Reflect before Act: Proactive Error Correction in Language Models
This addresses the problem of robust self-correction for LLMs in interactive tasks, offering incremental improvements over existing methods.
The paper tackles error accumulation in LLM-based decision-making by introducing a 'Reflect before Act' (REBACT) approach that adds a reflect step before actions, resulting in improved success rates by up to 24% on WebShop, 6.72% on ALFWorld, and 0.5% on TextCraft.
Large Language Models (LLMs) have demonstrated remarkable capabilities in interactive decision-making tasks, but existing methods often struggle with error accumulation and lack robust self-correction mechanisms. We introduce "Reflect before Act" (REBACT), a novel approach that enhances LLM-based decision-making by introducing a critical reflect step prior to taking the next action. This approach allows for immediate error correction, ensuring smooth action path and adaptibity to environment feedback. We evaluate REBACT on three diverse interactive environments: ALFWorld, WebShop, and TextCraft. Our results demonstrate that REBACT significantly outperforms strong baselines, improving success rates by up to 24% on WebShop (achieving 61%), 6.72% on ALFWorld (achieving 98.51%), and 0.5% on TextCraft (achieving 99.5%) using Claude3.5-sonnet as the underlying LLM. Further analysis reveals that REBACT's performance improvements are achieved with only a few modification steps, demonstrating its computational efficiency.