Technical Summary: Large Language Models Explain Experts Better Than Experts Themselves

1. Problem Statement

The core challenge addressed in this study is the externalization and transfer of tacit knowledge. Tacit knowledge, defined as the "know-how" embedded in experience, is notoriously difficult to articulate, making its transfer from experts to novices a significant bottleneck in organizations. This phenomenon is encapsulated in Polanyi's Paradox ("we can know more than we can tell"), where experts often cannot fully verbalize the underlying logic of their behaviors and decisions. Consequently, valuable expertise is frequently lost when experts leave, or it fails to disseminate effectively.

Traditional methods for capturing this knowledge—such as semi-structured interviews, cognitive task analysis, and mentoring—are costly, difficult to scale, and often require substantial human effort to structure. Furthermore, the knowledge experts can articulate is often an incomplete representation of the knowledge they actually use. This study investigates whether Large Language Models (LLMs) can overcome this articulation bottleneck by externalizing tacit knowledge directly from expert behaviors (conversations) rather than relying on expert self-reporting, and whether this externalized knowledge can improve downstream decision-making and novice training.

2. Methodology

The research employs a two-study experimental design within a math tutoring context, utilizing a dataset of real-world conversations between first- to fifth-grade students and tutors (Wang et al., 2024).

Data and Setup

• Dataset: 700 math tutoring conversations, including student mistakes and responses from both novice and expert tutors.
• Benchmark: A "Human Best Effort" benchmark derived from a Cognitive Task Analysis (CTA) where experts articulated a three-step decision-making model. This serves as the gold standard for expert-articulated knowledge.
• Participants: Novice tutors recruited via Prolific (screened for no prior math tutoring experience) in Study 2.

Study 1: LLM Response Generation

The objective was to test if LLM-externalized knowledge could guide an LLM to generate higher-quality responses to student mistakes.

• Knowledge Externalization Conditions: Three prompting strategies were tested to extract tacit knowledge from the dataset:
  1. LLM Baseline: No student-tutor conversations provided; the LLM extracts knowledge based on general instructions.
  2. LLM+Conversation: The LLM is provided with student-tutor conversations (observed behaviors) but no structural constraints.
  3. LLM+Both: The LLM is provided with conversations and required to output a 3-step structure mimicking the human expert benchmark.
• Comparison Conditions:
  • Human Best Effort: The LLM is prompted with the explicit, articulated knowledge from the human experts.
  • No Tacit Knowledge: A zero-shot baseline with no externalized knowledge.
• Evaluation: Responses were rated on a 1-10 scale for usefulness, caring, and human-sounding qualities.

Study 2: Knowledge Transfer to Novices

The objective was to determine if LLM-externalized knowledge could bridge the performance gap between human novices and human experts.

• Procedure: Novice tutors were randomly assigned to conditions where they were trained on the specific tacit knowledge representations generated in Study 1 (LLM Baseline, LLM+Conversation, LLM+Both, Human Best Effort, or No Tacit Knowledge).
• Task: Participants generated responses to student mistakes.
• Metric: The Expert-Novice Gap, defined as the difference in response quality between the trained novice and the human expert for the same conversation.

Robustness Checks

The study included several robustness analyses:

• Sample Size Sensitivity: Varying the number of training conversations (from 1 to 320) to verify if LLMs learn patterns rather than mimicking style.
• Alternative Knowledge: Testing domain-agnostic, out-of-context (writing), and mechanically domain-adapted knowledge to isolate the value of domain-specific substance.
• Retrieval-Augmented Generation (RAG): Replicating with the full dataset (420 conversations) using RAG instead of context-window subsampling.
• Model Replication: Re-evaluating results using different LLMs (xAI's Grok and Anthropic's Claude) for both generation and scoring.

3. Key Results

Response Generation (Study 1)

• Superiority of LLM-Externalized Knowledge: LLMs guided by tacit knowledge externalized from expert conversations (LLM+Conversation) generated responses that significantly outperformed those guided by expert-articulated knowledge (Human Best Effort).
• The Articulation Bottleneck: The "Human Best Effort" condition performed comparably to the "LLM Baseline" (no conversations), suggesting that experts struggle to articulate the full scope of their tacit knowledge.
• Structure vs. Flexibility: Imposing a rigid 3-step structure (LLM+Both) actually reduced response quality compared to the flexible LLM+Conversation approach. This indicates that forcing LLMs to mimic human articulation structures can hinder their ability to capture nuanced behavioral patterns.
• Quality Dimensions: The LLM+Conversation approach improved all three dimensions: usefulness, caring, and human-sounding.

Knowledge Transfer (Study 2)

• Bridging the Gap: Novices trained on LLM+Conversation knowledge achieved the smallest Expert-Novice Gap, significantly outperforming those trained on Human Best Effort or LLM Baseline knowledge.
• Novice Performance: Novices in the LLM+Conversation condition approached expert-level performance, particularly in the "caring" dimension, where the gap became statistically non-significant.
• Ineffectiveness of Human Articulation: Surprisingly, the Human Best Effort condition did not perform better than the LLM Baseline in transferring knowledge to novices, reinforcing the idea that human articulation is an incomplete proxy for expert capability.

Mechanism and Robustness

• Learning vs. Mimicry: A "learning curve" was observed where response quality increased with the number of training conversations, plateauing after a sufficient threshold. This confirms LLMs are extracting decision patterns rather than simply copying style.
• Domain Specificity: In-domain LLM-externalized knowledge outperformed out-of-context and mechanically adapted knowledge, primarily driven by improvements in usefulness (instructional guidance).
• Generalizability: Findings remained consistent across different LLM providers (GPT-4o, Grok, Claude) and retrieval methods (in-prompt vs. RAG).

4. Key Contributions

1. Scalable Externalization: The paper proposes and validates a scalable method for externalizing tacit knowledge using LLMs that bypasses the human articulation bottleneck.
2. Empirical Validation of Polanyi's Paradox: The study provides empirical evidence that LLMs can "know more" about an expert's decision-making process than the experts can "tell" through self-reporting.
3. Novice Training Efficacy: It demonstrates that knowledge derived from expert behaviors (via LLMs) is more effective for training novices than knowledge derived from expert articulations.
4. Condition Identification: The study identifies critical conditions for effective externalization: sufficient exposure to behavior data, flexible output structures (avoiding rigid human-like constraints), and the use of capable LLMs.

5. Significance and Claims

The authors claim that this work offers a practical pathway for organizations to leverage expert interaction data to mitigate knowledge loss and improve training. By shifting the focus from "asking experts what they know" to "analyzing what experts do," LLMs can serve as powerful tools for knowledge management.

The paper modestly concludes that while LLMs do not fully close the expert-novice gap, they significantly narrow it, often surpassing traditional methods of knowledge transfer. The findings suggest that the "know-how" embedded in expert interactions is a richer resource for training and decision support than the "know-what" that experts can explicitly document. This highlights the potential of LLMs not just as generative tools, but as analytical engines for organizational learning that can capture and transfer the implicit logic of expertise.