Abstract: LLM agents often fail in specialized domains requiring long-tail knowledge. We introduce AHCE (Active Human-Augmented Challenge Engagement), where the agent learns when and how to treat a human expert as an interactive reasoning tool rather than just a source of answers.
Key Insight: Successfully augmenting agents requires learning a policy for requesting expert reasoning. This moves beyond simple help requests toward "Active Human-AI Collaboration" where the agent manages its own uncertainty by querying a higher-level intelligence.
Relevance to RSI: Defines the interface for "Human-Assisted RSI," where the agent iteratively refines its own knowledge by strategic interaction with human supervisors, crucial for bootstrapping in specialized or safety-critical domains.