Reinforcement learning, multimodal generation, LLMs & agents, and physical AI.强化学习、多模态生成、LLMs & Agents 与 Physical AI。
TL;DR A benchmark for coding agents facing mid-execution changes. Across 120 scenarios and 12 change families, all 17 evaluated systems have lower success after the workspace changes.评测 Coding Agent 在执行过程中应对变化的能力。覆盖 120 个场景、12 类变化，17 个受测系统在工作环境变化后的成功率均下降。
TL;DR MindRL translates a shared reward into updates suited to autoregressive, parallel discrete, flow, and hybrid policies, with explicit controls on update reliability.MindRL 将共同奖励转化为适合自回归、并行离散、Flow 及混合策略的更新，并显式控制更新的可靠性。
TL;DR Intent-conditioned proposals and multi-intent GRPO expand and preserve diverse driving maneuvers for preference-based reinforcement learning.通过意图条件采样与多意图 GRPO，扩展并保留多样驾驶行为，为基于偏好的强化学习提供更丰富的候选轨迹。
TL;DR Temporal-aware sparse autoencoders reveal hierarchical concepts across diffusion timesteps, making diffusion transformers easier to interpret and control.用时间感知的稀疏自编码器提取扩散各时间步中的层次化概念，理解并控制 Diffusion Transformer 的生成过程。
TL;DR A black-box evaluation framework uses multi-objective optimization to study jailbreak success, contextual relevance, and detectability together.以多目标优化联合考察越狱成功率、上下文相关性与可检测性，构建黑盒大语言模型安全评测框架。
TL;DR The 3MAD dataset and MCM optimization expose cross-modal safety weaknesses in medical multimodal language models.通过 3MAD 数据集与 MCM 优化，揭示医疗多模态大语言模型在图文错配及跨模态攻击下的安全薄弱点。