Harmonizing Dense and Sparse Signals in Multi-turn RL: Dual-Horizon Credit Assignment for Industrial Sales Agents

Researchers propose Dual-Horizon Credit Assignment (DuCA), a new framework for optimizing large language models in industrial sales applications. The method addresses training instability by separately normalizing short-term linguistic rewards and long-term commercial rewards, achieving 6.82% improvement in conversion rates while reducing repetition and detection issues.

• →DuCA framework solves training instability in LLMs by separating optimization across different time scales for industrial sales applications.
• →The method achieved 6.82% relative improvement in conversion rate compared to baseline methods.
• →Inter-sentence repetition was reduced by 82.28% and identity detection rate lowered by 27.35%.
• →Horizon-Independent Advantage Normalization (HIAN) ensures balanced gradient contributions from both immediate and long-term objectives.
• →The research addresses the challenge of balancing commercial performance with natural language generation quality.