Large language models, or LLMs, have shown impressive capabilities in language understanding, coding and problem solving. Yet generating fluent and convincing responses is not the same as reasoning, despite appearances to the contrary.
Logical reasoning remains challenging for current LLMs because a single mistake can derail an entire reasoning process. Models must identify relevant information, apply the appropriate rules, track relationships across multiple steps and maintain logical consistency. Existing approaches often seek to improve reasoning by generating longer chains of thought, sampling multiple possible solutions or performing additional searches.
While effective in some settings, these methods can be computationally expensive and provide limited insight into how reasoning is represented and carried out inside the model.
Yuanyuan Lei, Ph.D., an assistant professor in the University of Florida’s Department of Computer & Information Science and Engineering, or CISE, and researchers in her lab are tackling the problem by uncovering the internal mechanisms that support logical reasoning and using that understanding to make LLMs more precise and efficient.
As LLMs are increasingly used in scientific, engineering and decision-making applications, this work could help make AI systems more capable while making their reasoning processes more consistent, reliable and trustworthy.
Three papers authored by members of Lei’s lab were recently accepted by the 2026 Meeting of the Association for Computational Linguistics, a leading international conference in natural language processing and artificial intelligence. The studies investigate logical reasoning by LLMs at three levels: where abstract logical knowledge is represented inside LLMs, how different reasoning abilities are organized and coordinated, and where reasoning trajectories are most vulnerable to failure.
The first study approaches LLM reasoning from a neuro-symbolic perspective. Neural language models are highly effective at processing natural-language expressions, while symbolic logic provides explicit and structured representations of reasoning. A longstanding goal in artificial intelligence is to combine the strengths of these two paradigms.
“Large language models learn from natural language, but logic itself is more abstract than language,” Lei said. “Our work shows that natural-language and symbolic reasoning may share an internal logical representation within LLMs. By identifying and strengthening this shared representation, we can better understand how models encode logic and directly guide them toward more reliable reasoning, offering a new neuro-symbolic perspective for connecting neural and symbolic reasoning.”
The second study examines logical reasoning from the perspective of reasoning specialization and coordination. Different reasoning tasks rely on different strategies: Deduction applies known rules to reach necessary conclusions, induction generalizes from observations and abduction identifies the most plausible explanation for observed evidence. Although these reasoning modes are distinct, complex problem solving often requires them to work together.
“Deduction, induction and abduction each contribute different strengths to problem solving, but in real-world reasoning they rarely operate in isolation,” Lei said. “By helping these specialized representations complement one another, we can move toward LLMs that reason in a more coordinated and flexible way.”
The third study investigates logical reasoning from the perspective of reasoning fragility. LLMs remain vulnerable in multi-step logical deduction, where a single flawed transition can propagate and cause the entire reasoning chain to fail. The researchers identified logical connectives such as “therefore,” “however,” “but,” “because” and “so” as key points of structural fragility, because these words help determine how one reasoning step relates to the next and can shape the direction of the entire reasoning process.
“Not every step in a reasoning chain is equally important,” Lei said. “Certain logical transitions can serve as critical decision points, where a small mistake may send the entire reasoning process in the wrong direction. By identifying these vulnerable moments and intervening selectively, we can guide LLMs toward more reliable reasoning while using computational resources more efficiently.”
Together, these studies provide a multi-level view of logical reasoning in LLMs . Collectively, the studies move from understanding how logical knowledge is represented, to how reasoning abilities are organized, to how reasoning failures can be controlled.
This body of work reflects a broader goal of Lei’s lab: develop AI systems that are not only capable of producing answers, but also able to reason in ways that are more effective, reliable and efficient. By uncovering the internal structure of machine reasoning and using that knowledge to design more principled methods for improving it, the researchers aim to move beyond treating LLM reasoning as a black box.
The three papers were published in the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held July 2–7, 2026, in San Diego, California.