Social Continual Learning: Empowering XSkill with Interactive Question-Asking for Causal Skill Acquisition

TL;DR: What if XSkill agents could ask questions like a curious child when they get stuck, using language to fill gaps in their knowledge and learn new skills more efficiently? By integrating dialogue-driven exploration and natural language oracles, XSkill could incrementally build causal models of its environment.

Research Question: Can augmenting XSkill with the ability to interactively ask questions and receive natural language feedback accelerate skill acquisition and causal reasoning in dynamic, partially observable environments?

Hypothesis: XSkill agents equipped with proactive question-asking capabilities and access to a natural language oracle will exhibit faster skill acquisition, better causal reasoning, and greater resilience to sparse or ambiguous feedback.

Experiment Plan: - Setup: Extend XSkill’s inference loop to include the SCOOP framework’s question-generation and causal reasoning modules, enabling the agent to query an oracle (simulated or human) when uncertainty is detected.

• Environments: Use partially observable, causally complex benchmarks requiring tool use and sequential reasoning.
• Comparisons: Compare standard XSkill, XSkill+QA, and baseline interactive agents without structured skill/experience streams.
• Measurements: Assess learning curves, causal inference accuracy, and efficiency of knowledge acquisition (e.g., fewer steps to mastery).
• Expected Outcome: The question-asking XSkill variant should learn more efficiently in complex, ambiguous settings, demonstrating the power of social continual learning.

References:

• Jiang, G., Su, Z., Qu, X., & Fung, Y. R. (2026). XSkill: Continual Learning from Experience and Skills in Multimodal Agents.
• Ognibene, D., Patania, S., Annese, L., Koyuturk, C., Garzotto, F., Vizzari, G., Ruggeri, A., & Colombani, S. (2025). SCOOP: A Framework for Proactive Collaboration and Social Continual Learning through Natural Language Interaction and Causal Reasoning. arXiv.org.