Disentangled Skill Representations for Predictive Human Modeling
Introduces SAIL (Skill Abstraction with Interpretable Latents), a method for modeling human skill as an interpretable, multi-dimensional construct inferred from naturalistic behavior over time Each individual is represented by a persistent skill embedding that controls a blend between expert and novice bases, trained using counterfactual subskill swaps for disentanglement The approach produces skill embeddings robust to transient performance fluctuations and learns transferable representations o
Analysis
TL;DR
- Introduces SAIL (Skill Abstraction with Interpretable Latents), a method for modeling human skill as an interpretable, multi-dimensional construct inferred from naturalistic behavior over time
- Each individual is represented by a persistent skill embedding that controls a blend between expert and novice bases, trained using counterfactual subskill swaps for disentanglement
- The approach produces skill embeddings robust to transient performance fluctuations and learns transferable representations of human subskills
- Demonstrates strong predictive performance across racing and baseball domains, consistently outperforming baselines in behaviorally grounded disentanglement
- Improves downstream AI coaching performance, showing practical utility for assistive and collaborative AI systems
Why It Matters
This work addresses a critical gap in AI systems designed to collaborate with, coach, or assist humans: the ability to accurately model and predict human skill. By treating skill as a persistent, compositional construct rather than a single-observation latent variable, SAIL enables more reliable and interpretable human modeling—essential for applications ranging from sports coaching to adaptive tutoring systems.
Technical Details
- Core Architecture: SAIL represents each individual with a persistent skill embedding that interpolates between expert and novice base representations, creating a structured latent space for skill
- Disentanglement Strategy: Counterfactual subskill swaps are used during training to encourage disentanglement of subskills, ensuring that learned representations are behaviorally grounded rather than confounded by transient performance factors
- Temporal Inference: Unlike typical latent variable estimation from single observations, skill is inferred from behavioral patterns over time, making the representation robust to noise and performance fluctuations
- Evaluation Domains: Validated across two distinct domains—racing and baseball—demonstrating generalization across different types of human skill expression
- Downstream Application: The skill-informed representations improve behavior prediction across in-domain contexts and enhance AI coaching performance
Industry Insight
- AI coaching and tutoring systems can leverage disentangled skill representations to provide more personalized, adaptive feedback by understanding the specific subskills a learner needs to develop
- The counterfactual swap training approach offers a generalizable technique for disentanglement in any domain where persistent latent traits must be inferred from sequential behavioral data
- As AI assistants become more embedded in collaborative human workflows, methods that produce interpretable skill models will be critical for building trust and enabling effective human-AI partnership
Disclaimer: The above content is generated by AI and is for reference only.