Editing Reasoning Models and the Architecture of the Thought (section)

== <span style="color: #FFFFFF;">Remembering</span> ==
* '''Reasoning Models''' — Advanced AI models designed specifically to execute complex, multi-step logical deduction, mathematics, and problem-solving, moving beyond simple pattern matching and text generation.
* '''System 1 vs. System 2 Thinking''' — A psychological framework by Daniel Kahneman applied to AI. *System 1*: Fast, automatic, intuitive (Standard LLMs). *System 2*: Slow, deliberate, analytical, step-by-step logic (Reasoning Models).
* '''Chain-of-Thought (CoT) Prompting''' — The foundational technique that unlocked AI reasoning. Instead of asking the AI for the final answer, the user prompts the AI to "think step-by-step," forcing the model to explicitly output its intermediate logical calculations.
* '''Latent Reasoning (Hidden Thoughts)''' — A feature of advanced reasoning models (like OpenAI's o1). Before generating the final output for the user, the model enters a hidden, internal loop where it generates, tests, and discards thousands of logical steps in the background.
* '''Tree of Thoughts (ToT)''' — An advanced reasoning architecture. Instead of a single, linear chain of logic, the AI generates multiple different, branching paths of logic simultaneously. It evaluates each branch, abandons the dead ends, and follows the most promising branch to the solution.
* '''Self-Correction''' — A critical capability of a true Reasoning Model. The ability to realize mid-thought that a previous logical step was mathematically incorrect, backtrack, fix the error, and resume the calculation.
* '''Search and Planning''' — The integration of classic computer science search algorithms (like Monte Carlo Tree Search, used in AlphaGo) with LLMs. The model treats reasoning as a massive maze, actively searching for the correct path to the goal.
* '''Compute-Optimal Scaling (Inference-Time Compute)''' — A massive shift in AI economics. Traditional models spend all their massive computing power during *training*. Reasoning models spend massive amounts of computing power during *inference* (the time it takes to answer the user), taking minutes or hours to "think" about a single prompt.
* '''Formal Logic & Math Validation''' — The benchmarks used to test reasoning models. While standard LLMs are tested on poetry or trivia, reasoning models are tested on elite, PhD-level physics, complex coding challenges, and formal mathematical proofs.
* '''The Verification Gap''' — The principle that it is vastly computationally easier to *verify* if a mathematical answer is correct than it is to *generate* the correct answer from scratch. Reasoning models exploit this by generating multiple answers and verifying them internally.
</div>

<div style="background-color: #006400; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">