并利用表征间的一致性优化模型，在无需额外标注的前提下提升图表、表格与代码表征之间的跨模态一致性。This work introduces CoCoEvolve, a method to improve consistency across chart, table, and code representations by defining explicit one-to-one correspondences and optimizing models using agreement between representations, without additional annotations.

提出 DRIFT（Denoising Redirection via Input perturbation of the Flow-matching Trajectory），一种测试期通用对抗补丁，贴在机器人夹爪上，针对现成策略的去噪速度场进行攻击。This work introduces DRIFT (Denoising Redirection via Input perturbation of the Flow-matching Trajectory), a test-time universal adversarial patch placed on the robot's gripper that attacks the denoising velocity field of an off-the-shelf policy.

文章论证了 Transformers 可作为医学图像分割任务的强大编码器，并通过与 U-Net 结合，恢复了局部空间信息以增强更精细的细节。It is argued that Transformers can serve as strong encoders for medical image segmentation tasks, with the combination of U-Net to enhance finer details by recovering localized spatial information.

文章提出了 SegFormer，一个简单高效且强大的语义分割框架，将 Transformers 与轻量级 MLP 解码器统一，并在 Cityscapes-C 上展示了出色的零样本鲁棒性。SegFormer is presented, a simple, efficient yet powerful semantic segmentation framework which unifies Transformers with lightweight multilayer perception (MLP) decoders and shows excellent zero-shot robustness on Cityscapes-C.

分析表明，VisualBERT 能够在无任何显式监督的情况下将语言元素对应到图像区域，甚至对句法关系敏感，例如能够跟踪动词与其论元对应图像区域之间的关联。Analysis demonstrates that VisualBERT can ground elements of language to image regions without any explicit supervision and is even sensitive to syntactic relationships, tracking, for example, associations between verbs and image regions corresponding to their arguments.

文章介绍了一种自监督视觉表征模型 BEiT（Bidirectional Encoder representation from Image Transformers），在图像分类和语义分割上的结果表明，该模型取得了与先前预训练方法相当的竞争性结果。A self-supervised vision representation model BEiT, which stands for Bidirectional Encoder representation from Image Transformers, is introduced, and results on image classification and semantic segmentation show that the model achieves competitive results with previous pre-training methods.

在输入和输出直接进行 4 倍下采样与上采样的设定下，实验表明，基于纯 Transformer 的 U 形编码器-解码器网络优于完全卷积或 Transformer 与卷积相结合的方法。Under the direct down-sampling and up-sampled of the inputs and outputs by 4x, experiments demonstrate that the pure Transformer-based U-shaped Encoder-Decoder network outperforms those methods with full Convolution or the combination of transformer and convolution.

BLIP 通过引导式 caption 方式有效利用含噪网络数据，由 captioner 生成合成 caption，并由 filter 去除噪声样本；在以零样本方式直接迁移到视频-语言任务时，展现出强大的泛化能力。BLIP effectively utilizes the noisy web data by bootstrapping the captions, where a captioner generates synthetic captions and a filter removes the noisy ones, and demonstrates strong generalization ability when directly transferred to video-language tasks in a zero-shot manner.

本文提出了首个用于睡眠阶段分类的深度学习方法，无需计算频谱图或提取手工特征即可端到端学习，利用了全部多变量多模态 PSG 信号（EEG、EMG、EOG），并能利用每个 30 秒窗口数据的时序上下文。This work introduces here the first deep learning approach for sleep stage classification that learns end-to-end without computing spectrograms or extracting handcrafted features, that exploits all multivariate and multimodal polysomnography (PSG) signals (EEG, EMG, and EOG), and that can exploit the temporal context of each 30-s window of data.

本文提出一种多任务深度学习的原则性方法，通过考虑各任务的同方差不确定性来加权多个损失函数，从而在分类与回归场景下同时学习具有不同单位或尺度的多种量。A principled approach to multi-task deep learning is proposed which weighs multiple loss functions by considering the homoscedastic uncertainty of each task, allowing us to simultaneously learn various quantities with different units or scales in both classification and regression settings.

