Enhancing Interactive Agents with Large Language Models: The SwiftSage Framework

With the rise of large language models (LLMs), reasoning capabilities have become central to model interaction. The article first explains the limitations of purely textual environments and introduces two benchmark suites: AFL World, a simple task set, and ScienceWorld, a complex, physics‑rich benchmark containing over 30 tasks and 200 objects.

Baseline methods such as DRNN reinforcement learning, knowledge‑augmented RL, and behavior‑cloning are discussed, highlighting their shortcomings in large action spaces and long‑term planning.

The core contribution is SwiftSage, a dual‑mode agent that uses a lightweight T5‑large model for fast action prediction and switches to a large LLM (e.g., GPT‑4) when the small model encounters difficulties. The system separates planning (LLM‑driven) from grounding (environment execution), reducing token usage and improving cost‑effectiveness.

SwiftSage’s architecture includes a small model that processes task descriptions, recent actions, observations, inventory, and environment state, while the large model answers five targeted prompts to generate detailed plans and sub‑goals. Experiments on ScienceWorld show that SwiftSage outperforms traditional baselines and other LLM‑based agents in both score and efficiency, especially on medium‑ and long‑duration tasks.

Limitations are noted: reliance on Oracle agents for training data, dependence on commercial LLM APIs, and challenges in obtaining rich feedback from real‑world robots. Future work aims to expand task domains, distill planning abilities into open‑source models, and integrate tighter perception‑action loops with real robotic hardware.

The article concludes with a Q&A addressing model choices, differences from Google’s multi‑agent work, and practical considerations for deploying the framework.