1
Sequential Data Poisoning in LLM Post-Training
大模型后训练阶段面临全新威胁：顺序数据投毒攻击，揭示安全漏洞与防御方法。
arXiv:2606.04929v1 Announce Type: new Abstract: LLM post-training proceeds through multiple stages, e.g., supervised fine-tuning (SFT) followed by rei…
2
Dynamic Latent Routing
提出General Dijkstra Search与动态潜路由后训练方法，实现MDP子策略最优时序组合，为语言模型强化学习提供新范式。
arXiv:2605.14323v1 Announce Type: cross Abstract: We investigate the temporal concatenation of sub-policies in Markov Decision Processes (MDP) with ti…