The field of large language models (LLMs) is moving towards a deeper understanding of their internal mechanisms and safety protocols. Recent research has focused on exploring the latent directions of reflection in LLMs, which has led to a better understanding of how these models evaluate and revise their own reasoning. Additionally, there is a growing interest in developing methods to detect deception and improve safety alignment in LLMs. Noteworthy papers in this area include: Unveiling the Latent Directions of Reflection in Large Language Models, which proposes a methodology to characterize reflection in LLMs and demonstrates the controllability of reflection. Safety Alignment Should Be Made More Than Just A Few Attention Heads, which introduces a targeted ablation method to identify safety-critical components in LLMs and proposes a novel training strategy to promote distributed encoding of safety-related behaviors. Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection, which proposes a white-box method to amplify a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace.