Synapse: Leveraging Few-Shot Exemplars for Human-Level Computer Control

This paper investigates the design of few-shot exemplars for computer automation through prompting large language models (LLMs). While previous prompting approaches focus on self-correction, we find that well-structured exemplars alone are sufficient for human-level performance. We present Synapse, an in-context computer control agent demonstrating human-level performance on the MiniWob++ benchmark. Synapse consists of three main components: 1) state-conditional decomposition, which divides demonstrations into exemplar sets based on the agent's need for new environment states, enabling temporal abstraction; 2) structured prompting, which filters states and reformulates task descriptions for each set to improve planning correctness; and 3) exemplar retrieval, which associates incoming tasks with corresponding exemplars in an exemplar database for multi-task adaptation and generalization. Synapse overcomes context length limits, reduces errors in multi-step control, and allows for more exemplars within the context. Importantly, Synapse complements existing prompting approaches that enhance LLMs' reasoning and planning abilities. Synapse outperforms previous methods, including behavioral cloning, reinforcement learning, finetuning, and prompting, with an average success rate of $98.5\%$ across 63 tasks in MiniWob++. Notably, Synapse relies on exemplars from only 47 tasks, demonstrating effective generalization to novel tasks. Our results highlight the potential of in-context learning to advance the integration of LLMs into practical tool automation.