Beyond Fixed Representations: The Vocabulary and Ve... | AI Research

Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI
Modern AI systems have become remarkably skilled at solving complex problems like coding, theorem proving, and scientific research. However, these systems operate within "fixed" frameworks where the rules, vocabulary, and success criteria are provided by humans in advance. This paper argues that to achieve true open-ended innovation—the kind that leads to genuine scientific and social breakthroughs—AI must move beyond simply searching for solutions within a pre-defined space. Instead, it must develop the ability to autonomously create, stabilize, and reuse new representational primitives that fundamentally change the space being searched.

The Two Gaps in Open-Ended Intelligence

The authors identify two primary obstacles preventing current AI from reaching this higher level of autonomy. The first is the "vocabulary gap," which refers to the difficulty of inventing new concepts or tools rather than just recombining existing ones. While current models can synthesize information, they struggle to create new, durable primitives—like the concept of "entropy" or "negative numbers"—that make previously impossible problems solvable.
The second is the "verifier gap." In current AI, success is usually measured by a clear, pre-existing standard, such as whether code runs or a proof is correct. In open-ended innovation, however, the value of a new concept may not be immediately obvious. A new idea might appear useless in the short term, only becoming valuable later as other concepts or technologies evolve. Because current AI lacks a way to judge the long-term potential of a "useless" idea, it often fails to retain the very primitives that could lead to future breakthroughs.

Intelligence as Cognitive Discrepancy Reduction

To better understand these gaps, the paper proposes a unified framework that views intelligence as "cognitive discrepancy reduction." In this view, intelligent behavior is a sequence of transformations aimed at closing the gap between a system’s current state and a desired goal.
When a system encounters a problem it cannot solve with its current tools, it must perform "generative transformations." These are not just searches within a fixed space, but actions that modify the space itself—such as creating new abstractions, analogies, or conceptual structures. By framing intelligence this way, the authors suggest that current AI models are stuck at the lower levels of an "innovation ladder," focusing on intra-space tasks while lacking the mechanisms to evolve their own representational frameworks.

Moving Toward Autonomous Innovation

To advance toward open-ended intelligence, the authors suggest that AI research needs to shift its focus. Rather than relying solely on next-token prediction, future systems require new objectives that explicitly reward the creation of useful, reusable primitives.
This approach would require architectural changes, such as persistent memory systems that allow models to store and refine new concepts over time, and adaptive verification mechanisms that can evolve alongside the representations they evaluate. By moving away from static, goal-oriented environments and toward systems that can autonomously expand their own conceptual vocabulary, AI may eventually be capable of the kind of open-ended innovation that has historically driven human progress.