DenoiseFlow: Formalizing Agentic Reasoning as Noisy MDP Reduces Compute 40-56% While Hitting 83.3% Accuracy Across Benchmarks

arXiv·medium signal

DenoiseFlow models multi-step LLM agent reasoning as a Noisy Markov Decision Process where errors compound across steps, then applies three coordinated stages: uncertainty estimation at each step, adaptive routing between fast execution and parallel exploration based on risk level, and targeted error recovery via root-cause localization. The result is 83.3% average accuracy across math, code generation, and QA benchmarks (+1.3% over best baseline) while simultaneously reducing compute by 40-56% by avoiding unnecessary exploration. The Noisy MDP framing provides a principled foundation for reliability/efficiency trade-offs in production agentic pipelines.