文章预览
LG - 机器学习 CV - 计算机视觉 CL - 计算与语言 1、[LG] Asking for Help Enables Safety Guarantees Without Sacrificing Effectiveness 2、[LG] Scaling Test-Time Compute Without Verification or RL is Suboptimal 3、[LG] LEAPS：A discrete neural sampler via locally equivariant networks 4、[LG] On Vanishing Gradients， Over-Smoothing， and Over-Squashing in GNNs：Bridging Recurrent and Graph Learning 5、[LG] Automated Hypothesis Validation with Agentic Sequential Falsifications 摘要：允许Agent寻求帮助既能保证安全又不牺牲效率、在没有验证或 RL 的情况下扩展推理时计算能力是次优的、通过局部等变网络实现的离散神经采样器、GNN 中的梯度消失过度平滑和过度挤压的统一视角、利用Agentic序列伪证自动验证假设 1、[LG] Asking for Help Enables Safety Guarantees Without Sacrificing Effectiveness B Plaut, J Liévano-Karim, S Russell [UC Berkeley] 允许Agent寻求帮助既能保证安全又不牺
………………………………
