Large language models (LLMs) and vision–language models (VLMs) have achieved remarkable performance across tasks ranging from natural language understanding and code generation to visual question answering and scientific reasoning. Yet beneath these impressive benchmarks lies a persistent and well-documented weakness: the reasoning processes that underpin model outputs remain largely opaque, inconsistent, and brittle. Models frequently arrive at correct answers through superficial pattern matching or sensitivity to prompt structure rather than through genuine logical inference. A minor rephrasing of a question, a subtle shift in context, or the introduction of irrelevant information can cause even the most capable systems to produce contradictory or nonsensical conclusions. This gap between surface-level fluency and deep reasoning reliability represents one of the most consequential open problems in modern AI research.

A central strand of Dr. Mohana’s research involves probing the internal architectures of these models to understand how reasoning capabilities emerge and where they break down. By analysing activations, attention patterns, and representations across different layers and model scales, the work seeks to identify which components contribute meaningfully to logical inference and which merely approximate it. These mechanistic insights directly inform the design of improved training strategies and more rigorous evaluation protocols. Rather than relying solely on end-task accuracy, the research emphasises evaluation under adversarial, noisy, and ambiguous conditions—testing whether models can maintain coherent reasoning when inputs are deliberately designed to expose shallow heuristics. The group also investigates structured prompting methods and intermediate reasoning representations, such as chain-of-thought and tree-of-thought frameworks, to understand how explicit reasoning scaffolds affect both performance and interpretability.

What becomes increasingly clear from this line of investigation is that entangling reasoning, language generation, and factual knowledge within a single monolithic architecture imposes fundamental limits on transparency and reliability. When a model simultaneously serves as a knowledge store, a language processor, and a reasoning engine, it becomes exceedingly difficult to diagnose errors, verify conclusions, or ensure that outputs are grounded in accurate information rather than plausible-sounding confabulation. This recognition motivates a complementary research direction: the development of modular AI architectures that explicitly separate these functions. In such systems, language models handle linguistic comprehension and generation, structured knowledge bases store and retrieve verified factual information, and dedicated reasoning modules operate over these inputs to perform transparent, auditable inference.

The potential advantages of this modular approach extend well beyond interpretability. By decoupling reasoning from language generation, individual components can be independently evaluated, updated, and improved without retraining the entire system. A knowledge base can be corrected or expanded without altering the reasoning engine; a reasoning module can be refined without disturbing the language model’s fluency. This composability also opens pathways toward systems that can explain why they reached a particular conclusion—not merely what that conclusion is. For domains where trust and accountability are paramount—healthcare, legal analysis, scientific discovery, and education—such explainability is not a luxury but a prerequisite for meaningful deployment.

The broader ambition of this research programme is to move the field beyond a paradigm in which reasoning is an emergent and poorly understood by-product of scale, toward one in which it is a deliberate, measurable, and improvable capability. The goal is not simply to build AI systems that produce fluent and persuasive outputs, but to develop systems that can retrieve grounded knowledge, reason over complex and multi-step problems, and produce explanations that are interpretable, reliable, and genuinely aligned with human understanding. As AI systems are entrusted with increasingly consequential decisions, ensuring that their reasoning is robust and transparent is no longer optional—it is essential.