TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning
arXiv:2603.12529v1 Announce Type: cross Abstract: Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage thes
arXiv:2603.12529v1 Announce Type: cross Abstract: Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.
Executive Summary
The article introduces TERMINATOR, an early-exit strategy designed to mitigate overthinking in Large Reasoning Models (LRMs) by determining optimal exit points for Chain-of-Thought (CoT) reasoning. By leveraging the predictability of the first arrival of an LRM's final answer, TERMINATOR achieves significant reductions in CoT lengths across four practical datasets. The approach outperforms current state-of-the-art methods, demonstrating its potential to improve the efficiency of LRMs in complex reasoning tasks.
Key Points
- ▸ TERMINATOR is an early-exit strategy for mitigating overthinking in LRMs
- ▸ The approach leverages the predictability of the first arrival of an LRM's final answer
- ▸ TERMINATOR achieves significant reductions in CoT lengths across four practical datasets
Merits
Efficiency Improvement
TERMINATOR's ability to reduce CoT lengths can lead to significant computational savings and improved efficiency in LRM-based reasoning tasks
Demerits
Task and Model Dependence
The optimal CoT lengths determined by TERMINATOR may be highly task and model-dependent, potentially limiting its generalizability across different scenarios
Expert Commentary
The introduction of TERMINATOR marks a significant step forward in addressing the issue of overthinking in LRM-based reasoning tasks. By providing a data-driven approach to determining optimal exit points, TERMINATOR has the potential to improve the efficiency and effectiveness of CoT reasoning. However, further research is needed to fully explore the generalizability and limitations of this approach, particularly in scenarios where the optimal CoT lengths may be highly task and model-dependent. As the field continues to evolve, it will be essential to consider the implications of TERMINATOR and similar approaches on the development of more explainable and transparent AI models.
Recommendations
- ✓ Further research should be conducted to explore the generalizability of TERMINATOR across different tasks and models
- ✓ The development of TERMINATOR should be accompanied by efforts to improve the explainability and transparency of LRM-based reasoning tasks