Expertise Aware Explanation Adaptation

This direction is merged into another direction: user-adaptive explanation. Its last version stays readable and citable here.

Version 1 of 152 papers60 conceptsfrozen 2026-10-04

Most explainable AI (XAI) systems treat transparency as a static property: a model produces a heatmap, a decision tree, or a text summary, and the user is expected to interpret it. However, research spanning five decades suggests that effective explanation is not a data dump, but a communicative act that must adapt to the recipient’s specific knowledge. A medical diagnosis explained to a surgeon requires technical precision and structural detail, while the same diagnosis explained to a patient requires functional analogies and the pruning of jargon.

The core intuition behind Expertise-Aware Explanation Adaptation is that the "optimal" explanation is the minimal set of information required to align a user’s mental model with the AI’s internal logic.

The Problem of Information Overload

The "one-size-fits-all" approach to XAI often leads to a usability-usefulness paradox. As Pradyumna Tambwekar and Matthew C. Gombolay have observed, users often prefer natural language for its ease of use, yet they are objectively better at predicting AI behavior when using structured logic like decision trees.

When a system provides too much detail to a novice, it causes cognitive overload; when it provides too little to an expert, it loses trust. Daniel Omeiza and colleagues found a "transparency paradox" in autonomous vehicles: highly specific explanations can actually increase passenger anxiety if they reveal every minor perceptual flicker, even if the vehicle is driving safely. The goal, therefore, is selective transparency.

Mechanisms of Adaptation

Researchers have developed several key mechanisms to bridge the gap between machine logic and human understanding:

1. Conceptual Sketchification and Pruning

Early systems like CADHELP, introduced by Richard E. Cullingford and his team, pioneered "conceptual sketchification." This process dynamically prunes explanations based on what the user already knows. Similarly, David N. Chin’s UCExpress modeled user expertise to omit "common knowledge" for experts while adding instructional similes for novices. By treating explanation as a process of "answer expression," these systems ensure brevity for the experienced and depth for the learner.

2. Hierarchical Model Reconciliation

In robotics, the challenge is often explaining why a specific plan was chosen over a human-suggested alternative (a "foil"). Sarath Sreedharan and Siddharth Srivastava proposed the HELM framework, which represents the user’s understanding as an abstraction of the robot’s complex model. Explanation then becomes a process of "model reconciliation": the robot only reveals the specific low-level constraints that make the user's alternative impossible, effectively "concretizing" the model only where necessary.

3. Progressive Disclosure

To manage cognitive load, modern frameworks like LEx by Ronal Singh and Tim Miller 0001 advocate for hierarchical layers of explanation. This is often implemented via progressive disclosure, where a system provides a high-level summary initially and allows the user to interactively "drill down" into technical details. Deepa Muralidhar argues that this user-centric design allows for "selective transparency" that matches the situational context of the task.

Modern Implementations: LLMs and Optimization

The advent of Large Language Models (LLMs) has provided new tools for this adaptation. Frameworks like LLMCheckup, developed by Qianli Wang and colleagues, allow users to "chat" with interpretability tools, adjusting the technical depth of the output based on self-reported expertise.

Beyond simple prompting, researchers are now using formal optimization to refine these explanations. Jacopo Teneggi and Tao Shu introduced a framework using Direct Preference Optimization (DPO) to generate concept-based explanations. By incorporating listener-specific profiles into the reward function, the system learns to select vocabulary and concepts that maximize "communicative utility" for a specific audience—ensuring that a doctor receives a different rationale than a patient for the same AI prediction.

Why It Works: The Cognitive Firewall

The ultimate goal of expertise-aware adaptation is proper trust calibration. Xuhai Xu and Haoran Zhang found that while AI explanations can lead laypeople to over-rely on incorrect AI, experienced clinicians possess a "cognitive firewall." They use explanations to verify the AI's logic against their own expertise.

By tailoring the abstraction level—switching between "structural" schemas for experts and "functional" schemas for novices, as suggested by Michael Lebowitz—adaptive systems ensure that every user receives the information they need to either confidently follow or correctly override the AI's advice.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_expertise_aware_explanation_adaptation_64115feb,
  title = {Expertise Aware Explanation Adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/expertise-aware-explanation-adaptation/rdr_64115febf7354024a620f9658195b6e0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_expertise_aware_explanation_adaptation_64115feb_v1,
  title = {Expertise Aware Explanation Adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/expertise-aware-explanation-adaptation/rdr_64115febf7354024a620f9658195b6e0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}