You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We introduce SwiftSage, a novel agent framework inspired by the dual-processtheory of human cognition, designed to excel in action planning for complexinteractive reasoning tasks. SwiftSage integrates the strengths of behaviorcloning and prompting large language models (LLMs) to enhance task completionperformance. The framework comprises two primary modules: the Swift module,representing fast and intuitive thinking, and the Sage module, emulatingdeliberate thought processes. The Swift module is a small encoder-decoder LMfine-tuned on the oracle agent's action trajectories, while the Sage moduleemploys LLMs such as GPT-4 for subgoal planning and grounding. We develop aheuristic method to harmoniously integrate the two modules, resulting in a moreefficient and robust problem-solving process. In 30 tasks from the ScienceWorldbenchmark, SwiftSage significantly outperforms other methods such as SayCan,ReAct, and Reflexion, demonstrating its effectiveness in solving complexreal-world tasks.
AkihikoWatanabe
changed the title
あ
SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex
Interactive Tasks, Bill Yuchen Lin+, N/A, arXiv'23
Jun 16, 2023
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: