Large language models may be good at mimicking reasoning, but they struggle to reason through simple causality. Neurosymbolic AI can help solve this conundrum.
It may seem counterintuitive to see an artificial intelligence (AI) model solve an International Mathematical Olympiad-level problem with flying colours yet be unable to tell the time accurately on an analogue clock.
However, this has indeed happened. Google DeepMind’s Gemini Deep Think was able to score 35 points at the 2025 International Mathematical Olympiad and earn a gold medal, all while thinking in natural language, according to Stanford’s 2026 AI Index Report.
The best model of the same generation, tested on visual reasoning AI benchmark ClockBench, read the clock correctly only 50.6% of the time, compared with 90.1% for humans. When the models fail, they tend to be one to three hours off, whereas a human would be about three minutes off.
One promising answer to addressing this inconsistency is neurosymbolic AI, which pairs the pattern learning of neural networks with the rule-following logic of symbolic systems.
Why jagged AI is an operational risk
Large language models (LLMs) have become extremely good at mimicking reasoning, which is why enterprises are building them into everyday work. They create long chains of thought that help produce good code, concise document summaries and original maths proofs. But what they still struggle to do is reliably identify properties of the world around them or reason through simple causality.
AI models that can accomplish one task can fail miserably at others. Researchers call this “jagged intelligence” – also known as the “jagged frontier” – which highlights the unpredictable boundary between tasks best suited for humans and those for machines.
This is important, as building workflows around AI that can fail at any given moment creates a vulnerability in an enterprise’s products and operations. Models can also insert hallucinated details into contracts, misunderstand a customer’s issue or create compliance violations.
The risk is measurable. In fact, no model scores above 70.2% on τ-bench, a benchmark that measures how well AI agents complete real-world tasks using tools while following policy rules. Models are falling short in real-world scenarios as well. On benchmarks from Vals AI, reported in the AI Index, the best models achieve only 69.4% accuracy when extracting information from mortgage tax certificates, 68.3% on long financial documents such as credit agreements and 63.3% on entry-level financial analyst tasks.
For business leaders, this means AI reliability is an operational risk that needs to be managed, and a model’s standout results offer minimal assurance of how consistently it will perform inside their own workflows.
Where neurosymbolic AI comes in
As its name indicates, neurosymbolic AI merges neural networks and symbolic AI. Neural networks recognize patterns in data, while symbolic AI uses formal representations of knowledge and reasoning to create a solution. Combined, the two offer the best of both worlds: using neural networks to process information and symbolic AI to reason over it creates a verifiable solution.
It is easy to think that bigger models will simply outgrow these weaknesses, and the 2025 olympiad gold can be seen as evidence to support this argument. But this same generation of models is still struggling with seemingly simpler tasks, such as reading clocks, which suggests that scale alone has not solved the reliability issue.
Symbolic reasoning has shown great promise in addressing jagged intelligence. On PlanBench, a benchmark designed to test AI planning ability rather than pattern recognition, researchers found that when the underlying structure of a problem was obfuscated, most frontier models solved fewer problems in several domains.
OpenAI’s GPT-5 dropped from 21 of 45 problems to 12 on Blocksworld, and from 15 to 7 on Sokoban, while DeepSeek’s open-source reasoning LLM DeepSeek-R1 dropped to solving 3 problems on Blocksworld and 0 problems on Floortile and Sokoban. The symbolic solver LAMA, however, stayed consistent in its ability to solve problems regardless of whether the problems were obfuscated, according to the AI Index. This shows that a structured approach can yield better results.
Major labs are starting to implement neurosymbolic reasoning. Google DeepMind’s AlphaGeometry combines neural language models with symbolic deduction engines to solve olympiad-level geometry problems while producing proofs that can be verified along the way. Reliable products will increasingly depend on models like these.
Regulation and AI agents raise the stakes
With the EU putting AI laws into effect, we are starting to see what regulations will look like for AI. Since 2 August 2026, the European Commission has been enforcing parts of the AI Act, including rules requiring that people be told when they are interacting with AI. Stricter rules for high-risk AI systems will apply from 2 December 2027.
As these requirements grow, AI built with symbolic reasoning can make its decision process easier to explain and document. Neural networks have historically struggled with this, so adding symbolic reasoning helps us build dependable AI.
The future of AI will be focused on agents that can perform tasks for companies. These agents will need to be able to think through problems and create reliable solutions. By adopting a neurosymbolic approach, we can enable AI models to verify that their actions are correct.
How companies should evaluate AI
What this means for companies is that they need to change their evaluation methods when looking to purchase these products. Business leaders will want to look at how a model performs on an AI benchmark that reflects their workload, with constraints applied to it.
They will also need to ensure their proofs of concept enable them to see how the model can think through problems and correct itself. The model should also be able to show how it reached its answer.
While AI capabilities continue to advance, questions around reasoning, consistency and reliability remain unresolved. We had expert systems that could think precisely but weren’t reliable. We now have deep learning that is reliable but cannot think precisely.
Neurosymbolic AI will enable us to take LLMs and make them trustworthy enough for tasks that demand reliability. By allowing neural networks to think symbolically, companies can achieve both precision and reliability.





