Researchers have introduced A*-Thought-V2, a framework designed to optimize Chain-of-Thought (CoT) reasoning by reducing the computational and context costs associated with long reasoning chains. Rather than utilizing hard pruning, the method models CoT as a hidden-state trajectory and employs an explicit-implicit interleaved latent architecture. By projecting question, step, and solution representations into a 3D PCA space, the system measures the alignment of local transitions against the global question-to-solution direction. Steps aligned with this direction are kept as explicit text, while deviating steps—often representing correction or branch exploration—are compressed into continuous latent tokens. To facilitate training, the authors implemented stepwise embedding forcing, which pools redundant steps into single latent embeddings, and label forcing, which utilizes soft multi-modal vocabulary distributions instead of one-hot labels. Evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks, A*-Thought-V2 increased average accuracy by up to 2.6% and improved Accuracy per Computation Unit by 2.29×. Furthermore, the method reduced response length by as much as 50%, decreased preprocessing time by 94.6%, and lowered training time by up to 80.3%. Analysis indicates that latent states form a compact region distinct from textual states, with higher entropy at latent positions facilitating richer step-level feature learning.