Neuro-symbolic AI: how to finally get language models to do arithmetic

The world's largest technology companies have invested tens of billions of dollars in language models capable of writing code, translating texts, and generating coherent reasoning on any topic. But ask any of them to divide 7429 by 17 — and the lottery begins. Sometimes the answer will be correct, sometimes not, and sometimes the model will produce a confident hallucination indistinguishable from the right answer. This is not a bug that can be fixed with a patch. It is a fundamental limitation of the architecture, and one researcher on the Habr platform not only pointed out the problem but proposed a working solution.

Before understanding the solution, it's worth understanding why the problem exists in the first place. Language models based on the Transformer architecture process information through continuous mathematical transformations — matrix multiplications, activation functions, attention mechanisms. All of this works brilliantly for tasks where patterns, context, and approximate estimates matter. But arithmetic is a discrete operation. Two plus two equals exactly four, not 3.97 and not 4.02. When a neural network tries to imitate precise calculations through approximation, it inevitably fails — especially on numbers that haven't appeared in the training set or fall outside familiar ranges.

The industry has tried workarounds. Chain-of-Thought prompting forces the model to reason step-by-step, which improves results but doesn't address the root problem — each step in the chain is still executed by the same approximating architecture. Fine-tuning on mathematical tasks helps in narrow scenarios but doesn't generalize. Increasing context window size and model scale yields marginal improvements at the cost of colossal computational expenses. Ultimately, companies arrived at a paradoxical solution: connecting ordinary calculators to the most powerful neural networks through tool calls. It works, but looks like a crutch attached to a spaceship.

The researcher took a different path, proposing a neuro-symbolic architecture — a hybrid in which neural network components do what they're truly strong at, while symbolic modules take on tasks requiring precision. The idea is not new in academic terms — neuro-symbolic AI has been discussed in scientific circles for several years, and researchers like Yoshua Bengio and Gary Marcus have long pointed to the need for combining two paradigms. But the distance from theoretical discussion to working implementation is enormous, and it is this distance that the author attempted to bridge.

The essence of the proposed architecture is a semantic neural network in which the model doesn't attempt to compute directly but recognizes the type of task and delegates execution to the corresponding symbolic module. The neural network acts as an interpreter of intentions and a router, while precise operations are performed by deterministic algorithms. This is fundamentally different from the external tools approach: instead of a cumbersome API call to a calculator, symbolic logic is embedded directly in the model's architecture, allowing it to work faster and more reliably.

The experimental results described by the author confirm the viability of the approach, though it is premature to speak of a revolution. The key question is scalability. Arithmetic is the simplest case of symbolic reasoning. It is far more complex to integrate symbolic modules for logical inference, planning, or fact verification. If the architecture can expand to these areas, it will truly change the landscape. If not, it will remain an elegant but niche solution.

For the industry, the significance of this work goes beyond the specific implementation. It highlights a trend gaining momentum in recent months: pure neural network scaling is hitting a ceiling, and the future lies with hybrid systems. OpenAI, Google, and Anthropic are already experimenting with various forms of neuro-symbolic integration, though they don't always speak about it openly. The fact that an independent researcher is proposing competitive ideas with open-source code speaks to the maturity of the community and the democratization of AI research.

Ultimately, the work poses the right question: shouldn't we stop forcing neural networks to do what they were not created for, and instead allow each component of the system to do what it does best? The answer seems obvious. All that's left is to implement it at scale.