Recall with Reasoning: Chain-of-Thought Distillation for Mamba's Long-Context Memory and Extrapolation
This addresses the problem of long-context memory for Mamba models, offering an incremental improvement without architectural changes.
The paper tackles Mamba's limited long-context memory in practice by proposing Recall with Reasoning (RwR), a method that distills chain-of-thought summarization from a teacher model to improve performance; experiments show it boosts Mamba's long-context results on benchmarks like LONGMEMEVAL and HELMET against baselines while preserving short-context abilities.
Mamba's theoretical infinite-context potential is limited in practice when sequences far exceed training lengths. This work explores unlocking Mamba's long-context memory ability by a simple-yet-effective method, Recall with Reasoning (RwR), by distilling chain-of-thought (CoT) summarization from a teacher model. Specifically, RwR prepends these summarization as CoT prompts during fine-tuning, teaching Mamba to actively recall and reason over long contexts. Experiments on LONGMEMEVAL and HELMET show RwR boosts Mamba's long-context performance against comparable Transformer/hybrid baselines under similar pretraining conditions, while preserving short-context capabilities, all without architectural changes.