Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models

TL;DR

Extending reasoning traces at test-time often leads to overthinking, which increases output variance and can decrease accuracy, but a parallel thinking approach can improve performance by selecting the most consistent response.

Contribution

The paper demonstrates that test-time extended thinking can be counterproductive and introduces a parallel thinking method that improves reasoning accuracy through majority voting.

Findings

Extended thinking initially improves performance but then declines due to overthinking.

Additional thinking increases output variance, creating an illusion of better reasoning.

Parallel thinking with majority voting outperforms extended thinking, achieving up to 20% higher accuracy.

Abstract

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like "Wait" or "Let me rethink" can improve performance. This raises a natural question: Does thinking more at test-time truly lead to better reasoning? To answer this question, we perform a detailed empirical study across models and benchmarks, which reveals a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to "overthinking". To understand this non-monotonic trend, we consider a simple probabilistic model, which reveals that additional thinking increases output variance-creating an illusion of improved reasoning while ultimately undermining precision. Thus, observed gains from "more thinking" are not true indicators of improved reasoning, but artifacts stemming from the connection between model uncertainty and evaluation metric. This suggests that test-time scaling through extended thinking is not an effective way to utilize the inference thinking budget. Recognizing these limitations, we introduce an alternative test-time scaling approach, parallel thinking, inspired by Best-of-N sampling. Our method generates multiple independent reasoning paths within the same inference budget and selects the most consistent response via majority vote, achieving up to 20% higher accuracy compared to extended thinking. This provides a simple yet effective mechanism for test-time scaling of reasoning models.