💻 computer science

Is it out-of-the-box? A geometric approach to measuring creative thinking in semantic space

This paper proposes a computational framework that formalizes the metaphor of "thinking outside the box" by defining conceptual boundaries as geometric regions in a latent semantic space, thereby enabling the measurement of creative deviation through a topic-normalized geometric distance metric.

Original authors: Anna Kruzenshtern, Viktor Dodonov, Leonid Chechurin

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anna Kruzenshtern, Viktor Dodonov, Leonid Chechurin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to measure how "out-of-the-box" a person's thinking is. Usually, we say someone is thinking outside the box when they come up with a wild, new idea. But here's the problem: nobody actually knows what the "box" looks like. Is it a small cardboard cube? A giant warehouse? Without a clear definition of the box, it's hard to measure how far someone has stepped outside of it.

This paper proposes a way to draw that box using math and geometry, specifically for written ideas (like scientific abstracts). Here is how they did it, explained simply:

1. Turning Words into a Map

First, the researchers needed a way to turn words into something they could measure. They used a technique called Latent Semantic Analysis (LSA).

The Analogy: Imagine a giant, invisible map where every word is a city. Words that mean similar things (like "engine" and "motor") are cities located very close together. Words that are unrelated (like "engine" and "banana") are cities far apart.
By doing this, every sentence or document becomes a collection of points on this map.

2. Drawing the "Box"

Next, they needed to define the "box" for a specific topic (like "Neuroscience" or "Engineering").

The Analogy: Imagine you take all the words related to a specific topic and drop them onto your map. You then stretch a rubber band around the outermost points to create a shape. In this paper, they use a simple, straight-edged shape (like a rectangular box) to enclose all the familiar words for that topic.
This box represents the established boundaries of what is considered "normal" or "conventional" for that subject.

3. Measuring the "Out-of-the-Box" Moment

Now, they take a new document (a new idea) and see where it lands on the map.

The Measurement: They calculate the size of the "box" formed by the new document's words.
The Comparison: They compare the size of the new document's box against the size of the "conventional" topic box.
The Result: If the new document's words stretch far beyond the edges of the conventional box, it gets a high score. This score is called Geometric Deviation. It doesn't mean the idea is "good" or "useful," but it does mean the idea covers a wider, more unusual area of the map than usual.

4. Checking Their Work

To make sure their "box" method actually works, they compared it to a different method used by computers (called Neural Surprisal).

The Analogy: Imagine a computer that has read millions of old papers. If you show it a new paper, the computer gets "surprised" if it sees words or phrases it has never seen before.
The Finding: They found that when their "box" method said an idea was "out-of-the-box," the computer's "surprise" meter also went up. This suggests their geometric method is a valid way to measure novelty, even though it looks at the shape of the idea rather than just how surprising the words are.

What This Means (and What It Doesn't)

The authors are clear about what their tool does:

It does: Provide a clear, visual way to see how much a document expands beyond the usual boundaries of a topic. It's like a ruler for "conceptual stretching."
It does not: Tell you if the idea is good, useful, or true. A nonsense sentence that uses weird words might get a high score because it's far from the "box," but that doesn't make it a brilliant invention.
Where it fits: The paper suggests this could be useful in inventive design education. Teachers could use it to track if students are gradually learning to stretch their thinking further over time, moving from small, safe ideas to bigger, more exploratory ones.

In short, the paper gives us a way to draw the "box" of normal thinking and measure exactly how far a new idea steps outside of it, using geometry instead of guesswork.

Technical Summary: Is it out-of-the-box? A geometric approach to measuring creative thinking in semantic space

Problem Statement
The metaphor of "thinking outside the box" is ubiquitous in describing creative reasoning, yet the "box" itself lacks an operational definition in computational frameworks. Without a measurable representation of conceptual boundaries, it is difficult to characterize how specific ideas deviate from established domains. Existing computational approaches to novelty assessment in scientific text rely primarily on probabilistic modeling (e.g., language model surprisal) or embedding-based similarity. While these methods capture unpredictability or semantic distance, they do not provide an explicit, interpretable notion of the extent of a document's conceptual space or how far it departs from historical conceptual boundaries.

