Abstract
If scientific discovery is one of the main driving forces of human progress, insight is the fuel for the engine, which has long attracted behavior-level research to understand and model its underlying cognitive process. However, current tasks that abstract scientific discovery mostly focus on the emergence of insight, ignoring the special role played by domain knowledge. In this concept paper, we view scientific discovery as an interplay between thinking out of the box that actively seeks insightful solutions and thinking inside the box that generalizes on conceptual domain knowledge to keep correct. Accordingly, we propose Mindle, a semantic searching game that triggers scientific-discovery-like thinking spontaneously, as infrastructure for exploring scientific discovery on a large scale. On this basis, the meta-strategies for insights and the usage of concepts can be investigated reciprocally. In the pilot studies, several interesting observations inspire elaborated hypotheses on meta-strategies, context, and individual diversity for further investigations.
Introduction: The Gap in Discovery Research
How do scientists arrive at novel ideas? Traditional psychology often isolates the "Aha!" or eureka moment, studying it through puzzles like the nine-dot problem. However, real-world scientific discovery is rarely an isolated spark; it is deeply embedded in domain knowledge. A paradigm shift requires a deep understanding of the existing paradigm. Current cognitive research often lacks the semantic richness and "path-finding" nature of actual science, where researchers navigate vast conceptual webs and jump to new semantic subspaces when faced with dead ends.
Mindle: A Sandbox for Scientific Thinking
To bridge this gap, the researchers introduced Mindle, a semantic searching game designed to mimic the cognitive demands of scientific discovery.
- Gameplay: Players start with a word and attempt to find a hidden "secret word." Each guess receives a similarity score (0-100) based on cosine similarity in a high-dimensional word embedding space (e.g., Word2Vec) from a 40,000-word vocabulary.
- Modeling: The game is framed as a Markov Decision Process (MDP). By analyzing player trajectories—the sequence of guesses—researchers can identify "Aha!" moments, defined as sudden, significant jumps in similarity scores following periods of stagnation.
- Knowledge Representation: Domain knowledge is modeled as a graph of concepts where edges represent semantic relationships derived from natural language corpora.
The Reciprocation of Insights and Concepts
The core finding of the Mindle pilot study is the Reciprocation Loop. Discovery is not a linear path but a cycle of "inside the box" local optimization and "outside the box" global jumps.
As illustrated in the figure above, a player might start with a specific conceptual framework (e.g., seeing circles in a Kanizsa illusion). When this framework fails to yield progress, an "outside the box" insight occurs (realizing there is a triangle). This insight then establishes a new "box"—a new conceptual subspace—where the player returns to "inside the box" thinking to optimize their solution.
The authors distinguish between two types of insights:
- Action-Level Insights: A meta-strategy where a player switches to global "random" jumping to escape a local minimum, regardless of word meaning.
- Semantic-Level Insights: A conceptual shift where a player realizes they are in the wrong category (e.g., shifting from "education" to "technology") and intentionally moves to a new semantic subspace.
This balance is hypothesized to be governed by uncertainty: high uncertainty about what to do leads to action-level jumps, while high uncertainty about the context leads to semantic-level reconstruction.
Results and Future Outlook
The pilot study involving 25 participants demonstrated that Mindle is an engaging tool for "citizen science," with 24 out of 25 participants finding the game interesting. Researchers successfully identified "eureka" moments where global random jumps landed on "hot solutions," which were then immediately exploited through local optimization.
Future work aims to use Inverse Reinforcement Learning to reverse-engineer how individuals organize conceptual knowledge. By scaling Mindle to larger populations, the researchers hope to determine if these meta-strategies for discovery are universal or vary significantly across individuals and contexts.
Summary updated