While artificial intelligence has achieved superhuman performance in structured environments like Go, Chess, and StarCraft, these successes represent a specific type of intelligence: crystallized intelligence. This paper argues that the next frontier for AI is fluid intelligence—the ability to solve novel, abstract problems without specific prior training or task-specific experience.
The Crystallized vs. Fluid Intelligence Gap
Modern AI excels at using learned knowledge and experience to solve problems within a known distribution. This is crystallized intelligence. In contrast, fluid intelligence is the capacity to reason and solve problems in unique situations using problem-independent skills.
The authors posit that current "advanced" paradigms—including transfer learning, meta-learning, and zero-shot learning—are fundamentally task-transductive. Rather than generating entirely new reasoning strategies on the fly, these systems find clever ways to map existing knowledge onto new data. They are essentially filling a new vessel with old wine.
A New North Star: The Giftedness Roadmap
To bridge this gap, the paper proposes a shift in evaluation metrics. Instead of focusing on niche games or massive datasets, the authors suggest using psychometric benchmarks designed for humans: the Naglieri Nonverbal Ability Test (NNAT) and Raven’s Progressive Matrices (RPM). These tests are "culture-neutral," requiring no reading or linguistic knowledge, making them ideal for testing pure machine reasoning.
Specifically, the authors target the Level D NNAT, which is used in California to identify "gifted" 3rd-grade students. To guide research, they propose three escalating challenges:
- The Two-Year Challenge (Verification): The AI must achieve 95% accuracy on a True/False version of the test, verifying if a given answer is correct.
- The Five-Year Challenge (Recognition): The AI must achieve 95% accuracy on the standard multiple-choice format without any human assistance in "coding" the images into symbolic logic.
- The Ten-Year Challenge (Recall/Generation): The AI must generate the correct missing tile from scratch. This prevents the model from using elimination tactics or exploiting artifacts in the multiple-choice options.
Key Insights and the "Human-in-the-Loop" Bias
A critical observation in the paper is that many current AI systems "solving" these puzzles are actually benefiting from human intervention. Researchers often translate visual puzzles into symbolic logic (e.g., "square inside circle") before feeding them to the AI. True fluid intelligence must handle raw visual input end-to-end, perceiving a pattern, inducing a rule, and applying it to a missing piece without symbolic pre-processing.
Current State of the Art and Limitations
Existing attempts to solve these puzzles highlight the difficulty of the task. For instance, a 2018 DeepMind study achieved 75% accuracy on RPM problems using deep learning. However, the model's performance plummeted when shapes or colors were slightly altered, indicating that the system had "crystallized" on the training set rather than learning a generalizable reasoning rule.
Other approaches using fractal geometry or self-similarity measures are often dismissed as "secondary mechanisms"—shortcuts that solve the puzzle without actual reasoning. Currently, no AI system can meet the "Five-Year Challenge" of end-to-end visual recognition and reasoning.
Conclusion: A Call to Action
The authors challenge the AI community to move away from massive training sets and task-specific optimization. Fluid intelligence should not require 100,000 examples to solve the 100,001st puzzle. By focusing on generative reasoning and raw visual perception, we can move toward an AI that is not just a master of games, but truly "gifted" in its ability to navigate the unknown.
Summary updated