Abstract
This paper addresses the gap between current AI pattern matching and human-like "eductive ability"—the capacity to solve novel problems through abstract reasoning. By proposing a research roadmap centered on Raven’s Progressive Matrices (RPM), the author advocates for integrating computational mental imagery and a dynamic interplay between perceptual and conceptual processing. The goal is to move AI beyond trivial interpolation toward nontrivial generalization, where abstract rules are successfully applied to entirely new visual attributes. This approach seeks to transition AI from "flat" generative models to complex production systems capable of iterative, multi-step visual reasoning.
Introduction: The Eductive Gap in Modern AI
Despite the "super efficiency" of modern AI, current models remain largely tethered to their training sets. While they can generate high-quality content, they rarely invent it in a way that demonstrates fluid intelligence. Fluid intelligence is defined as the ability to perceive complex relationships and reason abstractly in unfamiliar environments without relying on stored memory.
The core problem is that AI excels at interpolation—finding patterns within the distribution of its training data—but struggles with eductive ability. To bridge this gap, researchers must move toward architectures that incorporate mental imagery and a dynamic feedback loop between seeing (perceptual processing) and understanding (conceptual processing).
Raven’s Progressive Matrices and Fluid Intelligence
The primary benchmark for measuring this ability is Raven’s Progressive Matrices (RPM), a classic non-verbal intelligence test. Solving an RPM requires identifying abstract rules governing rows and columns of geometric shapes.
While some AI models can solve specific RPM datasets, they often fail at nontrivial generalization. For instance, a model might learn the concept of "progression" when applied to colors but fail to apply that same concept to shapes or sizes if those specific combinations were absent from the training set. True intelligence requires a "ladder of puzzles" where each step necessitates learning a new, higher-level pattern rather than just memorizing feature-rule pairs.
The Challenge of Nontrivial Generalization
The author highlights that true abstraction means giving "new meanings to learned perceptual elements." If a model cannot decouple a concept (e.g., "alternating") from a specific feature (e.g., "color"), it has not truly learned the abstraction.
As illustrated in the figure above, nontrivial generalization involves training on one set of concept-attribute pairs (like Progression + Color) and testing on unseen combinations (like Alternating + Size). Current benchmarks, such as those by Barrett et al. (2018), often focus on trivial generalization where models still struggle, indicating a fundamental flaw in how AI handles visual concepts.
A Roadmap for Imagery-Based Reasoning
To address these limitations, the paper proposes a multi-pronged research program:
- Psychometric Grounding: Reviewing human problem-solving strategies to identify the "learning potential" that AI currently lacks.
- Perceptual-Conceptual Interplay: Moving away from static feed-forward models toward dynamic systems where the identification of an abstract rule informs how the stimuli are perceived, and vice versa.
- Computational Mental Imagery: Utilizing "imagistic representations" that can be mentally manipulated. This allows an AI to "incarnate" abstract concepts onto visual stimuli to test hypotheses, providing a flexible workspace for "what-if" reasoning.
- Imagery-Based Production Systems: Shifting from "flat" generative models to complex systems that mimic human problem-solving by inspecting rows, generating intermediate results, making analogies, and iteratively backtracking when a hypothesis fails.
Conclusion and Future Work
The most significant limitation of current AI is that generative models are too "flat." They lack the hierarchical structure and multi-step reasoning ability required to orchestrate complex problem-solving. Future work focuses on extending generative models into full "production systems" that can inspect, apply, generate, and analogize in a loop. By treating imagery as a workspace rather than just an output, AI can begin to replicate the biological basis for higher cognitive abilities and achieve true visual abstract reasoning.
Summary updated