Teaching large language models how to absorb new knowledge
9 days ago
- #Machine Learning
- #Self-Adapting Models
- #Artificial Intelligence
- MIT researchers developed a method called SEAL (Self-Adapting LLMs) to enable large language models (LLMs) to permanently update their internal knowledge.
- SEAL allows LLMs to generate synthetic data (self-edits) from user inputs, similar to students creating study sheets, and then determine the best way to learn from this data.
- The model uses reinforcement learning to test and select the most effective self-edits, improving accuracy in tasks like question-answering and pattern recognition.
- SEAL outperformed baseline methods, increasing accuracy by nearly 15% in question-answering and over 50% in some skill-learning tasks.
- A limitation is catastrophic forgetting, where the model's performance on earlier tasks declines as it adapts to new information.
- Future work includes mitigating catastrophic forgetting and applying SEAL in multi-agent settings where LLMs train each other.
- The research aims to make LLMs more human-like by enabling continuous self-improvement in evolving environments.