The field of transformer research is moving towards improving the chain-of-thought reasoning capabilities of these models. Recent studies have focused on developing new methods for fine-tuning transformers to acquire this capability, including reinforcement learning and supervised fine-tuning. Notably, researchers have found that these approaches exhibit distinct learning behaviors, with reinforcement learning learning the whole chain-of-thought simultaneously and supervised fine-tuning learning the chain step-by-step. Another area of research is the development of new architectures and frameworks that can improve the stability and effectiveness of test-time adaptation in reasoning models. Overall, these developments are advancing the field of transformer research and improving the capabilities of these models. Noteworthy papers include: Transformers with RL or SFT Provably Learn Sparse Boolean Functions, But Differently, which provides theoretical insights into the underlying mechanisms of RL and SFT. SPINE: Token-Selective Test-Time Reinforcement Learning with Entropy-Band Regularization, which proposes a token-selective test-time reinforcement learning framework that improves Pass@1 over traditional test-time reinforcement learning methods. Softmax Transformers are Turing-Complete, which proves that length-generalizable softmax CoT transformers are Turing-complete. Learning When to Stop: Adaptive Latent Reasoning via Reinforcement Learning, which develops adaptive-length latent reasoning models that can reduce compute usage and improve compressive capabilities.