Reasoning Models: The Rise of System 2 Thinking

For the first few years of the LLM era, AI operated primarily on "System 1" thinking: fast, intuitive, and probabilistic. You prompt it, and it predicts the next token instantly. While impressive, this often led to "hallucinations" in logic, math, and complex coding.

The new frontier—Reasoning Models—introduces "System 2" thinking: slow, deliberate, and verifiable reasoning.

System 1 vs. System 2: Based on Daniel Kahneman's psychology. System 1 is your "gut instinct" (fast, automatic). System 2 is your "focused concentration" (slow, analytical). Reasoning models bring System 2 to AI.

⚙️ How Reasoning Models Work

Reasoning models don't just predict the next token; they search for the best path to an answer.

1. Chain-of-Thought (CoT)

Instead of going directly from

Question

$\rightarrow$

Answer

, the model generates a hidden (or visible) internal monologue:

Question

$\rightarrow$

Thought 1

$\rightarrow$

Thought 2

$\rightarrow$

Verification

$\rightarrow$

Answer

2. Test-Time Compute (The "Slow" Part)

Traditional models use the same amount of compute for a simple "Hi" as they do for a complex physics problem. Reasoning models use test-time compute, meaning they spend more processing power during the response phase to think through the problem.

3. Monte Carlo Tree Search (MCTS) and Verifiers

Advanced reasoning models (like the o1 series) explore multiple possible reasoning paths. They use a Verifier (a separate reward model) to score each path. If a path leads to a contradiction, the model discards it and tries another—much like a human solving a puzzle.

🔄 The Reasoning Loop

Prompt ──► [Internal Reasoning Loop] ──► Final Answer
               ▲            │
               │            ▼
        Check for Errors $\leftarrow$ Explore New Path

🛠️ Practical Application: When to Use Reasoning Models

Reasoning models are computationally expensive and slower. You shouldn't use them for everything.

Task	Use Standard LLM (System 1)	Use Reasoning Model (System 2)
Creative Writing	✅ Best (Faster, more fluid)	❌ Overkill
Email Drafting	✅ Best	❌ Too slow
Complex Coding	⚠️ Good for snippets	✅ Best for architecture/debugging
Advanced Math	❌ Prone to errors	✅ Best (Verifiable steps)
Strategic Planning	⚠️ Okay for brainstorming	✅ Best for logic-checking
Fact Retrieval	✅ Fast	⚠️ Only if verification is needed

💡 Summary & Key Takeaways

🎯 The End of the "Instant" Answer

We are trading speed for accuracy. "Thinking" time is now a feature, not a bug.

🎯 Verifiable Logic

Because these models show their work (or use internal verifiers), the results are significantly more reliable for technical tasks.

🎯 Compute Shift

The focus has shifted from just "bigger training sets" to "smarter inference-time processing."

❓ Knowledge Check

Q1: What is "Test-Time Compute"?

A) The power used to train the model on a GPU cluster.
B) The computational resources spent by the model while generating a response to a prompt.
C) A way to test if a model is hallucinating.
D) The time it takes to upload a prompt. Correct Answer: B

Q2: How does the Verifier contribute to the reasoning process?

A) It checks the grammar of the final answer.
B) It acts as a judge that scores different reasoning paths, allowing the model to discard incorrect ones.
C) It translates the prompt into different languages.
D) It increases the speed of the response. Correct Answer: B

Q3: Why is a Reasoning Model not ideal for writing a friendly greeting email?

A) It cannot write emails.
B) It is too expensive and slow for a task that requires intuition and fluidity rather than deep logic.
C) It doesn't understand social cues.
D) It can only output code. Correct Answer: B