FST Triples LLM Continual Learning Sample Efficiency

Fast-Slow Training (FST) treats model parameters as slow weights and context optimized via text feedback as fast weights.

In experiments, fast adaptation was handled by context, while general reasoning patterns remained in the base LLM, reducing catastrophic forgetting and the decline in plasticity.

• On reasoning tasks, it was up to 3x more sample-efficient than RL that updates only parameters.
• The final performance ceiling was also higher, reducing KL divergence from the base LLM by up to 70%.
• Adaptability was better preserved when moving on to the next task after learning one task, allowing it to absorb new tasks more stably than RL in continual learning settings.