1
Trust Region Policy Distillation
将信任区域优化引入策略蒸馏，解决模型压缩中的策略漂移问题。
arXiv:2607.04751v1 Announce Type: cross Abstract: Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust…
2
Factual Retrieval in LLMs Is a Redundant, Distributed and Non-Contiguous Process
揭秘LLM事实检索机制：冗余、分布式、非连续的计算过程，挑战传统认知。
arXiv:2606.21345v1 Announce Type: new Abstract: Large language models (LLMs) store and recall factual knowledge, yet the precise mechanism of how enti…
3
Loss Landscape Poisoning: Targeted Extraction of Unseen Training Data from LLMs
揭穿LLM数据隐私漏洞：通过训练数据投毒定向提取未见样本，AI安全新威胁。
arXiv:2606.17110v1 Announce Type: cross Abstract: Large Language Models are increasingly trained on proprietary or sensitive data, from private health…
4
Agentic Molecular Recovery via Molecule-Aware Exploration
新论文提出「分子感知探索」框架，让AI代理更智能地恢复和设计分子结构。
arXiv:2606.05847v1 Announce Type: new Abstract: Text-guided molecular generation with LLMs often yields invalid SMILES. We argue that invalid drafts s…
5
LLMs Need Encoders for Semantic IDs Too
论文揭示多模态LLM同样需要专用编码器处理语义ID，扩展了模型设计思路
arXiv:2606.00324v1 Announce Type: cross Abstract: Multimodal LLMs use dedicated encoders to bridge non-language modalities (vision encoders for images…
6
MetaphorVU: Towards Metaphorical Video Understanding
新赛道崛起：ICML 2026 spotlight论文首次提出隐喻视频理解任务，突破多模态理解边界。
arXiv:2605.25461v1 Announce Type: new Abstract: Metaphorical videos are prevalent across various real-world scenarios to convey complex ideas, and und…
7
MONA: Muon Optimizer with Nesterov Acceleration for Scalable Language Model Training
Muon优化器新升级MONA，融合Nesterov加速突破局部极小，提升大模型训练效率
arXiv:2605.26842v1 Announce Type: new Abstract: The Muon optimizer has recently offered a promising alternative to AdamW for large language model trai…
8
When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search
教你利用激活几何诊断转向瓶颈，GRACE框架让概念工程优化更高效。
arXiv:2605.16362v1 Announce Type: new Abstract: Activation steering offers a lightweight way to control LLMs without retraining, but its effectiveness…