📄 中文摘要
提出了一种稀疏推理时间对齐（SIA）方法，通过在生成轨迹的关键决策点进行干预，实现了对大型语言模型输出分布的精细控制。传统的标记级引导方法在每个解码步骤都依赖于密集干预，导致计算开销大且可能影响生成质量。SIA方法的关键在于只在重要的决策时刻进行干预，从而避免了不必要的密集干预，降低了计算成本，同时保持了模型的内在分布，提升了生成效果。该研究为推理时间对齐提供了一种更高效的解决方案。
稀疏交汇引导下的推理时间对齐
出处: Inference-time Alignment via Sparse Junction Steering
发布: 2026年2月26日
📄 English Summary
Inference-time Alignment via Sparse Junction Steering
This study proposes Sparse Inference-time Alignment (SIA), which enables fine-grained control over large language models' output distributions by intervening only at critical decision points along the generation trajectory. Traditional token-level steering methods rely on dense interventions at every decoding step, leading to substantial computational overhead and potential degradation in generation quality due to excessive deviation from the model's intrinsic distribution. The key insight of SIA is that dense interventions are unnecessary, allowing for a more efficient approach that maintains the model's inherent characteristics while reducing computational costs. This research offers a novel solution for inference-time alignment.