Methodology
The authors propose a computational framework that formalizes the "box" as a geometric region within a shared latent semantic space. The methodology proceeds through the following steps:

Semantic Space Construction:
- The study employs Latent Semantic Analysis (LSA) rather than contextual transformer embeddings to ensure a stable, shared coordinate system across all documents. This avoids the variability introduced by context-dependent embeddings.
- A TF-IDF matrix is constructed from the corpus, followed by Truncated Singular Value Decomposition (SVD) to generate a low-rank latent space.
- Documents are represented as finite sets of points (term embeddings) within this fixed $k$ -dimensional space.
Geometric Box Modeling:
- Document Box: For a given document $d$ , the set of unique term embeddings $E_d$ defines a geometric region. The extent of this region is quantified by an axis-aligned bounding box.
- Volume Measurement: The size of the conceptual region is calculated as the log-volume ( $\log BLV_{doc}$ ), defined as the sum of the logarithms of the axis-aligned widths ( $\Delta_j$ ) plus a small constant $\epsilon$ for numerical stability. This metric captures the semantic dispersion of the document without normalizing for length, as text length is not assumed independent of conceptual scope.
Reference "Box" Construction:
- To measure deviation, a reference "box" representing established boundaries is constructed for specific topics.
- The corpus is split into a training set (2015–2017) and an evaluation set (2018–2020).
- Latent Dirichlet Allocation (LDA) is trained on the training set to identify topics. For each topic, a vocabulary is selected based on relevance scores, and a topic-level box is constructed using the embeddings of these terms.
Metric Definition:
- The core metric is the Topic-Normalized Geometric Deviation ( $\Delta BLV$ ):
  $\Delta BLV(d) = \log BLV_{doc}(d) - \log BLV_{topic}(DT(d))$
- This value estimates how far a document extends beyond the historically observed conceptual boundaries of its dominant topic. Positive values indicate geometric expansion; negative values indicate confinement.
Validation:
- The proposed geometric deviation is compared against a surprisal-based novelty measure derived from the negative log-likelihood (NLL) of a GPT-2 language model fine-tuned on the same domain data.
- Correlations are computed between $\Delta BLV$ and token-level losses (specifically the top- $N$ highest-loss tokens) across different numbers of LDA topics ( $K$ ).

Key Results

Correlation with Surprisal: The study finds a consistent positive relationship between the geometric deviation ( $\Delta BLV$ ) and the surprisal-based novelty measure (NLL) across both Engineering (ENGI) and Neuroscience (NEUR) domains.
Statistical Significance: All individual correlations observed across varying numbers of topic boxes ( $K$ ) and token thresholds ( $N$ ) are statistically significant ( $p < 10^{-3}$ ).
Distinct Signals: While correlated, the results suggest that geometric deviation and probabilistic surprisal capture related but distinct aspects of novelty. The geometric approach measures the structural extent of a document's conceptual region, whereas surprisal measures deviation from learned token distributions.

Significance and Claims
The paper does not claim to provide a definitive or complete measure of creativity. Instead, it positions the framework as a tool for providing an interpretable structural signal of conceptual expansion.

Interpretability: Unlike "black box" probabilistic scores, the geometric approach offers a transparent visualization of how a document's semantic content spans a broader region of the latent space compared to historical norms.
Educational Application: The authors suggest the framework is particularly useful in inventive design education. It could support tracking the development of exploratory thinking over time by identifying when students move beyond familiar solution patterns (confinement within the box) to broader conceptual spaces (expansion beyond the box).
Limitations: The authors explicitly acknowledge limitations, including the potential underestimation of genuinely new terms not present in the training vocabulary, the simplification of semantic geometry via axis-aligned boxes, and the inability to distinguish between meaningful expansion and incoherent semantic dispersion.

In summary, the work contributes a complementary perspective to computational creativity by modeling novelty as geometric expansion in a shared semantic space, offering a tractable method to quantify how far a text departs from established conceptual boundaries.