Textual Equilibrium Propagation for Deep Compound AI Systems

Researchers introduce Textual Equilibrium Propagation (TEP), a new method to optimize large language model compound AI systems that addresses performance degradation in deep, multi-module workflows. TEP uses local learning principles to avoid exploding and vanishing gradient problems that plague existing global feedback methods like TextGrad.

• →Current global textual feedback methods like TextGrad degrade performance as AI system depth increases due to exploding and vanishing gradient problems.
• →TEP introduces a two-phase approach with local optimization followed by controlled global adaptation to maintain signal quality.
• →The method shows consistent accuracy and efficiency improvements over existing approaches, with gains increasing at greater system depths.
• →TEP preserves the practicality of black-box LLM components while enabling better optimization of complex compound AI systems.
• →The research addresses critical scalability challenges for deploying LLMs in multi-agent and tool-use workflows.