Understanding Synthetic Context Extension via Retrieval Heads

TL;DR

This paper investigates how synthetic data fine-tuning enhances long-context understanding in language models, revealing the role of retrieval heads and providing insights for improving synthetic data generation.

Contribution

It uncovers the relationship between synthetic data training and retrieval head development, offering mechanistic insights into long-context learning in LLMs.

Findings

Retrieval heads learned on synthetic data overlap with those from real data.

Strong correlation between retrieval head recall and downstream performance.

Retrieval heads are necessary but not sufficient for long-context tasks.

Abstract

Long-context LLMs are increasingly in demand for applications such as retrieval-augmented generation. To defray the cost of pretraining LLMs over long contexts, recent work takes an approach of synthetic context extension: fine-tuning LLMs with synthetically generated long-context data in a post-training stage. However, it remains unclear how and why this synthetic context extension imparts abilities for downstream long-context tasks. In this paper, we investigate fine-tuning on synthetic data for three long-context tasks that require retrieval and reasoning. We vary the realism of "needle" concepts to be retrieved and diversity of the surrounding "haystack" context, from using LLMs to construct synthetic documents to using templated relations and creating symbolic datasets. We find that models trained on synthetic data fall short of the real data, but surprisingly, the mismatch can be interpreted and even predicted in terms of a special set of attention heads that are responsible for retrieval over long context, retrieval heads (Wu et al., 2024). The retrieval heads learned on synthetic data have high overlap with retrieval heads learned on real data, and there is a strong correlation between the recall of heads learned and the downstream performance of a model. Furthermore, with attention knockout and activation patching, we mechanistically show that retrieval heads are necessary and explain model performance, although they are not totally sufficient. Our results shed light on how to interpret synthetic data fine-tuning performance and how to approach creating better data for learning real-world capabilities over long contexts.