ReDAct: Uncertainty-Aware Deferral for LLM Agents in Sequential Decision-Making
arXiv·medium signal
ReDAct addresses LLM agent hallucination in sequential tasks by introducing an uncertainty-aware deferral mechanism — when the agent's confidence drops below a threshold during multi-step execution, it defers to a fallback strategy rather than hallucinating forward. This is a practical pattern for production agent systems where one bad step cascades into catastrophic failure chains. The approach is model-agnostic and works across different agent architectures.