Some tasks cannot be answered reliably in one jump. Arithmetic, logic, and planning all require state to survive across several decisions. Reasoning techniques give the model or its surrounding system more room to construct, compare, and check those decisions.

The methods form a cost ladder. Chain-of-Thought elicits one path. Self-Consistency samples several paths and aggregates their answers. Tree of Thoughts adds an explicit search procedure with branching and backtracking. None guarantees correct reasoning, and a convincing trace is still only model output.

Chain-of-Thought Prompting

Wei et al. (2022) showed that worked reasoning examples can improve performance on sufficiently large models. Two forms became common:

  1. Few-shot CoT: include worked examples that show a reasoning process and final-answer format.
  2. Zero-shot CoT: append an instruction such as “Let’s think step by step” (Kojima et al. 2022).

An external scratchpad can turn one difficult generation into a sequence of smaller continuations. That often helps on benchmark problems with explicit intermediate state. It also creates more tokens in which an early mistake can spread, so the trace needs verification rather than trust.

Example (same task, with and without CoT):

Question: I buy 10 apples, give away 4, buy 5 more, then eat 1. How many remain?
 
Direct answer (no CoT): 11  (incorrect)
 
CoT-style answer:
Start with 10.
Give away 4 -> 6.
Buy 5 -> 11.
Eat 1 -> 10.
Final answer: 10  (correct)

Visible reasoning also has a product boundary. Some model APIs return concise answers or summaries rather than hidden internal reasoning, and exposing a long scratchpad is rarely the right user interface. Production systems usually need verifiable intermediate results, tool outputs, or a concise rationale.

Self-Consistency

Self-Consistency (Wang et al. 2022) samples several reasoning paths instead of accepting one greedy path. The system groups equivalent final answers and selects the most frequent one.

The method helps when independent samples fail differently and the final answer can be compared mechanically. It is much weaker for open-ended outputs where two correct answers may not match and a majority can repeat the same misconception.

Question: When I was 6, my sister was half my age. Now I'm 70. How old is she?
 
Sampled path 1 -> 67
Sampled path 2 -> 67
Sampled path 3 -> 35
 
Majority vote -> 67 (correct)

Cost and token use grow roughly with the sample count. Calls can run in parallel to reduce wall-clock latency, but they still consume capacity, and the aggregation rule becomes another part of the system to test.

Tree of Thoughts

Tree of Thoughts (ToT), proposed by Yao et al. (2023), wraps the model in a search algorithm. The model proposes candidate states and may help score them. Breadth-first, depth-first, or beam-like search decides which branches continue. Dead ends can be abandoned instead of becoming the rest of one irreversible completion.

flowchart TD
    S[Problem] --> A[Thought A]
    S --> B[Thought B]
    A --> A1[Evaluate A1]
    A --> A2[Evaluate A2]
    B --> B1[Evaluate B1]
    B --> B2[Evaluate B2]
    A2 --> X[Dead end]
    X --> A[Backtrack]
    B1 --> G[Goal reached]

ToT fits problems with meaningful alternatives, delayed consequences, and a state that can be evaluated along the way. Straightforward extraction has no useful search tree. Ordinary arithmetic is usually cheaper to send to a calculator.

Tradeoffs

TechniqueCallsAccuracy profileBest use caseMain downside
Chain-of-Thought1Strong baseline for many reasoning tasksArithmetic, logic, structured step-by-step tasksCan lock into one bad chain
Self-ConsistencyNBetter than single CoT on many verifiable tasksTasks with a clear final answer suitable for votingHigher cost and latency
Tree of ThoughtsMany (branching)Can improve tasks with a useful search stateProblems needing exploration, lookahead, backtrackingMost expensive and operationally complex

Use the cheapest method that clears the evaluation target:

  1. Start with one direct call and an answer format that can be checked.
  2. Add CoT-style decomposition when the task genuinely has intermediate state.
  3. Use self-consistency when final answers can be normalized and voted on.
  4. Use ToT only when the problem contains an actual search space.

Questions

References