How Individuals Efficiently Search for Creative Concepts: A Value-Guided Search Account

Technical Summary: How Individuals Efficiently Search for Creative Concepts: A Value-Guided Search Account

Problem Statement
Creative idea generation requires navigating a vast space of possible concepts to identify those worth combining into novel and appropriate ideas. While associative theories explain how concepts become activated, they leave open the mechanism by which individuals efficiently search among numerous possibilities to select candidate concepts for development. The central problem addressed is how the search for candidate concepts is guided before a complete creative idea is formed. Specifically, the authors investigate whether this search can be modeled as a value-guided process regulated by an exploration–exploitation balance, where "value" is defined as a goal-dependent signal indicating a concept's potential for further development.

Methodology
The research employs a two-study design combining behavioral experiments with computational modeling and simulation-based validation.

• Study 1 (Value Construction):

  • Participants: 61 undergraduate students.
  • Task: Participants evaluated 90 pre-selected concepts (derived from Alternative Uses Task responses) under two distinct contexts: a "general use" condition and a "creative use" condition.
  • Measures: For each concept, participants provided ratings for novelty, appropriateness, and "concept value" (operationalized as willingness to develop the concept into a final response). They also made binary selection decisions.
  • Analysis: Hierarchical Bayesian modeling was used to compare competing models of concept selection and value construction. The study tested whether concept value predicts selection better than novelty or appropriateness alone and whether the weighting of these attributes changes based on task context.

• Study 2 (Search Dynamics):

  • Participants: 86 undergraduate students.
  • Task: A semantic exploration paradigm where participants navigated an artificially constructed semantic network (nodes grouped into 15 semantic clusters). Participants used arrow keys to move between neighboring concepts and a space bar to select a concept for generating a response.
  • Measures: The study tracked step-by-step navigation transitions, reaction times, and concept selection events. Post-task ratings of novelty, appropriateness, and value were collected for encountered concepts.
  • Analysis:
    • Model Comparison: Competing softmax models (value-guided, novelty-only, appropriateness-only, semantic proximity, random) were fitted to predict transition probabilities between concepts.
    • Dynamic Regulation: The study tested whether the "temperature" parameter (controlling stochasticity vs. value-guided exploitation) remains constant or shifts dynamically within "answer intervals" (the sequence of steps between two selected concepts).
    • Simulation: A semantic network was reconstructed from 1,004 real AUT responses. Five activation models (value, distance, random, jump, reset) were simulated to predict which concepts participants would generate in a standard AUT.

Key Results

1. Task-Dependent Value Construction (Study 1):

   • Concept value was the strongest predictor of concept selection (accuracy = 83%), outperforming models based solely on novelty or appropriateness.
   • Value is constructed via a nonlinear integration of novelty and appropriateness.
   • Context Modulation: In creative contexts, the weight assigned to novelty ($\alpha$) was significantly higher, while appropriateness ($\beta$) was weighted higher in general contexts. Furthermore, the integration parameter ($\tau$) was lower in creative contexts, indicating that novelty and appropriateness must be jointly satisfied (less compensatory) for a concept to be deemed valuable in a creative task.

2. Value-Guided Search Trajectories (Study 2):

   • Transitions between candidate concepts were best predicted by the value-guided softmax model (Exceedance Probability $X_p = 1.00$), significantly outperforming models based on semantic proximity or random choice.
   • Exploration–Exploitation Trade-off: Stronger value guidance (lower temperature $T$) correlated with fewer redundant revisits but longer search durations and fewer unique concepts generated, suggesting a trade-off between efficiency and exploratory breadth.

3. Dynamic Regulation of Stochasticity (Study 2):

   • Search dynamics follow a U-shaped reaction time pattern within answer intervals: faster responses early in the interval and slower responses as a concept is selected.
   • Temperature Shift: Model comparison strongly favored a dynamic-temperature model over a fixed-temperature model. Search begins with higher temperature (broader, more stochastic exploration) and shifts to lower temperature (stronger value sensitivity/convergence) as the search progresses within an interval.
   • This shift is most pronounced during category switches, where a transient increase in exploration is followed by value-sensitive convergence in the new semantic region.

4. Simulation Validation:

   • In simulations using real AUT data, the value-guided model outperformed distance-based and random models in predicting human-generated concepts, particularly at mid-range neighborhood sizes.
   • Models incorporating an early high-to-low temperature sequence (HT+LT) provided the best fit, confirming that efficient retrieval relies on early variability followed by rapid value-sensitive convergence.

Significance and Claims
The paper claims to provide a process-level account of creative concept search, moving beyond the view of creativity as merely the retrieval of complete ideas. The authors argue that:

• Value is a directional constraint: Concept value does not replace associative search but provides a directional bias, helping individuals navigate open-ended conceptual spaces by prioritizing concepts that satisfy current task goals.
• Coupled Mechanisms: Efficient creative search depends on the coupling of value guidance with dynamic exploration–exploitation regulation. Specifically, search begins with broad, stochastic exploration to avoid local fixation and transitions to value-sensitive convergence once promising concepts are identified.
• Temporal Dynamics: The study demonstrates that this regulation occurs at the moment-to-moment level of semantic transitions, not just in the final selection of an idea.
• Generalizability: The findings suggest that value-guided search operates most clearly at the local scale of candidate activation, before broader structural regularities of the semantic network dominate the search process.

The authors maintain a modest stance, noting that the study relies on post-task value ratings and a constrained semantic network, and that future work is needed to determine if evaluative weights change online during search or to generalize these principles beyond the Alternative Uses Task.