DenoiseFlow: Uncertainty-Aware Denoising for Reliable LLM Agentic Workflows

Researchers introduce DenoiseFlow, a framework that addresses reliability issues in AI agent workflows by managing uncertainty through adaptive computation allocation and error correction. The system achieves 83.3% average accuracy across benchmarks while reducing computational costs by 40-56% through intelligent branching decisions.

• →DenoiseFlow addresses 'accumulated semantic ambiguity' where small errors compound in multi-step AI agent reasoning chains.
• →The framework uses three coordinated stages: sensing uncertainty, regulating computation allocation, and correcting errors through root-cause analysis.
• →Testing across six benchmarks shows DenoiseFlow achieves highest accuracy (83.3% average) with significant cost reduction of 40-56%.
• →The system performs online self-calibration without requiring ground-truth labels for continuous improvement.
• →The approach formalizes multi-step reasoning as a Noisy MDP with adaptive branching between single-path and parallel exploration.