Reward Mediation and Centralized Knowledge Transfer for Multi-Agent Enterprise Search

TL;DR: Give all the little search agents a “reward coach” who helps share tips and tricks across many types of tasks, making everyone learn faster—like a team sharing pointers from a coach. This involves building a centralized reward shaping module (“reward agent”) for multi-task RL in search settings.

Research Question: Can a centralized reward agent facilitate more effective knowledge sharing and transfer between diverse search tasks and agents in large-scale enterprise environments?

Hypothesis: A centralized reward agent (as proposed by Ma et al., 2024) that dynamically shapes and distributes auxiliary rewards based on global task knowledge will improve multi-task RL efficiency and adaptation to new/unseen tasks compared to decentralized or static reward schemes.

Experiment Plan: - Implement a centralized reward agent in the KARL framework, providing shaped rewards to distributed task-specific policy agents.

• Benchmark learning speed, transfer efficiency to new tasks, and reward signal informativeness compared to standard RL and decentralized reward approaches.
• Evaluate on KARLBench and new composite enterprise search tasks, analyzing reward signal propagation and its effect on multi-task generalization.

References:

• 1. Chang, J. D., et al. (2026). KARL: Knowledge Agents via Reinforcement Learning.
• 2. Ma, H., Luo, Z., Vo, T. V., Sima, K., & Leong, T.-Y. (2024). Centralized Reward Agent for Knowledge Sharing and Transfer in Multi-Task Reinforcement Learning.