Chain-of-Thought (CoT) has become a cornerstone of reasoning in large language models, yet its effectiveness is constrained by the limited expressiveness of discrete token sampling.Recent latent reasoning approaches attempt to alleviate this limitation by replacing discrete tokens with soft embeddings (probabilityweighted mixtures of token embeddings) or hidden states, but they commonly suffer from two issues: (1) global activation injects perturbations into high-confidence steps, impairing reasoning stability; and (2) soft embeddings quickly collapse toward the highest-probability token, limiting exploration of alternative trajectories.To address these challenges, we propose SeLaR (Selective Latent Reasoning), a lightweight and training-free framework.Se-LaR introduces an entropy-gated mechanism that activates soft embeddings only at lowconfidence steps, while preserving discrete decoding at high-confidence steps.Additionally, we propose an entropy-aware contrastive regularization that pushes soft embeddings away from the highest-probability token's direction, encouraging sustained exploration of multiple latent reasoning paths.Experiments on five reasoning benchmarks demonstrate that SeLaR consistently outperforms standard CoT and state-of-the-art training-free methods.

Accurate identification and classification of white blood cells are essential for diagnosing hematological malignancies and analyzing blood disorders. Existing approaches predominantly leverage masked autoencoders (MAEs) to extract intrinsic blood cell features through image reconstruction as a pretext task. However, these methods encounter two critical challenges: (1) their generalization performance deteriorates under domain shifts caused by variations in staining techniques, illumination conditions, and microscope settings, and (2) the learned data distribution often deviates from the true distribution of blood cell features. To overcome these limitations, we propose CD-CBC, a novel framework for cross-domain blood cell image classification that integrates contrastive representation learning with a denoising mechanism. CD-CBC consists of two key components: a LoRA-based segmentation anything model (LoRA-SAM) and a contrastive masked autoencoder (CMAE). LoRA-SAM mitigates shortcut learning in contrastive learning by eliminating background noise and platelet interference, while CMAE captures fine-grained semantic features and models spatial relationships, enhancing cross-domain robustness. Additionally, we introduce a denoising mechanism in the latent space, which guides the model to focus on unmasked patches during reconstruction, allowing it to better capture the true distribution of blood cell features. Extensive experiments on two benchmark blood cell datasets demonstrate that CD-CBC achieves superior cross-domain performance, reaching an average accuracy of 62.47%, which is 3.17% higher than the current state-of-the-art, thereby confirming its strong generalization capability.