---
title: Knowledge Dispersion Index (KDI)
url: https://www.emergentmind.com/topics/knowledge-dispersion-index-kdi
type: topic
---

# Knowledge Dispersion Index (KDI)

The Knowledge Dispersion Index (KDI) refers to a class of quantitative metrics designed to capture the spread, diversity, or reliability of knowledge within a system, model, or organization. Multiple KDI formulations exist in the research literature, each contingent on the underlying domain—LLM response evaluation, intellectual capital, scientific influence, field-mapping, or information-theoretic model comparison. All share the goal of operationalizing the concept of “dispersion” as it applies to knowledge, information flow, or citation structure.

## 1. KDI for LLM Response Dispersion

The most recent instantiation of KDI is as a black-box metric for benchmarking domain-specific knowledge in large language models (LLMs) without requiring labeled datasets or access to model internals [2408.13624]. Here, the KDI—termed “response dispersion”—quantifies the variability in an LLM’s answers to repeated prompts within a fixed domain. The precise definition is as follows:

- Run N independent generations of an LLM on a fixed opinion-style prompt within a domain (e.g., N=100, prompt asks for a favorite single-word topic in a category).
- Embed each single-word response $t_i$ into a vector in $\mathbb{R}^d$ using one of:
  - OpenAI’s text-embedding-3-large (d≈1536)
  - Reference-Sentence-Similarity (RSS) embeddings (d=N, based on pairwise string similarity).
- Stack responses into an $E\in\mathbb{R}^{N\times d}$ embedding matrix.
- Compute the singular value decomposition $E = U\Sigma V^\top$, with singular values $\sigma_1\ge\sigma_2\ge\dots$.
- Let $\tau=0.95$ (95% of total variance). Define $\mathit{Var}(k)=\frac{\sum_{i=1}^{k}\sigma_i^2}{\sum_{i=1}^m\sigma_i^2}$, $m=\min(N,d)$. The KDI is $k^* = \min\{k:\mathit{Var}(k)\ge \tau\}$.

Thus, KDI specifies the number of principal directions required to account for 95% (or another chosen threshold) of the response space variability for a topic+model pair.

Empirical findings indicate an average Spearman rank correlation of approximately $-0.59$ between KDI and gold-standard QA accuracy across models and domains (lower KDI, signifying less diverse answers, corresponds to higher QA accuracy). For pairwise model selection, KDI matches or approximates accuracy-based rankings in roughly 74–89% of cases, depending on chosen tolerance for accuracy trade-off [2408.13624].

## 2. KDI in Organizational Intellectual Capital

KDI was originally introduced to quantify knowledge flow and intellectual capital at the micro- (organizational) and macro- (economic sector) levels [1106.2601]. The construction is based on aggregating normalized indicators:

- **Micro-KDI:** Weighted sum of up to 23 human- and function-oriented organizational metrics (e.g., patents pending, % R&D in basic research, IT capacity). Raw values $M_i$ are normalized $m_i=(M_i-\min_i)/(\max_i-\min_i)$, then KDI$_{micro}=\sum_{i=1}^{23} w_i m_i$ with weights $w_i\ge 0$ s.t. $\sum w_i=1$.
- **Macro-KDI:** Aggregates flow metrics for industrial (inter-firm technical knowledge) and consumer (public broadcast and media dissemination) sectors. Let $I_{industrial},I_{consumer}$ be suitable flow indices, then KDI$_{macro} = \beta I_{industrial} + (1-\beta) I_{consumer}$.
- **Total KDI:** $KDI_{total} = \alpha KDI_{micro} + (1-\alpha) KDI_{macro}$, with analyst-chosen weights $\alpha,\beta\in[0,1]$.

Theoretical extension involves mapping knowledge flow as a network graph with nodes, capacity constraints, and self-correcting equilibria (modeled analogously to max-flow/min-cut). Robustness is assessed by studying network “super-families” and propagation under perturbations [1106.2601].

## 3. KDI via Influence Dispersion in Citation Networks

Another formalization of KDI is tied to the topological structure of a scientific paper’s influence in citation networks, notably through the Influence Dispersion Tree (IDT) model [1904.07579]. The construction is as follows:

- Build the IDT for a focal paper $P$ with $n$ citing papers $C_P$, assigning edges so as to maximize tree depth (parent selection favors farthest ancestor).
- Compute the Influence Dispersion Index (IDI): $IDI(P) = \sum_{\ell\in\text{Leaves}(T_P)} \text{dist}_T(P,\ell)$.
- The Normalized Influence Divergence (NID) is defined as $NID(P) = \frac{IDI(P)-n}{IDI^{max}(n)-n}$, where $IDI^{max}(n)$ is the maximal dispersion tree given $n$.
- The Knowledge Dispersion Index is then $KDI(P) = 1 - NID(P)$, so $KDI\in[0,1]$ with higher values denoting balanced dispersion (depth $\approx$ breadth $\approx \sqrt{n}$).

