ReCache:面向工具增强型LLM智能体的高效键值缓存重用与压缩
该框架独立缓存资源表示,用资源级注意力与剪枝实现键值缓存重用压缩,内存降九成,首词加速近四倍,性能几乎无损。
完整、可扩展且稳健的优先级规划:最大容量下多机器人有序存取
提出最大容量多机器人有序存取优先级规划:可扩展、无死锁、近线性加速,且对扰动稳健。
局部令牌化的鲁棒时间序列水印生成模型
提出局部令牌化生成模型L-VQVAE与水印法LVQMark,抵御后编辑攻击,稳定检测力与假阳性。
AutoOR:规模化后训练大语言模型以自动形式化运筹学问题
提出AutoOR,用合成数据与强化学习后训练LLM,自动形式化运筹优化问题,达领先效果。
ICICLE:通过上下文文档扩展检索
提出一种上下文索引框架,将新增文档作为推理时证据,结合复制机制与校准,提升新文档检索并保持旧文档性能。
DELOS:对比深度学习用于开普勒测光中的低信噪比盲凌星搜索
DELOS用对比评分搜索开普勒低信噪比凌星,合成数据精度99.3%,比BLS/TLS性能提升15.5%/11.25%,快3-5倍/74-80倍。
截断低分,提升高分:基于排名分类的BoN式蒸馏
提出TUP策略:截断低排名样本,仅加权保留高排名尾部,闭式训练高效,理论与实验验证有效。
TempJail:利用字幕调度对大型视觉语言模型进行时间越狱攻击
TempJail优化字幕时间调度,利用时间漏洞对LVLM进行黑盒视频越狱,攻击成功率显著提升。
多模态视觉问答中的问题引导式证据获取
Q-Guide按问题定向获取证据,在DocVQA2026和Manga109超越基线,增益来自感知定向。
SAGE-XGBoost:空间增强图嵌入——数据稀缺下自然灾害易发性制图的机器学习框架
提出SAGE-XGBoost,融合空间图嵌入与数据增强,应对数据稀缺下的滑坡/野火易发性制图,AUC约0.97/0.95,优于现有模型。
RecPFN:面向基于上下文学习推荐的先验拟合网络
RecPFN将上下文学习引入序列推荐,基于合成因果先验预训练,少量域序列即可单次预测,性能最佳且鲁棒。
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Open-Vocabulary 3D Object Detection with Co-Distillation Discovery and Dual Guidance Robust Training
Core-KAN: Continuous Vision Kernels with Kolmogorov-Arnold Networks
Coupled Optimal Transport with Landmark Constraints
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
LoRA-GA$^2$: Low Rank Adaptation with Multi-step Gradient Adaptive Alignment
Inadvertent Context Leakage in Language Models
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
Auditing Cross-Lingual Fairness in Language Model Watermarking
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
Online Test-Time Adaptation for Generalizable Dynamic Graph Anomaly Detection
AutoLumNet: Monotone Optimal Transport for Single-Shot Exposure Correction
Towards Clinically Faithful Medical Image Captioning via Enhanced Vision-Language Alignment
A 360-Degree Vision Dataset for Learning Yaw Control on GPS-Denied Micro-UAVs in Disaster-Response-Relevant Environments
A knowledge-guided agentic framework for mitigating patient-context ambiguity in health queries
Stopping and Routing LLM Judge Panels
Separating Covariate Shift from Mechanism Change with Two Discriminators: CJSD, a Conditional Discrepancy with an Exact Covariate-Concept Decomposition
Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models
AvatarDynamizer: From Static to Dynamic Human Avatars via Generative Dynamic Textures
Dynamic Gated Cross-Modal Fusion with Sarcastic-aware Contrastive Regularization for Multimodal Sarcasm Detection
Answer-Level Trust Selection for Physical Vision-Language Reasoning
Unified and Efficient Point-Line Local Features
Natural Language Code Retrieval for 1C:Enterprise: An Open Benchmark and Efficient Bi-Encoder
DIFFCZSL: Compositional Zero-Shot Learning Regularized by Diffusion Representations
Robust Incomplete Multimodal Sentiment Analysis via Iterative Proxy Correction
Flow Matching Meets 3D Curvilinear Structure Segmentation in Medical Imaging
HealMed: Multilingual Evaluation of Large Language Models in Medicine
Breaking Planner Integrity Boundary: Enviroment State-Text Injection Attack on LLM-Driven Embodied Agents
PETA:面向虚拟筛选的参数高效测试时自适应
提出PETA,在测试时仅调整0.03%参数(LayerNorm)适配靶点,优于全量重训基线。
V-REX:面向兽医X光的高效专家VLM训练
兽医放射学中,从零训练的专用VLM以少量参数和数据超越大型基础模型,高效生成诊断报告。
STEP:基于分数的时间能量用于人体姿态视频异常检测
提出STEP框架,用PCA将姿态序列投影到白化PC空间,避免噪声产生不合理姿态,提升视频异常检测性能。
ID-VTG:图像消歧的视频时间定位
提出图像消歧视频时间定位任务,用图文查询定位实例动作,构建基准并验证框架最优。
G-MARK:基于知识图谱的协同驾驶多智能体接地推理
提出G-MARK知识图谱协同驾驶推理,遮挡精度+42.2%,控制误差-13.1%,通信负载减25.6倍。
已知光照下基于点的稀疏视图三维重建
提出基于β面元与伴随光传输的稀疏视图重建,仅用267个基元,倒角距离降低28.5%
从街景图像到街道质量指标:面向郊区15分钟城市的视觉语言推理
用视觉语言模型分析法国尼斯郊区街景,评估步行友好性,发现优质街道稀缺,展示AI支持大规模城市诊断的潜力。
G-CARL:面向患者的医学报告解读的基于证据的检查表对齐奖励学习
提出面向患者的医学报告解读任务,构建奖励学习框架与基准,提升事实准确性和需求契合度。