Small Language Models (SLMs)

Small language models (SLMs) are expected to play a central role in large-scale agent orchestration thanks to their low inference cost, yet they face challenges of their own — at 3B parameters and below, limited memory capacity and a sharp drop in agentic abilities such as tool use. We study post-training methods that draw out the capabilities of SLMs, and their integration with non-parametric external memory that compensates for what they lack. Robot foundation models also fall into this size class, making them a promising application.