提出一种用于符号音乐的分层自监督"世界模型"，采用 2.55M 参数的 Swin V2 编码器，在 MIDI 钢琴卷帘图像上以 JEPA 风格目标（音高与时间平移等变性、掩码嵌入预测以及分布正则化）训练，无需标签与乐理词汇。A hierarchical self-supervised ``world model'' for symbolic music is presented, using a 2.55M-parameter Swin V2 encoder trained on MIDI piano-roll images with JEPA-style objectives (pitch- and time-shift equivariance, masked embedding prediction, and a distributional regularizer), using no labels and no music-theory vocabulary.

EnvACE 是一种 Agentic 强化学习方法，用 world rehearsal 替代训练中的外部环境交互，将 world rehearsal 确立为突破外部环境约束、扩展 LLM Agent 训练的新路径。EnvACE is introduced, an agentic reinforcement learning method that replaces external environment interaction during training with world rehearsal, establishing world rehearsal as a new path toward scaling LLM agent training beyond the constraints of external environments.

CalibForge 是一个面向终端任务的自动合成系统，利用已验证的求解器行为，通过对抗式求解器校准来修订候选任务；消融实验表明，两种策略都比仅靠人工撰写加验证、或普通单求解器反馈产生更有效的监督信号。CalibForge is presented, an autonomous terminal-task synthesis system that uses verified solver behavior to revise candidate tasks through adversarial solver calibration and ablations show that both strategies yield more effective supervision than authoring and validation alone or ordinary single-solver feedback.

PaDoc 是一个布局驱动的解析器，将预测布局视为共享页面表示上的分支结构，在五个并发级别下均为最快的 end-to-end 解析器。PaDoc is proposed, a layout-grounded parser that treats the predicted layout as a branching structure over a shared page representation that is the fastest end-to-end parser at five concurrency levels and is the fastest end-to-end parser at five concurrency levels.

World-to-Wrist VLA (W2-VLA) 是一个用于细粒度机器人操作的 VLA 模型，具备任务条件下的未来腕部建模；W2-CoT 是一个合成流水线，生成描述操作进度、物理过渡线索和腕部局部证据的结构化标注。World-to-Wrist VLA (W2-VLA), a VLA model for fine-grained robot manipulation with task-conditioned future wrist modeling, and W2-CoT, a synthesis pipeline that produces structured annotations describing manipulation progress, physical transition cues, and wrist-local evidence.

SmartMage 是一个统一的 MLLM，动态调度异构模态以实现语义感知的 3D 场景理解，在五个 3D 场景理解基准上达到 SOTA，并在仅 RGB 视频理解基准上取得具有竞争力的结果。SmartMage is proposed, a unified MLLM that dynamically orchestrates heterogeneous modalities for semantic-aware 3D scene understanding and achieves state-of-the-art performance across five 3D scene understanding benchmarks, and attains competitive results on RGB-only video understanding benchmarks.

识别出像素级嵌入的不可见元数据痕迹，表明大规模语义监督（无论是类别标签还是亿级规模的 caption）在预训练中会自然引发元数据-语义相关性，导致模型将低层信号转化为预测特征。Invisible metadata traces embedded at the pixel level are identified, suggesting that large-scale semantic supervision, whether through categorical labels or billion-scale captions, naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features.

提出 OSReward，一个面向真实场景的高质量基准，用于评估 VLM 评判器在 CUA 轨迹上的表现；同时发布一个面向 CUA 社区的、带推理标注的开放轨迹判断语料库，以弥补大规模可靠 CUA 奖励的缺口。OSReward is introduced, a realistic, high-quality benchmark that evaluates VLM judges on CUA trajectories, and an open corpus of reasoning-annotated trajectory judgments for the CUA community, to close the gap in reliable CUA reward at scale.

提出构建经济世界模型（作为生成式引擎）的实施路线图：异质 Agent 在其中行动、交互、适应并与市场和制度共同演化，由此从内部生成经济动态。This paper develops an implementation roadmap for building economic world models as generative engines in which heterogeneous agents act, interact, adapt, and co-evolve with markets and institutions, thereby producing economic dynamics from the inside.