Soft latent reasoning extends chain-of-thought reasoning beyond discrete token sequences by using continuous concept tokens. However, effectively training models in this space remains challenging: conventional reinforcement learning methods mainly optimize sampled rollout-action likelihoods and therefore underuse the full distributions produced during soft rollouts. We propose **Full-Distribution Self-Distillation (FDSD)**, an off-policy training framework that treats stochastic soft rollout distributions as dense pseudo-labels. FDSD aligns the current model distribution with advantage-weighted rollout targets via KL minimization and importance correction, enabling the model to learn from the full rollout distribution rather than only from sampled-action likelihoods. Experiments on GSM8K, MATH-500, OlympiadBench, and out-of-domain benchmarks show that FDSD achieves consistent average improvements in mathematical reasoning across model scales while preserving OOD generalization, with only modest computational overhead. These results suggest that full-distribution alignment is a promising training direction for soft latent reasoning.

Recent advancements have demonstrated the advantage of converting pretrained large language models into powerful text encoders by enabling bidirectional attention in transformer layers. However, existing methods often require extensive training on large-scale datasets, posing challenges in low-resource, domain-specific scenarios. In this work, we show that a pretrained large language model can be converted into a strong text encoder without additional training. We first conduct a comprehensive empirical study to investigate different conversion strategies and identify the impact of the attention sink phenomenon on the performance of converted encoder models. Based on our findings, we propose a novel approach that enables bidirectional attention and suppresses the attention sink phenomenon, resulting in superior performance. Extensive experiments on multiple domains demonstrate the effectiveness of our approach. Our work provides new insights into the training-free conversion of text encoders in low-resource scenarios and contributes to the advancement of domain-specific text representation generation. Our code is available at https://github.com/bigai-nlco/Look-Both-Ways-and-No-Sink.

Semantic similarity between two sentences is inherently subjective and can vary significantly based on the specific aspects emphasized. Consequently, traditional sentence encoders must be capable of generating conditioned sentence representations that account for diverse conditions or aspects. In this paper, we propose a novel yet efficient framework based on transformer-style language models that facilitates advanced conditioned sentence representation while maintaining model parameters and computational efficiency. Empirical evaluations on the Conditional Semantic Textual Similarity and Knowledge Graph Completion tasks demonstrate the superiority of our proposed framework.