Empirical analysis on large bibliometric corpora shows that $KDI$ is more predictive than simple citation counts in forecasting future impact and identifying highly influential papers [1904.07579].

## 4. KDI as Diversity-Coherence Composite in Knowledge Mapping

KDI is also used as a composite metric integrating cognitive diversity and coherence, especially in knowledge integration and diffusion studies [1412.6683]. The method proceeds as:

- **Diversity** (Rao–Stirling): $D_{RS} = \sum_{i=1}^N\sum_{j=1}^N p_i p_j d_{ij}$, where $p_i$ is categorical share and $d_{ij}$ the cognitive distance (often $1$ minus cosine similarity).
- **Coherence:** $C_{RS} = \sum_{i=1}^N\sum_{j=1}^N \ell_{ij} d_{ij}$, where $\ell_{ij}$ indicates relational intensity (e.g., citation, co-occurrence).
- **Composite KDI:** $KDI = w_D(D_{RS}/D_{RS}^{max}) + w_C(C_{RS}/C_{RS}^{max})$, with normalized terms and $w_D+w_C=1$.

The approach is sensitive to the categorization scheme and the definition of cognitive distance. Visualization overlays map diversity and coherence values onto “basemaps” of science or technology [1412.6683].

## 5. KDI as Dispersion Index in Information Theory

In information theory, KDI describes the variance of the pointwise Kullback–Leibler (KL) divergence between two probability distributions $f$ and $g$ [2106.12292]. The definition is:

$$
\mathrm{VarK}(f\!:\!g) = \int f(x) \log^2 \frac{f(x)}{g(x)}\, dx - \left[\int f(x) \log \frac{f(x)}{g(x)}\, dx \right]^2.
$$

This quantity measures the spread (reliability) of the instantaneous KL divergence and is used to supplement the mean divergence in model selection via a mean–variance trading rule. KDI is always nonnegative, vanishes iff $f=g$ almost surely, and is not a metric. It is particularly recommended when models have similar average divergence but differ in variance of fit [2106.12292].

## 6. Cross-Contextual Properties and Implementation Considerations

Despite domain diversity, KDIs share systematic features:

- **Normalization and Intelligibility:** KDIs are either directly interpretable as (dimension-count, network) indices or are normalized to $[0,1]$ for comparison.
- **Embedding and Representation:** Construction may depend on data embedding (e.g., response embeddings for LLMs, vectorized categories for knowledge mapping, trees for citation analysis).
- **Robustness:** KDI metrics must be stress-tested against category choice, sample size, and parametrization (e.g., choice of thresholds $\tau$ or normalization bases).
- **Empirical Utility:** Where validated, KDIs serve as surrogates or complements to more direct but expensive measures (e.g., QA accuracy, human-judged impact, manual diversity analyses).
- **Limiting Factors:** Predictive power and robustness are context sensitive—e.g., KDI may underperform in domains with inherently high answer variability or ill-defined conceptual distances.

## 7. Summary Table: Major KDI Variants

| Context                      | Definition/Formula                                        | Application Domain           |
|------------------------------|----------------------------------------------------------|------------------------------|
| LLM Response Dispersion      | $\min\{k: \mathit{Var}(k)\ge \tau\}$ in SVD of response embeddings | LLM benchmarking [2408.13624]|
| Organizational Knowledge     | Weighted sum of normalized micro/macro indicators        | Intellectual capital [1106.2601] |
| Citation Influence           | $1 -$ Normalized Influence Divergence (NID)              | Scholarly impact [1904.07579]|
| Diversity-Coherence Mapping  | $w_D D_{RS}/D_{RS}^{max} + w_C C_{RS}/C_{RS}^{max}$      | Science mapping [1412.6683]  |
| Information-Theoretic        | $\mathrm{VarK}(f\!:\!g)$ variance of $\log(f/g)$         | Model selection [2106.12292] |

Each KDI variant is anchored in a precise mathematical formulation and an explicit workflow, as detailed in the cited works. Adherence to these domain-specific definitions is necessary for valid computation, interpretation, and comparative analysis.

Source: https://www.emergentmind.com/topics/knowledge-dispersion-index-kdi