1.4. Frameworks for Reasoning — From Chain-of-Thought to Tool Use
🪄 Step 1: Intuition & Motivation
Core Idea: Even though LLMs can mimic reasoning, they’re still “text completion engines.” They don’t truly plan or verify their thoughts unless we help them do it. That’s where reasoning frameworks come in — methods like Chain-of-Thought, Tree-of-Thought, and ReAct.
These frameworks are like training wheels for model thinking — structured ways to help an LLM “slow down,” “reflect,” and “act deliberately” instead of blurting out the first plausible answer.
Simple Analogy: Imagine trying to solve a math problem in your head. If you just guess, you’ll likely be wrong. But if you write down each step, check your reasoning, and sometimes use a calculator — you’ll be far more accurate.
That’s what these frameworks do for LLMs: they turn wild guessing into structured thought.
🌱 Step 2: Core Concept
Let’s explore the three big reasoning scaffolds — CoT, ToT, and ReAct — each evolving from the last to fix a limitation.
1️⃣ Chain-of-Thought (CoT) — Teaching the Model to Think Step-by-Step
The Chain-of-Thought technique asks the model to show its work by generating intermediate reasoning steps before giving the final answer.
Example Prompt:
“Let’s think step-by-step.”
This simple cue encourages the model to unpack logic incrementally:
Question: “If there are 3 red apples and 2 green apples, how many apples total?” Answer (CoT): “There are 3 red and 2 green, 3 + 2 = 5 apples total.”
Here, the model explicitly thinks out loud — improving accuracy on math, logic, and multi-step tasks.
2️⃣ Tree-of-Thought (ToT) — When One Line of Thinking Isn’t Enough
While CoT is linear, Tree-of-Thought introduces exploration. Instead of committing to one reasoning path, ToT lets the model branch out — like exploring multiple possible solutions before picking the best one.
How it works:
- Each reasoning step becomes a node.
- The model explores several branches (“what if” paths).
- A heuristic (another LLM or a scoring rule) ranks which branch seems most promising.
- The best path is expanded further, forming a search tree of reasoning.
This structure resembles search algorithms like depth-first or breadth-first traversal — but applied to thoughts instead of states.
3️⃣ ReAct — Blending Reasoning with Action
Even CoT and ToT have limits: they only think, but never do. ReAct (Reason + Act) merges both — the model reasons, performs an external action (like a tool call), observes the result, then continues reasoning.
Example flow:
Thought: I need to look up the capital of Canada.
Action: Search["Capital of Canada"]
Observation: Found 'Ottawa'.
Thought: Great, Ottawa is the capital of Canada.
Answer: Ottawa.This creates a feedback loop:
Reason → Act → Observe → Reason again.
It allows the model to use calculators, databases, APIs, or retrieval systems to ground its reasoning in factual evidence.
When to Use Each Framework
| Framework | Ideal For | Drawback |
|---|---|---|
| CoT | Math, logic, structured reasoning | Costly (more tokens) |
| ToT | Planning, multi-step decision-making | Slow (branch explosion) |
| ReAct | Retrieval, data-backed tasks | Needs tool orchestration |
📐 Step 3: Mathematical Foundation
Self-Consistency Sampling (Enhancing CoT Reliability)
To reduce reasoning errors, we can run CoT multiple times and aggregate answers — this is self-consistency.
Formally,
$$ y^* = \text{mode}{f_\theta(x, z_i)}_{i=1}^k $$where each $z_i$ is a different random reasoning path sampled by the model. The final answer $y^*$ is chosen by majority or consensus.
🧠 Step 4: Key Ideas & Assumptions
- CoT assumes reasoning can be improved by making it explicit.
- ToT assumes multiple paths may lead to better solutions than one.
- ReAct assumes reasoning is stronger when paired with external grounding tools.
- All three rely on prompt scaffolding — guiding structure for “how” the model should think, not just “what” it should answer.
⚖️ Step 5: Strengths, Limitations & Trade-offs
✅ Strengths:
- Makes model reasoning traceable and interpretable.
- Boosts accuracy on multi-step or logic-heavy tasks.
- Allows integration with external data and tools.
⚠️ Limitations:
- High token cost (more reasoning steps = more text).
- Latency increases — slower responses.
- Risk of verbosity and circular reasoning loops.
⚖️ Trade-offs:
- CoT improves reliability but increases inference cost.
- ToT increases exploration depth but reduces speed.
- ReAct improves grounding but requires complex orchestration and safety controls.
🚧 Step 6: Common Misunderstandings
🚨 Common Misunderstandings (Click to Expand)
- “CoT = guaranteed correctness.” → No, it just exposes reasoning — still depends on internal accuracy.
- “ReAct means real understanding.” → Not quite; the model doesn’t truly “know,” it just learns to use tools effectively.
- “ToT always beats CoT.” → Only when the problem benefits from exploring multiple paths — otherwise it’s overkill.
🧩 Step 7: Mini Summary
🧠 What You Learned: How reasoning frameworks like Chain-of-Thought, Tree-of-Thought, and ReAct scaffold structured thinking in LLMs.
⚙️ How It Works: CoT encourages linear reasoning; ToT explores reasoning trees; ReAct lets the model use external tools during reasoning.
🎯 Why It Matters: These frameworks help transform LLMs from text predictors into structured problem-solvers — a crucial step toward reliable reasoning in applied systems.