潜在思想调优：通过融合信息在潜在标记中架起上下文与推理的桥梁

The research proposes Latent Thoughts Tuning (LT-Tuning), a framework designed to bridge context and reasoning in large language models (LLMs) by utilizing fused information in latent tokens. While explicit Chain-of-Thought (CoT) enhances reasoning capabilities, it confines model thoughts to discrete vocabulary, limiting flexibility. Recent advancements in reasoning within continuous latent spaces offer a promising alternative, allowing for more robust inference and computation. However, existing latent paradigms often encounter issues such as feature collapse and instability due to distribution mismatches when repeatedly using hidden states as input embeddings, as well as alignment challenges when relying on assistant models. LT-Tuning aims to address these issues by redefining the reasoning process in latent spaces, enhancing the overall performance of LLMs.