ReBalance: Efficient Reasoning with Balanced Thinking (2026) logo

ReBalance: Efficient Reasoning with Balanced Thinking (2026)

Free

Detects overthinking/underthinking via confidence variance and applies steering vectors to redirect reasoning — ICLR 2026; works on DeepSeek-R1, QwQ, o3-class models

FreeFree tier
Inputs: textOutputs: text
Type
Open Source

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

Detects overthinking through high confidence variance
Detects underthinking via consistent overconfidence
Aggregates hidden states into reasoning mode prototypes
Applies steering vectors to guide reasoning trajectories
Dynamic control function modulates steering vector strength and direction based on real-time confidence
Training-free and plug-and-play integration
Reduces output redundancy while improving accuracy
Works on models ranging from 0.5B to 32B parameters

Pros & Cons

Pros
  • 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
Cons
  • 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

Best For

Math reasoningGeneral question answeringCoding tasks

FAQ

What is ReBalance?
ReBalance is a training-free framework for efficient reasoning in large reasoning models, designed to detect and mitigate overthinking and underthinking by using confidence variance as an indicator and applying steering vectors.
How does ReBalance work?
It uses confidence variance to identify overthinking (high variance) and underthinking (consistent overconfidence), aggregates hidden states from a small dataset into reasoning mode prototypes, then applies a steering vector with a dynamic control function to adjust reasoning depth in real-time.
Is ReBalance training-free?
Yes, ReBalance is a training-free framework that does not require fine-tuning of the underlying model. It only uses a small-scale dataset to construct reasoning mode prototypes and then applies steering vectors during inference.
What models does ReBalance support?
ReBalance has been tested on four models ranging from 0.5B to 32B parameters, including variants in math, QA, and coding domains. It is designed as a plug-and-play strategy and is expected to work with any autoregressive large reasoning model that provides access to hidden states.
How is ReBalance evaluated?
Extensive experiments are conducted on nine benchmarks covering math reasoning, general question answering, and coding tasks, demonstrating reduced output redundancy and improved accuracy.