It’s Abstraction, Stupid!

TL;DR: The ultimate question to intelligence is not how to reason, but how to organise experiences.

In a recent development of the large language model, the ability to reason has been highlighted, at least by their advertisement, after the release of OpenAI’s o1 and the Deepseek-R1. These researches enable us to increase performance by increasing the number of computes and samples with sophisticated algorithms such as search algorithms or reinforcement learning. With this so-called reasoning mechanism, our algorithm can solve sophisticated puzzles, math problems, and many complex benchmarks that we consider the trace of intelligence. But the real question I want to ask is, have we really found the last piece of intelligence? If not, what are we currently missing?

Reasoning is the skill for fast and precise problem-solving, which is treated as the fundamental trait of intelligence that allows us to rapidly adapt to different environments and efficiently get what we want. In this procedure, we generally identify the problem, decompose it, and apply known tactics or knowledge. Alternatively, we can use intuition to reduce our search space. However, whether you use reasoning or intuition, both require foundational knowledge. This suggests that our intelligence relies on the representation of accumulated experiences. So, how do we accumulate such experiences in a meaningful way? Do we have a good model for this? Sadly, we have not reached that point yet.

The urge to organise information has existed since the birth of civilization. In ancient Mesopotamia, workers recorded tax incomes on clay tablets. In the digital age, we developed relational databases. But a database does not learn from the data it holds; it is a passive container. This limitation led to the rise of machine learning, the quest to create machines that can actively learn from data.

To solve this, Symbolicists argue that core abstractions can be defined by human-designed rules or axioms, allowing complex problems to be solved through logical deduction. This is the foundation for automatic proof-checkers like Lean. But the real world is too irregular for precise definition. More importantly, we can arbitrarily change the rules, which suggests a symbolic system is not an absolute truth but a tool for fault-tolerant communication and inference.

On the other hand, Connectionists argue for using artificial neural networks to solve irregular patterns by finding their basis. This method has been spectacularly successful, from AlexNet to the current large language model era. Yet, despite gigantic amounts of data and computation, our state-of-the-art machines struggle to abstract experiences in a way that allows them to understand unprecedented situations. For example, the poor performance of these models on the Abstraction and Reasoning Corpus (ARC) is telling. Models that can process trillions of words still fail at simple visual puzzles that a human can solve in seconds, indicating a fundamental gap in their ability to form and manipulate abstract concepts.

So how can we make the system learn good abstraction to reason? A good example is our brain, especially our hippocampus. The hippocampus makes it possible to store long-term memory from short-term memory while sleeping, but that is not its only role. The hippocampus is for combining multiple experiences (or contexts) into a single point of view. For instance, it allows us to create a cognitive map an abstract mental model of our environment that is far more flexible than a simple recording of sensory data.

This ability to organise experience is rooted in what neuroscientists call relational binding: the process of linking disparate pieces of information (a person’s face, their name, the place you met) into a single, cohesive episodic memory, often from a single exposure. However in this case, a standard causal transformer struggles because its nature is perceiving its nature as autoregressive way. In contrast, an encoder-only architecture like ENTP, with its bidirectional attention, can view the entire context at once. This allows it to form a much richer representation where each element is understood in relation to the whole, making it inherently better at relational binding. Diffusion-based models take this a step further. Research suggests that by learning to iteratively refine an entire block of text from a noisy state, they operate on the context in a non-linear, flexible manner that is conceptually closer to the hippocampus’s ability to synthesise disparate inputs.

In conclusion, while the reasoning capabilities of the latest models are impressive, I think they are a symptom of intelligence, not its cause. The foundational problem is one of representation. The current connectionist paradigm has shown fundamental limits in its ability to create robust abstractions from experience, a weakness tied directly to its architectural priors. The missing piece, and therefore the next direction for research, is the development of architectures that are explicitly designed for dynamic, structured model-building. I think we need to think more about organising experiences, whether by referencing our nature or by creating its own, and investigate its mechanisms so that we can build holistic context rather than just sequential prediction for our mechanically interpretable intelligence.