Skip to content
TheChatArchiveThe Chat Archive
Geminigemini-2.5-pro-preview-03-25· GoogleHugging Face LMArena Human Preference 140K

How thinking model learn different from normal models

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4.1-2025-04-14’s answer.
  1. User

    How thinking model learn different from normal models

  2. Gemini Assistant

    Okay, let's break down the difference between how "normal" AI models learn and how models exhibiting "thinking" (often associated with techniques like Chain-of-Thought, or CoT) appear to learn or operate.

    It's crucial to understand that "thinking models" aren't usually a fundamentally different type of base model architecture. They are often the same powerful base models (like GPT-4, PaLM, Llama) as "normal" models, but they are either prompted differently or fine-tuned specifically to exhibit step-by-step reasoning.

    Here's a comparison focusing on the "learning" aspects:

    1. "Normal" Model Learning (Standard Pre-training & Fine-tuning):

    • Goal: Predict the next word/token based on the preceding context, or follow instructions directly.
    • Learning Process (Pre-training):
      • Fed massive amounts of text data from the internet, books, etc.
      • Learns statistical patterns, grammar, facts, and some level of implicit reasoning by predicting missing words or the next word in sequences.
      • The model learns associations: "Paris" is associated with "France," "capital," "Eiffel Tower." It learns that 2+2 is usually followed by = 4.
    • Learning Process (Fine-tuning / Instruction Tuning):
      • Trained on curated datasets of instructions and desired outputs (e.g., questions and good answers, summaries, translations).
      • Often uses techniques like Reinforcement Learning from Human Feedback (RLHF) to align the model's outputs with human preferences (helpfulness, honesty, harmlessness).
    • Output Generation (Inference): When given a prompt (e.g., "What is the capital of France?"), it directly uses its learned associations and patterns to generate the most probable answer ("Paris"). For complex problems, it might try to directly generate the final answer, sometimes making logical leaps or errors because it hasn't explicitly mapped out the intermediate steps.

    2. "Thinking" Model Learning/Operation (e.g., using Chain-of-Thought):

    • Goal: Solve complex problems by breaking them down into intermediate steps, showing the reasoning process before arriving at the final answer.
    • Learning Process (Base Model):
      • Starts with the same pre-trained base model as the "normal" model. It has the same foundational knowledge and implicitly learned reasoning capabilities.
    • Key Difference - Eliciting Reasoning (Not necessarily new learning):
      • A) Prompting Techniques (Most Common - e.g., Chain-of-Thought):
        • This isn't "learning" in the sense of updating the model's weights. Instead, it's about how you ask the model to perform the task at inference time.
        • You explicitly instruct the model to "think step by step" or provide examples (few-shot prompting) where the desired output includes the reasoning process.
        • Example Prompt: Q: Roger has 5 tennis balls. He buys 2 cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now? A: Let's think step by step. Roger started with 5 balls. He bought 2 cans, and each can has 3 balls, so he bought 2 * 3 = 6 balls. In total, he now has 5 + 6 = 11 balls. The final answer is 11.
        • By prompting this way, you guide the existing model to allocate its computational resources towards generating intermediate steps, leveraging its learned knowledge in a more structured sequence. It forces the model to "show its work."
      • B) Fine-tuning for Reasoning:
        • This is a form of learning (updating weights).
        • The base model is further fine-tuned on a dataset specifically containing problems and their step-by-step solutions (the reasoning chain).
        • The model learns to explicitly generate these reasoning steps as part of its output for relevant problems. It learns that producing the intermediate steps is the desired behavior for certain types of queries.
    • Output Generation (Inference): When given a complex prompt (and potentially the "think step by step" instruction), the model first generates the intermediate reasoning steps based on its training or the prompt's guidance, and then derives the final answer from those steps. This often leads to more accurate results for multi-step problems.

    Summary Table:

    Feature"Normal" Model"Thinking" Model (e.g., using CoT)
    Base ModelStandard LLM (e.g., GPT, PaLM)Often the same standard LLM
    Pre-trainingLearns patterns, facts, implicit reasoningSame
    Fine-tuningGeneral instruction following, alignment (RLHF)Can be the same, OR specifically fine-tuned on reasoning datasets
    Key MechanismDirect answer generation based on learned patternsEliciting/Generating intermediate steps via prompting or specialized fine-tuning
    "Learning" FocusLearning direct input-output mappingsLeveraging base knowledge + learning to structure output as a reasoning chain (either via prompts or fine-tuning)
    InferenceInput -> Direct OutputInput -> Reasoning Steps -> Final Output

    In essence: "Thinking" models aren't usually learning fundamentally different things during pre-training. The difference lies in how they are guided (prompted) or specifically trained (fine-tuned) to structure their output, making their implicit reasoning capabilities explicit and sequential. This structured approach often improves performance on tasks requiring complex, multi-step logic.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.