
Beyond LLMs: Agents, Emergent Abilities, Intermediate-Guided Reasoning, BabyLM
Steven Feng, Div Garg, and Karan Singh present a session from Stanford's CS25: Transformers United, looking past single large language models toward systems built around them. The talk covers autonomous agentic AI systems that coordinate LLMs to act on multi-step tasks, and the emergent abilities that appear only once models are scaled past a certain size. It surveys intermediate-guided reasoning techniques, including chain-of-thought prompting, that break text generation into explicit steps before producing a final answer. The speakers also introduce BabyLM, a research effort to train small, efficient language models on roughly the amount of data a human child encounters, as a counterpoint to ever-larger pretraining runs. The session closes with discussion of the ethical stakes of deploying these systems. Delivered as part of Stanford's ongoing CS25 seminar series on transformer architectures.