The field of AI safety and alignment is rapidly evolving, with a growing focus on developing innovative methods to ensure that large language models (LLMs) and vision-language models (VLMs) behave safely and align with human values. Recent research has explored the use of influence functions to prune harmful training examples, simplified reinforcement learning frameworks to incentivize safety alignment, and cost-constrained runtime monitors to detect and prevent misaligned outputs. Another area of research has focused on steering out-of-distribution generalization and improving safety via intent awareness for VLMs.

Noteworthy papers in this area include AlphaAlign, which proposes a simple yet effective pure reinforcement learning framework with verifiable safety reward to incentivize latent safety awareness. Other notable works include GrAInS, which introduces a gradient-based attribution method for inference-time steering of LLMs and VLMs, and SafeWork-R1, which demonstrates the coevolution of capabilities and safety in a multimodal reasoning model. LoRA is also a notable method for safety alignment of reasoning LLMs, which achieves high safety levels without compromising reasoning abilities. Layer-Aware Representation Filtering is another method that identifies safety-sensitive layers within the LLM and leverages their representations to detect which data samples in the post-training dataset contain safety-degrading features.