Why Does the Effective Context Length of LLMs Fall Short?
This addresses a critical bottleneck in LLMs for users needing long-context processing, offering a significant performance boost without retraining.
The paper tackles the problem that large language models (LLMs) have effective context lengths shorter than their training lengths, attributing it to a left-skewed frequency distribution of relative positions. It introduces STRING, a method that shifts position embeddings during inference to improve performance, achieving over 10-point gains on benchmarks like RULER and InfiniteBench for models such as Llama3.1 70B and Qwen2 72B, surpassing commercial models like GPT-4-128K.
Advancements in distributed training and efficient attention mechanisms have significantly expanded the context window sizes of large language models (LLMs). However, recent work reveals that the effective context lengths of open-source LLMs often fall short, typically not exceeding half of their training lengths. In this work, we attribute this limitation to the left-skewed frequency distribution of relative positions formed in LLMs pretraining and post-training stages, which impedes their ability to effectively gather distant information. To address this challenge, we introduce ShifTed Rotray position embeddING (STRING). STRING shifts well-trained positions to overwrite the original ineffective positions during inference, enhancing performance within their existing training lengths. Experimental results show that without additional training, STRING dramatically improves the performance of the latest large-scale models, such as Llama3.1 70B and Qwen2 72B, by over 10 points on popular long-context benchmarks RULER and InfiniteBench, establishing new state-of-the-art results for open-source LLMs. Compared to commercial models, Llama 3.1 70B with \method even achieves better performance than GPT-4-128K and clearly surpasses Claude 2 and Kimi-chat.