LLM is not a reasoning – they are really good at planning

Large language models (LLMs), such as OpenAI’s O3, Google’s Gemini 2.0, and DeepSeek’s R1, have shown significant progress in solving complex problems, producing human-like texts and even writing code accurately. These advanced LLMs are often called “Inference Model” Extraordinary ability to analyze and solve complex problems. But actually these models need to be made reasonOr are they just good at it planning? This distinction is subtle and profound, and it has a significant impact on how we understand LLM’s capabilities and limitations.

To understand this difference, let’s compare two situations:

• reasoning: Detectives investigating crime must piece together contradictory evidence, infer which are false, and draw conclusions based on limited evidence. This process involves inference, contradiction resolution and abstract thinking.
• planning: A chess player calculates the best action sequence to check the opponent’s opponent.

Although both processes involve multiple steps, detectives conduct in-depth reasoning to infer, evaluate contradictions and apply general principles to specific cases. On the other hand, chess players are mainly involved in the program, selecting the best action sequence to win the game. As we will see, LLM functions more like a chess player than a detective.

Understanding the Difference: Reasoning and Planning

To realize why LLM is good at planning rather than reasoning, it is important to first understand the difference between the two terms. Reasoning is the process of drawing new conclusions from a given premise using logic and reasoning. It involves identifying and correcting inconsistencies, generating novel insights, not just providing information, making decisions in ambiguous situations, and engaging in causal understanding and counterfactual thinking, such as the “if?” scheme.

On the other hand, the plan focuses on building a range of actions to achieve specific goals. It relies on breaking complex tasks into smaller steps, following known problem-solving strategies, adjusting previously learned patterns to similar problems, and performing structured sequences instead of gaining new insights. Although both reasoning and planning involve step-by-step processing, reasoning requires deeper abstraction and reasoning, while planning follows established procedures without generating new knowledge.

How LLM handles “reasoning”

Modern LLMs, such as OpenAI’s O3 and DeepSeek-R1, are equipped with a technology called “operation chain” (COT) reasoning to improve their problem-solving capabilities. This approach encourages the model to break down the problem into intermediate steps, mimicking the way humans think logically through the problem. To see how it works, consider a simple math problem:

If a store sells apples for $2 per piece, but if you buy more than 5 apples, how much does 7 apples cost?

A typical LLM using COT prompts might solve this:

1. Determine the normal price: 7 * $2 = $14.
2. Confirm the discount applies (starting from 7>5).
3. Calculate the discount: 7 * $1 = $7.
4. Subtract the discount from the total: $14 – $7 = $7.

By devising a series of steps explicitly, the model minimizes the chance of mistakes resulting from attempting to predict answers in one go. While this step-by-step breakdown makes LLM look like reasoning, it is essentially a form of solving the problem, just like following a step-by-step recipe. On the other hand, a true reasoning process might identify a general rule: If the discount applies to 5 apples, each apple will be $1. One can immediately infer such a rule, but LLM cannot simply follow a structured sequence of computations.

Why is the business chain planning rather than reasoning

While thinking chains (COT) improve the performance of LLMS on logical tasks such as mathematical word problems and coding challenges, it does not involve real logical reasoning. This is because COT follows procedural knowledge and relies on structured steps rather than generating novel insights. It lacks a true understanding of causality and abstract relationships, which means that the model will not engage in counterfactual thinking or consider hypothetical situations that require intuition. Furthermore, COT cannot fundamentally change its approach rather than a trained model, thus limiting its ability to be creative or adaptable to unfamiliar situations.

What does LLM need to become a real reasoning machine?

So, what does LLM need to reason as truly as humans? These key areas require improvements and potential ways to achieve it:

1. Symbolizing understanding: Humans make reasons by manipulating abstract symbols and interpersonal relationships. However, LLM lacks a true symbolic reasoning mechanism. Integrating symbolic AI or hybrid models that combine neural networks with formal logic systems can enhance their ability to truly reason.
2. Causal inference: True reasoning requires understanding causality, not just statistical correlation. One reason must deduce the underlying principle from the data, not just predicting the next token. A study of causal AI, which explicitly simulates causality and can help LLMS transition from planning to reasoning.
3. Self-reflection and metacognition: Humans evaluate their thinking process through inquiries “Is this conclusion meaningful?” On the other hand, LLMS has no mechanism for self-reflection. Building models that can critically evaluate their own output will be a step towards real reasoning.
4. Common sense and intuition: Even if LLMs can gain a lot of knowledge, they often struggle with basic common sense reasoning. This happens because they don’t have real-world experience to shape their intuitions, and they cannot easily recognize the absurdity that humans will immediately gain. They also lack ways to bring real-world dynamics into decision-making. One way to improve this situation may be to build models using common sense engines, which may involve integrating sensory inputs from the real world or using knowledge graphs to help the model better understand the way humans are.
5. Counterfactual thinking: Human reasoning often involves asking, “What if the situation is different?” LLMs struggle in such “if” schemes because they are limited by the data they train. To make the model more human in this case, they will need to simulate the hypothetical situation and understand how changes in variables affect the results. They also need a way to test different possibilities and come up with new insights, not just predicting based on what they have seen. Without these abilities, LLMS would not be able to really imagine other futures—they could only work with what they learned.

in conclusion

While LLMs may make sense, they actually rely on planning techniques to solve complex problems. Whether solving mathematical problems or engaging in logical inferences, they organize known patterns primarily in a structured way, rather than gaining insight into the principles behind them. This distinction is crucial in AI research, because if we mistake complex plans for real reasoning, we may overestimate the true capabilities of AI.

In addition to token prediction and probability planning, the path to real AI reasoning will require basic progress. It will require breakthroughs in symbolic logic, causal understanding and metacognition. Until then, LLM will remain a powerful tool to solve structured problems, but they won’t really think in a human way.