Retrieval-Aware Distillation for Transformer-SSM Hybrids

State-space models (SSMs) offer efficient sequence modeling but lag behind Transformers on benchmarks that require in-context retrieval. Prior work links this gap to a small set of attention heads, termed Gather-and-Aggregate (G&A), which SSMs struggle to reproduce. We propose retrieval-aware distillation, which converts a pretrained Transformer into a hybrid student by preserving only these retrieval-critical heads and distilling the rest into recurrent heads. We identify the essential heads via ablation on a synthetic retrieval task, producing a hybrid with sparse, non-uniform attention placement. We show that preserving just 2% of attention heads recovers over 95% of teacher performance on retrieval-heavy tasks (10 heads in a 1B model), requiring far fewer heads than hybrids that retain at least 25%. We further find that large recurrent states often compensate for missing retrieval: once retrieval is handled by these heads, the SSM backbone can be simplified with limited loss, even with an 8× reduction in state dimension. By reducing both the attention cache and the SSM state, the resulting hybrid is 5–6× more memory-efficient than comparable hybrids, closing the Transformer–SSM gap at a fraction of the memory cost.

This paper studies how to make large language models remember information from long passages more efficiently. Today’s strongest models can look back over previous text, but doing so can require a lot of memory. We show that much of this remembering ability depends on a very small number of attention components inside the model. By keeping only those important components and replacing the rest with more memory-efficient alternatives, we can preserve most of the original model’s performance while using much less memory. This could help make long-context language models cheaper and easier to run.