ReBalance: Efficient Reasoning with Balanced Thinking (2026)
FreeDetects overthinking/underthinking via confidence variance and applies steering vectors to redirect reasoning — ICLR 2026; works on DeepSeek-R1, QwQ, o3-class models
About ReBalance: Efficient Reasoning with Balanced Thinking (2026)
ReBalance is a training-free framework for large reasoning models that detects and mitigates overthinking (redundant computational steps on simple problems) and underthinking (insufficient exploration of reasoning paths) by using confidence variance as a continuous indicator. It aggregates hidden states from a small-scale dataset into reasoning mode prototypes and applies a steering vector with a dynamic control function that modulates its strength and direction based on real-time confidence. This allows the framework to prune redundancy during overthinking and promote exploration during underthinking, improving both efficiency and accuracy. Accepted at ICLR 2026, ReBalance is demonstrated on four models ranging from 0.5B to 32B parameters across math reasoning, general question answering, and coding benchmarks, offering a plug-and-play strategy for efficient LRM deployment.
Key Features
Pros & Cons
- Training-free framework requiring no fine-tuning
- Improves reasoning efficiency by reducing redundant computation
- Enhances accuracy by balancing exploration and exploitation
- Plug-and-play integration with existing large reasoning models
- General strategy validated across multiple benchmarks and model sizes
- Requires a small-scale dataset for constructing reasoning mode prototypes
- May introduce additional computational overhead for steering vector computation
- Effective only on models with accessible hidden states