Beyond Fixed Representations: The Vocabulary and Verifier Gaps in Open-Ended AI

Noosaga Placements

• Artificial Intelligencesubfield95%
  The paper is a conceptual position paper on open-ended AI, directly addressing the nature of intelligence and innovation, and is categorized under cs.AI.
  arXiv:2607.09560v1 [cs.AI]This paper argues that building stronger intelligent systems capable of open-ended innovation requires additional classes of operations...We characterize the distance between current AI systems and genuinely open-ended intelligence through two gaps.
• Deep Learningframework80%
  The paper critiques current deep learning systems (including LLMs) for operating within fixed representational frames and not being able to autonomously expand their vocabulary or evaluators.
  Modern AI systems are increasingly being evaluated for their ability to reason, code, prove theorems... but they share a structural limitation: the representational frame... is typically fixed and supplied in advance.Current AI systems are powerful searchers and recombiners within a provided vocabulary, but they rarely decide, on their own, how and when the vocabulary should grow.
• Behavior-Based AIframework70%
  The paper discusses open-endedness and innovation, which are related to behavior-based AI's emphasis on emergent capabilities, but does not directly use or extend that framework.
  Open-ended innovation is what allowed humans to create knowledge...We propose a ladder of innovation autonomy...
• Machine Learningsubfield60%
  The paper discusses training objectives (e.g., next-token prediction) and learning from data, but does not propose new ML algorithms or empirical results.
  The prevalent next-token prediction objective can be interpreted as a specific instantiation of discrepancy reduction...Enrichment of objectives requires corresponding change in data.
• Foundation Modelsframework70%
  The paper critiques foundation models (including LLMs) for their inability to perform open-ended innovation, as they are trained on fixed distributions and operate within fixed representational spaces.
  Existing models are increasingly powerful at solving the type-one problems that operate within a given and frozen representational framework...Current models do form concepts and build vocabularies, but these are inherited from the training distribution rather than initiated autonomously.
• Knowledge Representation Reasoningsubfield50%
  The paper focuses on representational primitives and conceptual spaces, which are central to knowledge representation, but does not engage with formal KR systems.
  the representational frame within which the model operates, including its conceptual vocabulary...the creation, stabilization, and reuse of new representational primitives
• Reinforcement Learningframework50%
  The paper mentions reinforcement learning as one of the frameworks that can be seen as a special case of cognitive discrepancy reduction, but does not use or extend RL directly.
  reinforcement learning (RL) formalizes action selection under goals and rewardsEach provides an effective account of a particular cognitive operation under a particular cognitive prior.

Abstract

Modern AI systems are increasingly being evaluated for their ability to reason, code, prove theorems, use tools, and long-horizon research tasks. These are powerful capabilities, but they share a structural limitation: the representational frame within which the model operates, including its conceptual vocabulary, the space of admissible solutions it can search, and the criteria by which success is evaluated, is typically fixed and supplied in advance. This paper argues that building stronger intelligent systems capable of open-ended innovation requires additional classes of operations: the creation, stabilization, and reuse of new representational primitives, which alter the space being searched rather than simply searching within it. We characterize the distance between current AI systems and genuinely open-ended intelligence through two gaps. The first is the vocabulary gap, the difficulty of inventing and stabilizing new representational primitives rather than merely recombining existing ones. The second is the verifier gap, the difficulty of judging the value of a new primitive when its full payoff may be visible only after future reuse. We interpret both gaps through a unified framework of intelligence as cognitive discrepancy reduction. By viewing intelligent behaviors as a sequence of cognitive transformations, we distinguish intra-space transformations which operate within a fixed representational frame, from generative transformations which may modify the frame itself. On this basis, we propose a ladder of innovation autonomy and outline several directions for advancing open-ended AI, including objectives that reward useful representational change, persistent memory architectures for invented primitives, and adaptive verification mechanisms capable of evolving alongside the representations they evaluate.

Paper Context