Beyond the Last Answer: Your Reasoning Trace Uncovers More than You Think

TL;DR

This work addresses the limitation of evaluating LLM reasoning solely by the final answer. It introduces Subthought Reasoning, a test-time framework that segments an initial reasoning trace into subthoughts, generates completions from intermediate states, and aggregates the resulting answers via the mode to improve accuracy on challenging problems like the AIME datasets. Across seven open-weight models and both AIME2024/2025, mode-based aggregation yields substantial gains (up to around 13% on AIME2024 and 10% on AIME2025) and reveals that lower answer-distribution entropy correlates with correct solutions. The findings suggest practical benefits for reliability estimation and decoding strategies, and point to non-greedy exploration as a robust avenue for leveraging internal reasoning dynamics to obtain more trustworthy results.

Abstract

Large Language Models (LLMs) leverage step-by-step reasoning to solve complex problems. Standard evaluation practice involves generating a complete reasoning trace and assessing the correctness of the final answer presented at its conclusion. In this paper, we challenge the reliance on the final answer by posing the following two questions: Does the final answer reliably represent the model's optimal conclusion? Can alternative reasoning paths yield different results? To answer these questions, we analyze intermediate reasoning steps, termed subthoughts, and propose a method based on our findings. Our approach involves segmenting a reasoning trace into sequential subthoughts based on linguistic cues. We start by prompting the model to generate continuations from the end-point of each intermediate subthought. We extract a potential answer from every completed continuation originating from different subthoughts. We find that aggregating these answers by selecting the most frequent one (the mode) often yields significantly higher accuracy compared to relying solely on the answer derived from the original complete trace. Analyzing the consistency among the answers derived from different subthoughts reveals characteristics that correlate with the model's confidence and correctness, suggesting potential for identifying less reliable answers. Our experiments across various LLMs and challenging mathematical reasoning datasets (AIME2024 and AIME2025) show consistent accuracy improvements, with gains reaching up to 13\% and 10\% respectively. Implementation is available at: https://github.com/hammoudhasan/SubthoughtReasoner.

Figures (6)

• Figure 1: Subthought Analysis. We show that by examining intermediate reasoning steps and their corresponding answers ($A_1, \dots, A_5$), taking the mode of these answers ($A_{mode}$) often leads to better performance than using only the final answer ($A_{last}$), as is typically done. This figure illustrates a case where $A_{mode}=96$ is correct, while $A_{last}=50$ is not.
• Figure 2: Answer Evolution Across Models on AIME2024 (Greedy Completion). Each row corresponds to a different model.
• Figure 3: Mean Entropy of Subthought Answer Distributions on AIME2024 (Greedy Completion). Comparison between problems answered correctly ($A_{last} = A_{true}$) and incorrectly ($A_{last} \ne A_{true}$). Lower entropy correlates with correct answers.
• Figure 4: Accuracy Comparison: Last Answer vs. Most Frequent Answer. Comparison of Last Answer Accuracy ($\text{Acc}_{Last}$, blue) with Most Frequent Answer Accuracy ($\text{Acc}_{MostFreq}$, orange) using Greedy and Non-Greedy subthought completions across various models and AIME datasets. Numbers above bars show the absolute gain ($\text{Acc}_{MostFreq} - \text{Acc}_{Last}$). Green upward triangles indicate improvement, red downward triangles indicate decrease. Our method consistently improves or matches baseline accuracy.
• Figure 5: Mean Entropy of Subthought Answer Distributions on AIME2024 (Greedy Completion). Comparison between problems answered correctly ($A_{last} = A_{true}$) and incorrectly ($A_{last} \ne A_{true}$). Lower entropy correlates with correct answers.