CoPE: Clipped RoPE as A Scalable Free Lunch for Long Context LLMs
This addresses the challenge of efficiently extending context lengths in LLMs, which is crucial for applications like long-document processing, but it is an incremental improvement over existing RoPE adaptation methods.
The paper tackled the problem of scaling Rotary Positional Embedding (RoPE) for long-context Large Language Models by proposing CoPE, a soft clipping method that unifies out-of-distribution mitigation and semantic modeling, resulting in significant performance gains up to 256k context length and establishing it as a new state-of-the-art.
Rotary Positional Embedding (RoPE) is a key component of context scaling in Large Language Models (LLMs). While various methods have been proposed to adapt RoPE to longer contexts, their guiding principles generally fall into two categories: (1) out-of-distribution (OOD) mitigation, which scales RoPE frequencies to accommodate unseen positions, and (2) Semantic Modeling, which posits that the attention scores computed with RoPE should always prioritize semantically similar tokens. In this work, we unify these seemingly distinct objectives through a minimalist intervention, namely CoPE: soft clipping lowfrequency components of RoPE. CoPE not only eliminates OOD outliers and refines semantic signals, but also prevents spectral leakage caused by hard clipping. Extensive experiments demonstrate that simply applying our soft clipping strategy to RoPE yields significant performance gains that scale up to 256k context length, validating our theoretical analysis and establishing CoPE as a new state-of-the-art for length generalization. Our code, data, and models are available at https://github.com/hrlics/CoPE.