Reproduce “Dynamic Long Context Reasoning over Compressed Memory via End-to-End Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhuoen Chen, Dongfang Li, Meishan Zhang, Baotian Hu, Min Zhang
A dual-system framework called LycheeMemory enables large language models to process and reason over ultra-long contexts, up to 1.75 million tokens, by combining compressed memory representations with dynamic, state-dependent recall and iterative reasoning. This approach delivers a 6x inference speedup and 2x reduction in peak GPU memory while maintaining competitive multi-hop reasoning accuracy.
Key insights
- Cognitively inspired architectures that separate compressed long-term memory from dynamic working memory can effectively scale LLMs to ultra-long contexts.
- Dynamic and state-dependent memory recall, conditioned on the evolving reasoning state, is crucial for accurate multi-hop reasoning over compressed information, outperforming static retrieval.
- Jointly optimizing memory compression and the reasoning process via end-to-end reinforcement learning significantly enhances the quality of compressed representations for downstream tasks, yielding optimal performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.