🛑 Answer Convergence as a Signal for Early Stopping in Reasoning

Authors: Xin Liu, Lu Wang (University of Michigan)

Code & Resources: GitHub Repository | ArXiv Paper

💡 Demo Description

This interactive demo illustrates the core concept of our Early Stopping strategy.

• Left Panel: Shows the model's full Chain-of-Thought (CoT) reasoning process.
• Right Panel: Shows the reasoning process truncated by our method.

Key Insight: Models often reach Answer Convergence (the correct answer) well before completing the full reasoning chain. Subsequent steps are often redundant self-verification, which can be safely skipped to reduce inference costs.

📊 Key Results (from Paper)

Our experiments across five benchmarks (including NQ, GSM8K, GPQA) reveal substantial redundancy in standard CoT:

• NaturalQuestions (NQ): Token reduction of over 40% with improved accuracy using Learn-to-Stop.
• GSM8K: Token reduction of ~45% with minimal or no accuracy drop.
• Methods: We propose three strategies: Answer Consistency (Unsupervised), Think Token Adjustment (Unsupervised), and Learn-to-Stop (Supervised).