Principle:Iamhankai Forest of Thought Majority Vote Consensus

Knowledge Sources	Self-Consistency Forest-of-Thought
Domains	Ensemble_Methods, Decision_Theory
Last Updated	2026-02-14 03:00 GMT

Overview

An aggregation strategy that selects the most frequently occurring answer across multiple independent reasoning paths to improve prediction reliability.

Description

Majority Vote Consensus applies the wisdom-of-crowds principle to LLM reasoning. When multiple reasoning trees in a forest produce candidate answers, the most common answer is selected as the final prediction. This approach is robust to individual tree failures because errors are unlikely to consistently produce the same wrong answer across independent paths.

In FoT, majority voting serves two roles:

• Early stopping: If a strict majority (> N/2) of trees agree before all trees finish, the forest stops early to save computation
• First-pass selection: In CGDM post-processing, majority voting is the primary answer selection mechanism, with ties escalated to an LLM judge

Usage

Majority voting is used implicitly within FoT evaluation whenever multiple trees produce answers. It is the primary mechanism in the "majority" stopping strategy and the first pass in CGDM.

Theoretical Basis

Based on the Condorcet Jury Theorem: if each independent voter (tree) has probability p > 0.5 of being correct, the majority vote accuracy approaches 1 as the number of voters increases:

$P(\text{majority correct})=\sum _{k=\lceil N/2\rceil }^N(\genfrac{}{}{0ex}{}{N}{k})p^k(1-p{)}^{N-k}$

Early stopping condition: $\text{stop if }\underset{a}{\max }\text{count}(a)>\frac{N_{\text{trees}}}{2}$

This is equivalent to self-consistency decoding applied to tree-level outputs rather than token-level outputs.

Related Pages

Implemented By

• Implementation:Iamhankai_Forest_of_Thought_Most_Frequent_Elements

Uses Heuristic

• Heuristic:Iamhankai_Forest_of_Thought_Early_Stop_Majority_Vote