研究人员推出了一种名为离散潜在推理（DLR）的新颖方法，旨在提高大型语言模型中潜在推理的可解释性和效率。DLR将连续潜在状态转换为离散令牌，其灵感来源于基于渲染的压缩技术。该方法旨在通过将离散符号监督与离散潜在令牌对齐，来解决连续潜在方法中常见的که instability和可解释性不足的问题。在Qwen3-VL和LLaMA-3模型上进行的多个推理基准测试表明，DLR的压缩率高达20倍，同时保持了可解释的推理轨迹，优于现有的潜在推理基线。 AI
影响 该方法可能带来更高效、更易于理解的LLM推理，从而可能降低推理成本并改善模型对齐。
排序理由 该集群包含一篇详细介绍LLM推理新方法的学术论文。
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
arXiv:2606.31779v1 Announce Type: cross Abstract: Language models typically reason via explicit chain-of-thought (CoT), generating intermediate steps token-by-token. Latent CoT offers an alternative: it performs multi-step reasoning in the model's hidden states, replacing decoded…
Language models typically reason via explicit chain-of-thought (CoT), generating intermediate steps token-by-token. Latent CoT offers an alternative: it performs multi-step reasoning in the model's hidden states, replacing decoded tokens with continuous representations for greate…
arXiv:2606.29712v1 Announce Type: new Abstract: Large language models achieve high reasoning performance via explicit chain-of-thought and reinforcement learning, but require long output sequences and extended inference time. Latent reasoning reduces this cost by shifting computa…
Large language models achieve high reasoning performance via explicit chain-of-thought and reinforcement learning, but require long output sequences and extended inference time. Latent reasoning reduces this cost by shifting computation into a latent space; however, continuous la…