When we evaluate the ability of large language models (LLMs) to solve planning tasks, we usually attribute performance differences to a supposed general “difficulty” of the task. However, recent research, such as that summarized in arXiv:2607.11197, suggests this view is incomplete: variable performance across tasks may be due to LLMs possessing two distinct cognitive skills, not a single ability spectrum. Specifically, two latent dimensions are identified: operational reasoning, which evaluates local action applicability and immediate state transitions; and structural enumeration, which concerns reasoning about goal reachability and landmark structure. This finding has profound implications for software development based on artificial intelligence and for companies seeking to integrate intelligent agents into their processes.
From a technical perspective, operational reasoning improves significantly with model scaling and longer reasoning chains, while structural enumeration remains relatively insensitive to these factors. This means that simply increasing model size or inference time does not improve all facets of planning; a competency-specific approach is required. For companies developing custom software, this distinction is crucial when designing virtual assistants, recommendation systems, or process automation tools. An agent that must guide a user step by step (e.g., in a customer service flow) will benefit from fine-grained operational reasoning; conversely, a strategic planning system requires robust structural enumeration to identify distant goals and optimal paths.
At Q2BSTUDIO, as a software development and technology company, we apply these insights to deliver solutions that go beyond mere LLM integration. Our AI services are designed taking into account the specific competencies each project needs. For instance, when building a conversational agent for customer support, we prioritize operational reasoning through structured prompting and chain-of-thought techniques; for a logistics planning engine, we enhance structural enumeration with landmark networks and reinforcement learning. This competency-based approach yields more predictable performance aligned with business needs, avoiding the temptation to rely solely on model scaling.
Furthermore, optimizing these cognitive skills does not happen in a vacuum. The underlying infrastructure plays a fundamental role. At Q2BSTUDIO, we offer cloud services on AWS and Azure that provide the scalability and low latency required to run complex reasoning models. Cybersecurity is also a key factor, especially when AI agents handle sensitive data or make automated decisions; our teams integrate security protocols from the design phase. On the other hand, business intelligence (BI) with Power BI allows visualizing and monitoring the performance of these planning systems, facilitating informed decisions about which competencies to improve and how.
The multidimensional research proposed in the original study uses an item response theory (IRT) model to decompose LLM performance into the two mentioned dimensions. This analysis reveals that, under varying reasoning budgets, larger models show asymmetric gains: they improve in operational reasoning but barely in structural enumeration. This suggests that current scaling efforts may be overlooking a fundamental skill for long-term planning tasks. For companies developing custom software, this finding indicates that investing in specialized prompting techniques, training with structured examples, or even hybrid architectures (such as combining LLMs with classical planners) may be more cost-effective than simply acquiring a larger model.
In the context of AI agents, planning ability is a central pillar. An agent that cannot enumerate structural alternatives to reach a goal will fail in complex scenarios. Therefore, at Q2BSTUDIO, we design agents that integrate dual reasoning modules: a fast component for operational decisions (based on LLM) and a slower, more deliberate one for structural analysis (based on search or symbolic planning). This architecture, inspired by the dual-system theory, allows leveraging the strengths of each competency without depending on a single model.
Adopting this competency-based approach also has implications for evaluation and deployment. Instead of asking “Does the model improve?” we must ask “Which competency improves, under what conditions, and why?” For IT managers and innovation directors, this means traditional benchmark tests can be misleading; it is necessary to design evaluations that isolate each cognitive dimension. Our AI consulting services help companies design these tests and interpret the results to make better investment decisions.
In summary, the evidence of two distinct cognitive skills in LLM planning opens a new path for intelligent software development. At Q2BSTUDIO, we are committed to applying these advances practically, offering custom software, cloud, cybersecurity, and BI solutions that enhance the true potential of artificial intelligence. The next time you evaluate a language model for a planning task, remember that it is not a single skill: it is a spectrum of competencies that deserve to be understood and optimized separately.
