Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2509.17000 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912596849852416 |
|---|---|
| author | Jiang, Shuhao Wang, Songbo Qiao, Yang Xu, Chun Zheng, Chaoyang Zhou, Shengyi Wang, Huanjun Li, Fangming Zhang, Cong Wang, Jiyu |
| author_facet | Jiang, Shuhao Wang, Songbo Qiao, Yang Xu, Chun Zheng, Chaoyang Zhou, Shengyi Wang, Huanjun Li, Fangming Zhang, Cong Wang, Jiyu |
| contents | Large Reasoning Models (LRMs) often suffer from computational inefficiency due to overthinking, where a fixed reasoning budget fails to match the varying complexity of tasks. To address this issue, we propose Adaptive Overclocking, a method that makes the overclocking hyperparameter $α$ dynamic and context-aware. Our method adjusts reasoning speed in real time through two complementary signals: (1) token-level model uncertainty for fine-grained step-wise control, and (2) input complexity estimation for informed initialization. We implement this approach with three strategies: Uncertainty-Aware Alpha Scheduling (UA-$α$S), Complexity-Guided Alpha Initialization (CG-$α$I), and a Hybrid Adaptive Control (HAC) that combines both. Experiments on GSM8K, MATH, and SVAMP show that HAC achieves superior accuracy-latency trade-offs, reducing unnecessary computation on simple problems while allocating more resources to challenging ones. By mitigating overthinking, Adaptive Overclocking enhances both efficiency and overall reasoning performance. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_17000 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Adaptive Overclocking: Dynamic Control of Thinking Path Length via Real-Time Reasoning Signals Jiang, Shuhao Wang, Songbo Qiao, Yang Xu, Chun Zheng, Chaoyang Zhou, Shengyi Wang, Huanjun Li, Fangming Zhang, Cong Wang, Jiyu Machine Learning Artificial Intelligence Large Reasoning Models (LRMs) often suffer from computational inefficiency due to overthinking, where a fixed reasoning budget fails to match the varying complexity of tasks. To address this issue, we propose Adaptive Overclocking, a method that makes the overclocking hyperparameter $α$ dynamic and context-aware. Our method adjusts reasoning speed in real time through two complementary signals: (1) token-level model uncertainty for fine-grained step-wise control, and (2) input complexity estimation for informed initialization. We implement this approach with three strategies: Uncertainty-Aware Alpha Scheduling (UA-$α$S), Complexity-Guided Alpha Initialization (CG-$α$I), and a Hybrid Adaptive Control (HAC) that combines both. Experiments on GSM8K, MATH, and SVAMP show that HAC achieves superior accuracy-latency trade-offs, reducing unnecessary computation on simple problems while allocating more resources to challenging ones. By mitigating overthinking, Adaptive Overclocking enhances both efficiency and overall reasoning performance. |
| title | Adaptive Overclocking: Dynamic Control of Thinking Path Length via Real-Time Reasoning Signals |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2509.17000 |