Large language models (LLMs) are powerful creative tools for writing, code generation, video generation, and companionship. Yet while LLMs may seem the smartest, they can also look the dumbest. So, even though today’s models are remarkably capable, they are still inconsistently correct. They can solve hard math problems and then fumble trivial logic. One of their major bottlenecks is reasoning and reliability, and this is the root cause of why we still cannot fully trust these models in high-stakes agentic settings. Solving this would unlock autonomous scientific research, automation, and decision-making at scale.

There are different technical paths to address the reasoning and reliability bottleneck. These include architectural innovation, training and optimization, interpretability and internal representations in LLMs, failure-mode analysis, and neuro-symbolic or hybrid approaches.

Some of the technical questions involve architectural challenges, such as whether the transformer architecture has a fundamental expressiveness ceiling for multi-step logical inference, or whether it is sufficient in principle with enough depth and width.

Some efforts focus on training and optimization, asking questions such as: “When models fail at reasoning, is it a representation failure, where the right internal structure was never learned, or a retrieval failure, where the structure exists but is not activated correctly?” They also ask: “Does next-token prediction as a training objective fundamentally misalign the model toward fluency over correctness? Can these even be separated in natural language?”

There are also more fundamental questions about whether models learn abstract reasoning schemas that transfer across domains, or whether they are learning a very large library of superficially similar pattern completions.

Other research focuses on interpretability and internal representations in LLMs. Do models have stable, localizable internal representations of logical relations such as implication, negation, and quantification, and if so, where? Is there evidence of multi-step computation happening across layers, or do most reasoning steps collapse into single-layer lookups? Can probing classifiers tell us whether a model “knows” the correct answer before it generates incorrect tokens, and what does that imply about where errors originate?

Some research also focuses on failure modes, asking: “What is the mechanistic cause of hallucination during reasoning? Is it attention over the wrong context, failure of uncertainty representation, or something else?”

Others, including us at CodeX, are exploring neuro-symbolic and hybrid approaches. Can formal logic systems be integrated with neural networks without destroying the flexibility that makes neural networks useful in the first place? Is there a way to give models access to a provably correct symbolic reasoner for subproblems while retaining neural generalization for the rest? What is lost and what is gained by externalizing reasoning into explicit program execution, such as code interpreters or tool use, versus keeping it in the weights? Can neuro-symbolic architectures generalize their hybrid reasoning to truly novel domains, or do they require hand-engineered symbolic scaffolds for each domain?

There are fundamental research questions in neuro-symbolic AI, mainly around the right architecture for integrating neural networks and symbolic AI in a way that leverages the best of both worlds: the flexibility and creativity of neural networks, and the reliability and verifiability of symbolic and automated reasoning approaches.

At CodeX, we are interested in both fundamental research and applications of this approach in law, contracts, compliance, and beyond. Cutting-edge enterprises and startups are already leveraging neuro-symbolic approaches to build more reliable solutions in automated theorem proving, drug discovery, task and motion planning in robotics, workflow automation, and more.

Get in touch if you are interested in collaborating: Sign Up Here