LLMs Are Not Reasoning—They’re Just Really Good at Planning

Large language models (LLMs) like OpenAI’s o3, Google’s Gemini 2.0, and DeepSeek’s R1 have shown remarkable progress in tackling complex problems, generating human-like text, and even writing code with precision. These advanced LLMs are often referred as “reasoning models” for their remarkable abilities to analyze and solve complex problems. But do these models actually reason, or are they just exceptionally good at planning? This distinction is subtle yet profound, and it has major implications for how we understand the capabilities and limitations of LLMs.

To understand this distinction, let’s compare two scenarios:

• Reasoning: A detective investigating a crime must piece together conflicting evidence, deduce which ones are false, and arrive at a conclusion based on limited evidence. This process involves inference, contradiction resolution, and abstract thinking.
• Planning: A chess player calculating the best sequence of moves to checkmate their opponent.

While both processes involve multiple steps, the detective engages in deep reasoning to make inferences, evaluate contradictions, and apply general principles to a specific case. The chess player, on the other hand, is primarily engaging in planning, selecting an optimal sequence of moves to win the game. LLMs, as we will see, function much more like the chess player than the detective.

Understanding the Difference: Reasoning vs. Planning

To realize why LLMs are good at planning rather than reasoning, it is important to first understand the difference between both terms. Reasoning is the process of deriving new conclusions from given premises using logic and inference. It involves identifying and correcting inconsistencies, generating novel insights rather than just providing information, making decisions in ambiguous situations, and engaging in causal understanding and counterfactual thinking like “What if?” scenarios.

Planning, on the other hand, focuses on structuring a sequence of actions to achieve a specific goal. It relies on breaking complex tasks into smaller steps, following known problem-solving strategies, adapting previously learned patterns to similar problems, and executing structured sequences rather than deriving new insights. While both reasoning and planning involve step-by-step processing, reasoning requires deeper abstraction and inference, whereas planning follows established procedures without generating fundamentally new knowledge.

How LLMs Approach “Reasoning”

Modern LLMs, such as OpenAI’s o3 and DeepSeek-R1, are equipped with a technique, known as Chain-of-Thought (CoT) reasoning, to improve their problem-solving abilities. This method encourages models to break problems down into intermediate steps, mimicking the way humans think through a problem logically. To see how it works, consider a simple math problem:

If a store sells apples for $2 each but offers a discount of $1 per apple if you buy more than 5 apples, how much would 7 apples cost?

A typical LLM using CoT prompting might solve it like this:

1. Determine the regular price: 7 * $2 = $14.
2. Identify that the discount applies (since 7 > 5).
3. Compute the discount: 7 * $1 = $7.
4. Subtract the discount from the total: $14 – $7 = $7.

By explicitly laying out a sequence of steps, the model minimizes the chance of errors that arise from trying to predict an answer in one go. While this step-by-step breakdown makes LLMs look like reasoning, it is essentially a form of structured problem-solving, much like following a step-by-step recipe. On the other hand, a true reasoning process might recognize a general rule: If the discount applies beyond 5 apples, then every apple costs $1. A human can infer such a rule immediately, but an LLM cannot as it simply follows a structured sequence of calculations.

Why Chain-of-thought is Planning, Not Reasoning

While Chain-of-Thought (CoT) has improved LLMs’ performance on logic-oriented tasks like math word problems and coding challenges, it does not involve genuine logical reasoning. This is because, CoT follows procedural knowledge, relying on structured steps rather than generating novel insights. It lacks a true understanding of causality and abstract relationships, meaning the model does not engage in counterfactual thinking or consider hypothetical situations that require intuition beyond seen data. Additionally, CoT cannot fundamentally change its approach beyond the patterns it has been trained on, limiting its ability to reason creatively or adapt in unfamiliar scenarios.

What Would It Take for LLMs to Become True Reasoning Machines?

So, what do LLMs need to truly reason like humans? Here are some key areas where they require improvement and potential approaches to achieve it:

1. Symbolic Understanding: Humans reason by manipulating abstract symbols and relationships. LLMs, however, lack a genuine symbolic reasoning mechanism. Integrating symbolic AI or hybrid models that combine neural networks with formal logic systems could enhance their ability to engage in true reasoning.
2. Causal Inference: True reasoning requires understanding cause and effect, not just statistical correlations. A model that reasons must infer underlying principles from data rather than merely predicting the next token. Research into causal AI, which explicitly models cause-and-effect relationships, could help LLMs transition from planning to reasoning.
3. Self-Reflection and Metacognition: Humans constantly evaluate their own thought processes by asking “Does this conclusion make sense?” LLMs, on the other hand, do not have a mechanism for self-reflection. Building models that can critically evaluate their own outputs would be a step toward true reasoning.
4. Common Sense and Intuition: Even though LLMs have access to vast amounts of knowledge, they often struggle with basic common-sense reasoning. This happens because they don’t have real-world experiences to shape their intuition, and they can’t easily recognize the absurdities that humans would pick up on right away. They also lack a way to bring real-world dynamics into their decision-making. One way to improve this could be by building a model with a common-sense engine, which might involve integrating real-world sensory input or using knowledge graphs to help the model better understand the world the way humans do.
5. Counterfactual Thinking: Human reasoning often involves asking, “What if things were different?” LLMs struggle with these kinds of “what if” scenarios because they’re limited by the data they’ve been trained on. For models to think more like humans in these situations, they would need to simulate hypothetical scenarios and understand how changes in variables can impact outcomes. They would also need a way to test different possibilities and come up with new insights, rather than just predicting based on what they’ve already seen. Without these abilities, LLMs can’t truly imagine alternative futures—they can only work with what they’ve learned.

Conclusion

While LLMs may appear to reason, they are actually relying on planning techniques for solving complex problems. Whether solving a math problem or engaging in logical deduction, they are primarily organizing known patterns in a structured manner rather than deeply understanding the principles behind them. This distinction is crucial in AI research because if we mistake sophisticated planning for genuine reasoning, we risk overestimating AI’s true capabilities.

The road to true reasoning AI will require fundamental advancements beyond token prediction and probabilistic planning. It will demand breakthroughs in symbolic logic, causal understanding, and metacognition. Until then, LLMs will remain powerful tools for structured problem-solving, but they will not truly think in the way humans do.