ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoning
arXiv:2605.05737v1 Announce Type: cross
Abstract: 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 tasks. As a result, erro…