Reproduce “In-Memory Learning: A Declarative Learning Framework for Large Language Models” 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
Bo Wang, Tianxiang Sun, Hang Yan, Siyin Wang, Qingyuan Cheng, Xipeng Qiu
A declarative learning framework called In-Memory Learning (IML) enables large language model agents to continuously improve their performance by distilling, refining, and applying knowledge from past experiences using natural language within their memory components. This approach facilitates autonomous self-enhancement without human-labeled data, yielding accuracy improvements over traditional in-context learning baselines.
Key insights
- LLMs exhibit distinct internal abilities for inference, induction, and revision, which are crucial for self-improvement, with varying strengths across different models (e.g., GPT-3.5-turbo strong in inference, Llama2-70b-chat strong in induction).
- The stability and effectiveness of the IML process are influenced by parameters such as 'momentum' (prompting strategies for note revision) and 'accumulation steps' (the amount of data processed before revision).
- Agents can encounter situations akin to 'local minima,' where they struggle to update existing notes despite contradictory new experiences, potentially due to the transformer's copy mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.