Explainable Skill Chunking: Interpretable Action Abstractions for Transparent Long-Horizon LLM Agent Reasoning

TL;DR: Imagine if your agent could not only solve complex tasks but also explain its steps in a way humans (and other agents) can understand—kind of like showing its work on a math test. By integrating explainable skill chunking (SkillTree, Wen et al., 2024) with KLong, we enable the agent to learn, execute, and verbalize discrete, meaningful skill abstractions as part of its long-horizon reasoning.

Research Question: How does embedding interpretable skill-based abstractions within trajectory-splitting and RL fine-tuning affect the transparency and transferability of LLM agents solving long-horizon tasks?

Hypothesis: Agents trained to represent and verbalize their plans as discrete, explainable skill chunks will not only match performance of black-box models but will be easier to debug, adapt, and transfer across domains.

Experiment Plan: - Augment trajectory-splitting SFT with annotations that map sub-trajectories to human-interpretable skill labels.

• During RL, incorporate an auxiliary loss that encourages skill chunking and verbalization, leveraging differentiable decision trees as in SkillTree.
• Evaluate performance, interpretability (measured by human raters), and transfer to new but structurally similar tasks.

References:

• Wen, Y., Li, S., Zuo, R., Yuan, L., Mao, H., & Liu, P. (2024). SkillTree: Explainable Skill-Based Deep Reinforcement Learning for Long-Horizon Control Tasks. AAAI Conference on Artificial Intelligence.