Human-in-the-Loop Evaluation and Interactive Feedback for RL-Based Knowledge Agents

TL;DR: Let real users—not just benchmarks—give feedback to search agents, so they get better at answering tough or weird questions and explain themselves—think of it as letting students ask questions in class and grade the answers. The experiment will build interactive human feedback loops into KARL’s training and evaluation.

Research Question: How does integrating human-in-the-loop feedback and evaluation into the RL training pipeline impact the reliability, explainability, and user trust of knowledge agents in enterprise search?

Hypothesis: Incorporating live or asynchronously collected human feedback and corrections during RL training (as in FlowXpert and human preference learning) will improve answer quality, explainability, and trustworthiness, especially on ambiguous or hard-to-verify queries.

Experiment Plan: - Deploy KARL-powered agents in a controlled enterprise search setting with real users providing feedback, correction, and uncertainty flags.

• Incorporate this feedback into reward shaping, policy updates, and answer explanation modules.
• Compare performance, user satisfaction, and explainability metrics with standard RL-trained agents, including on new, ambiguous, or high-stakes queries.
• Analyze which types of feedback (e.g., corrections, uncertainty, explanations) most improve agent performance and user trust.

References:

• 1. Chang, J. D., et al. (2026). KARL: Knowledge Agents via Reinforcement Learning.
• 2. Shi, B., Luo, Y., Wang, J., Zhao, Y., Zhang, S., Hao, B., Zhao, C., Sun, Y., Zhang, Z., Sun, R., Li, H., Chen, X., Miao, J., Pei, D. (2025). FlowXpert: Expertizing Troubleshooting Workflow Orchestration with Knowledge Base and Multi-Agent Coevolution. Knowledge Discovery and Data Mining.