Go to ICLR 2025 Conference homepageOutput Alignment: A Top-down Approach to Length Generalization
Recently, large language models have exhibited impressive performance and surprising emergent properties. However, their abilities remain constrained by the preset context window of the Transformer architecture, and they continue to struggle with length generalization. In this work, we propose a new perspective on length generalization by focusing on the output distribution rather than the input, as most prior studies have done (e.g., through positional encodings or data structure). First, through case studies on simple synthetic tasks, we highlight the importance of output alignment---the consistency of output distributions across sequences of varying lengths. We then extend this observation to natural language tasks and introduce a metric named Long-Short Misalignment to quantify output alignment, finding a strong correlation between this metric and length generalization performance. Based on these insights, we propose a regularization loss during training that improves output alignment. Extensive experiments confirm the effectiveness of this approach. Overall, our work provides a novel perspective for understanding and enhancing length generalization in large language models.