0
Act 2

Understanding

6 / 8

Chain-of-Thought

Act 2 · ~4 min

Theory

When a model emits only the final answer, errors from early steps propagate silently. Chain-of-thought (CoT) makes intermediate reasoning explicit in the context window.

Without CoT

Prompt: "4 apples at $0.50 + 2 oranges at $0.75 — total?"

Output: $3.25 ❌

No visible scaffolding. Bad arithmetic step is invisible.

With CoT

Prompt: "...Let's think step by step."

Output:

  1. apples = 4 × $0.50 = $2.00
  2. oranges = 2 × $0.75 = $1.50
  3. total = $3.50 ✓

Each token conditions on correct prior state.

  1. Identify the sub-problems — apples cost, oranges cost, total.
  2. Solve each in order — write the arithmetic for each step.
  3. Aggregate — combine the intermediate results into the final answer.

CoT helps on arithmetic, symbolic, and multi-hop tasks. It adds latency on lookups. Reach for it on complex reasoning, not as a universal fix.