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

When using a thinking model, if you do not let it think, is it possible for it to still…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with o3-mini’s answer.
  1. User

    When using a thinking model, if you do not let it think, is it possible for it to still be more accurate than normal models?

  2. Gemini Assistant

    Yes, it is possible for a model designed or trained for "thinking" (like using Chain-of-Thought, CoT, or similar reasoning techniques) to be more accurate than a "normal" model, even when you don't explicitly prompt it to show its thinking steps.

    Here's why:

    1. Underlying Training Improvements: Models fine-tuned on reasoning tasks (like CoT datasets) often learn better internal representations and reasoning pathways. This training can improve their general ability to handle complex questions, even when they aren't explicitly asked to "think step-by-step." They might implicitly follow a more logical internal process to arrive at the answer.
    2. Better Generalization: The process of learning to break down problems and reason sequentially can lead to better generalization on various tasks, including those where only a direct answer is requested. The model becomes inherently "smarter" at connecting information.
    3. Implicit Reasoning: Even without an explicit prompt like "Let's think step-by-step," a model trained for reasoning might still perform some degree of internal, implicit reasoning before generating the final answer, especially if the question is complex enough to trigger those learned patterns.
    4. Correlation with Overall Capability: Often, models designated as "thinking models" (like advanced versions of GPT or Gemini) are also simply larger, more capable base models with more extensive training overall. So, their higher accuracy might stem from their general superiority, which is further enhanced by reasoning-specific training.

    However, there are important caveats:

    • Full Potential Not Realized: While it might be more accurate than a "normal" model, it will likely not be as accurate as it could be if you did explicitly prompt it to think step-by-step, especially for complex problems where explicit reasoning prevents errors. The explicit prompt forces deliberation and allows the model to allocate computational effort (token generation) to the reasoning process itself.
    • Task Dependency: For simple tasks, the difference might be negligible. The advantage of a "thinking model" (even when not explicitly prompted) becomes more apparent as task complexity increases.
    • Definition of "Normal Model": The comparison depends heavily on what the "normal model" is. If it's a significantly smaller or less capable model, the "thinking model" will likely outperform it regardless. If it's a similarly sized base model without the reasoning fine-tuning, the "thinking model" might have a slight edge due to that specialized training.

    In summary: Training a model to "think" can improve its fundamental reasoning capabilities, potentially leading to higher accuracy even on direct-answer tasks compared to models without such training. However, explicitly prompting the model to use its reasoning steps usually unlocks its maximum accuracy potential for complex problems.

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

Identical first prompt