Paper 2603.12372
Efficient Reasoning with Balanced Thinking
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 3
- GitHub
- 333 stars
01 In brief
Summary
Large Reasoning Models (LRMs) often overthink (redundant steps on simple problems) or underthink (insufficient exploration), causing inefficiency and inaccuracy.
Existing mitigation methods, like suppressing reflection keywords, can induce underthinking.
The paper proposes REBALANCE, a training-free framework to balance these extremes.
It uses stepwise confidence and confidence variance as indicators: high variance signals overthinking, while consistent overconfidence signals underthinking.
REBALANCE extracts a steering vector from hidden states of overthinking and underthinking prototypes, computed offline from a small dataset.
A dynamic control function adjusts the steering vector's strength and direction in real-time based on confidence, pruning redundancy during overthinking and promoting exploration during underthinking.
Experiments on four models (0.5B-32B) and nine benchmarks (math, QA, coding) show REBALANCE reduces output length by up to 35.4% while improving accuracy by up to 10 points, outperforming baselines.
It is a plug-and-play solution that generalizes across domains and models, including NPU devices, without additional training or inference stages.
02 From the paper
Abstract
Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Project page and code are available at https://rebalance-ai.github.io .