When an intelligent system explains its reasoning, plans, or instructions, it communicates across a hidden chasm: the user's prior knowledge. If the system provides a highly detailed, step-by-step breakdown, a novice might find it essential, but an expert will experience a "preoccupation cost" [Paleja et al., 2021], finding the redundant information patronizing or actively detrimental to task efficiency [Peng et al., 2021; Gu et al., 2023]. Conversely, if the system outputs a high-level conceptual summary, an expert can seamlessly integrate it, but a novice will face cognitive friction or complete task failure due to missing prerequisite knowledge [Köhn et al., 2020; Doran et al., 2017].
This fundamental trade-off has driven a long-standing research direction: user-adaptive explanation. The goal is to dynamically calibrate the complexity, format, and granularity of an explanation to match the recipient's specific background. While early intelligent agents struggled to map linguistic variations to a user's internal state [Jacobs, 1983; Rissland et al., 1984], modern systems leverage structured planning and alignment techniques to prevent "interpretation collapse"—the tendency of general-purpose models to default to a generic, one-size-fits-all description that satisfies neither beginners nor specialists [Joshi et al., 2025; Wang et al., 2025].
Modeling Expertise as Abstraction Hierarchies
To adapt an explanation, a system must first formalize what "expertise" means. Rather than treating it as a vague categorical label, researchers model user knowledge as a specific level of abstraction or fidelity within a domain's constraint network [Sreedharan et al., 2018].
An expert's mental model operates with highly compressed conceptual units; they understand the underlying sub-steps implicitly. A novice's model requires these units to be explicitly unpacked. To bridge this gap, adaptive systems maintain a dynamic user model to track what has already been communicated or assumed [Vilnat et al., 1985; Hipp et al., 1993].
Expert Model: [Execute Task] -> (High-Level Action)
│
Novice Model: [Execute Task] -> [Subgoal A] ──> [Subgoal B]
│ │
[Step a1 -> a2] [Step b1 -> b2]
In plan-based instruction generation, user competence acts as a threshold for plan decomposition [Ardissono et al., 1993]. For example, using Hierarchical Task Networks (HTNs), a system can evaluate a user's profile before generating text: if the user is flagged as understanding a complex sub-component, the planner selects a high-level communicative action; otherwise, it recursively decomposes the component into atomic, low-level steps [Köhn et al., 2020]. This prevents information overload while ensuring sufficient grounding for beginners.
Structural vs. Process-Oriented Schemas
Adapting an explanation is not merely a matter of adjusting text length or truncating sentences; it fundamentally alters what information is prioritized. Foundational work in natural language generation established that technical communication strategies must pivot between entirely different organizational schemas based on user knowledge:
- Experts prefer structural descriptions: They require constituency-based schemas that focus on the hierarchical arrangement of components, parts, and properties [Paris, 1985; Lebowitz, 1985]. They look for specific, convergent data and constraints to refine their execution [Recipe 2.0, 2023].
- Novices require process traces: They benefit from functional or temporal schemas detailing the causal chain of how a system operates [Paris, 1985; Lebowitz, 1985]. Novices rely on generalizable rules of thumb to navigate complex decision spaces without becoming overwhelmed.
This divergence extends directly to the placement and complexity of illustrative examples. Mittal et al. demonstrated that beginners require simple, single-feature examples placed immediately after the concept is introduced to anchor their understanding. Experts, conversely, find simple examples insufficient and prefer complex, multi-feature instances grouped at the end of a description as a reference manual.
Pruning and Textual Recursion
To implement these shifts dynamically, systems use text-pruning mechanisms like "conceptual sketchification" [Cullingford et al., 1981] and answer expression filtering [Chin, 1988]. By mapping user expertise to a taxonomy of fact difficulty, a system can omit redundant information that qualifies as common knowledge for a specialist, drastically shortening the output while maintaining clarity.
This is achieved through textual recursion. As argued by McKeown et al., discourse units can be expanded conditionally. When a system generates an explanation, it monitors for concepts that fall outside the user's estimated knowledge profile. If an unfamiliar concept is detected, the system triggers a recursive sub-explanation schema to expand that single descriptive predicate into a complete definition, leaving it implicit for more advanced readers [Holland-Minkley et al., 1999].
Progressive Disclosure in Modern AI
In modern complex domains—such as cybersecurity operations [Stock et al., 2018; Jalalvand, 2025] or clinical search engines [Contempré et al., 2021]—this adaptive mechanism is deployed via progressive disclosure.
For instance, in a Security Operations Center (SOC), generic model-internal explanations often overwhelm analysts during high-pressure triages. Frameworks like CORTEX [Pant et al., 2025] and ICoT [Zeng et al., 2025] modulate their technical depth based on the user's operational role. They provide high-level, natural language "attack narratives" for general users or quick triage, while allowing deep-dives into raw feature attributions, saliency maps, or protocol-level code patches for forensic experts [Wang et al., 2024].
[User Profile Detected]
│
├─► Novice/Triage ──► High-Level Attack Narrative (Causal Flow)
│
└─► Forensic Expert ─► Saliency Maps & Protocol-Level Code Patches
Rather than relying purely on hand-crafted rules or static templates, contemporary architectures frame explanation generation as a cooperative game optimized via Direct Preference Optimization (DPO) [Wang, 2025]. By mapping a core set of semantic claims to different levels of technical vocabulary, the system learns to select the precise subset of information that minimizes a specific listener's uncertainty. The explanation works because it treats communication not as a broadcast of raw data, but as a targeted update to an individual's mental model.