Improving LLM Long Context Understanding via Synthetic Data and Adaptive Compression

Abstract

Recent innovations in large language models (LLMs) have led to their widespread use, but the long context problem remains a fundamental challenge. Transformer-based LLMs are constrained by the quadratic scaling of the self-attention mechanism, which restricts most popular LLMs to a context length of several thousand tokens. Many methods have been introduced to extend the context of LLMs, including the Activation Beacon approach. In this work, we propose two key advancements to the existing methodology. First, we generate long context synthetic data across a variety of tasks for training context-extended models, which can supplement or even replace expensive human-annotated data. Second, we introduce a novel two-pass, adaptive compression technique for more intelligent compression of long contexts. We find that the two strategies lead to orthogonal performance improvements on real-world long context tasks, resulting in an overall 4.2% increase in accuracy compared to the previous benchmark.