ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning

Fan Huang

Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tas…