ltzheng/Synapse

Implements in-context learning for web automation by retrieving semantically similar trajectory exemplars via FAISS-indexed memory, then constructing prompts that demonstrate task solutions to LLMs. Evaluated on MiniWoB++ and Mind2Web benchmarks with support for both zero-shot prompting and supervised fine-tuning on CodeLlama via LoRA, enabling agents to ground language models in interactive web environments.