Efficient Reasoning with Balanced Thinking

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Large Reasoning Models (LRMs) often suffer from overthinking, expending redundant computational steps on simple problems.limitation 90%
  Evidencepartial

  they often suffer from overthinking, expending redundant computational steps on simple problems

  Implicationpartial

  Directly stated in the abstract as a core problem addressed by the paper

  Verificationpartial

  partial

• Large Reasoning Models (LRMs) often suffer from underthinking, failing to explore sufficient reasoning paths despite inherent capabilities.limitation 90%
  Evidencepartial

  or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities

  Implicationpartial

  Directly stated in the abstract as a core problem addressed by the paper

  Verificationpartial

  partial

• Existing methods to mitigate overthinking may inadvertently induce underthinking, compromising accuracy.limitation 85%
  Evidencepartial

  Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy

  Implicationpartial

  Directly stated in the abstract as motivation for the proposed method

  Verificationpartial

  partial

• ReBalance is a training-free framework that achieves efficient reasoning with balanced thinking.method 95%
  Evidencepartial

  we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking

  Implicationpartial

  Directly stated as the main contribution of the paper in the abstract

  Verificationpartial

  partial

• ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence.method 90%
  Evidencepartial

  ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence

  Implicationpartial

  Directly stated in the abstract describing the core mechanism of the method

  Verificationpartial

  partial

• ReBalance effectively reduces output redundancy while improving accuracy across multiple benchmarks.result 85%
  Evidencepartial

  ReBalance effectively reduces output redundancy while improving accuracy

  Implicationpartial

  Directly stated in the abstract with reference to extensive experiments

  Verificationpartial

  partial

• ReBalance was evaluated on four models ranging from 0.5B to 32B parameters and across nine benchmarks in math reasoning, general question answering, and coding tasks.result 95%
  Evidencepartial

  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

  Implicationpartial

  Directly stated in the abstract with specific details about the experimental setup

  Verificationpartial

  partial

• ReBalance offers a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment.technical 90%
  Evidencepartial

  offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment

  Implicationpartial

  Directly stated in the abstract as a key advantage of the proposed method

  Verificationpartial

  partial

Startup potential card