First Try Matters: Revisiting the Role of Reflection in Reasoning Models

TL;DR

This paper investigates the role of reflections in reasoning models by analyzing eight models across five mathematical benchmarks to determine whether post-answer reflections improve accuracy. It introduces an LLM-based candidate-answer extractor to separate forward reasoning from reflective steps, revealing that reflections are predominantly confirmatory and contribute little to correcting initial mistakes. Training with reflection-rich data improves first-answer correctness rather than correcting errors, suggesting reflections mainly diversify problem-solving exposure and improve generalization. To enhance inference efficiency, the authors propose a question-aware early-stopping framework that uses a candidate-answer detector and a question-aware reflection controller, achieving significant token reductions (up to ~29.9%) with small accuracy penalties (~3-4%). Overall, the work provides practical guidance for data design and inference strategies in reasoning models, highlighting that “the first try matters” while offering efficient mechanisms to curb unnecessary reasoning.

Abstract

Large language models have recently demonstrated significant gains in reasoning ability, often attributed to their capacity to generate longer chains of thought and engage in reflective reasoning. However, the contribution of reflections to performance improvement remains unclear. In this paper, we systematically analyze the rollouts of eight reasoning models on five mathematical datasets. We focus on reflective behaviours where the model has already produced an answer but continues reflecting before finalizing its output. Our analysis reveals that reflections are predominantly confirmatory and rarely alter the model's initial answer, a pattern consistent across models and datasets. To understand the role of reflections in training, we construct supervised fine-tuning (SFT) datasets with varying amounts of reflection steps. We observe that training models on rollouts with more reflection steps primarily enhances first-answer correctness rather than the ability to correct initially wrong answers through reflections. This motivates us to propose a question-aware early-stopping method that enhances inference-time token efficiency by stopping the reasoning process once a few plausible candidate answers are generated, thereby reducing unnecessary reflection steps. Motivated by this, we further propose to dynamically truncate the reflections after a candidate answer has appeared during generation, which reduces reasoning tokens by 24.5% across five mathematical datasets, within a 2.9% drop in accuracy.

Figures (13)

• Figure 1: Illustration of a long CoT and the extraction result of candidate answers.
• Figure 2: Distribution of first candidate answer positions across different LLMs and prompts. The x-axis denotes the relative position of the first candidate answer (line index divided by total lines), and the y-axis shows the proportion of rollouts in each bin.
• Figure 3: Reflections type statistics of long CoTs of different models. Long CoTs are collected on AIME2024 and AIME2025 (32 rollouts per question), AMC (4 rollouts per question), Olympiad Bench, and Math500 (1 rollout per question). Statistics are compiled for the union of all rollouts. More detailed breakdown of each dataset can be found in Figure \ref{['fig:transitions_stacked_bar_full']} of Appendix \ref{['sec:breakdown_trans']}.
• Figure 4: Breakdown of long CoTs: orange bars show the token count up to the first candidate answer, and blue bars show the token count in subsequent reflections. Numbers on bars indicate the accuracy of the first candidate answer, and the accuracy improvement brought by reflections.
• Figure 5: Left: Average number of candidate answers per rollout across different datasets. Right: Relative position of the first candidate. Values are averaged over 8 models.