README.md

metadata

title: Answer Convergence Early Stopping
emoji: 🛑
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: 5.9.0
app_file: app.py
pinned: false
license: mit
short_description: Demo for EMNLP Paper "Answer Convergence as a Signal..."

🛑 Answer Convergence as a Signal for Early Stopping in Reasoning

[EMNLP Accepted] | Paper (arXiv) | GitHub Code

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

---

💡 What is this?

This Space demonstrates the core concept of our paper: Large Language Models often internally converge on an answer long before they finish generating the full reasoning chain.

By detecting this Answer Convergence, we can stop the generation early, saving 40%+ of inference costs without sacrificing accuracy.

🚀 Key Methods

We propose three strategies to detect this signal:

1. Answer Consistency: Unsupervised method checking if the answer stabilizes across reasoning chunks.
2. Think Token Adjustment: Encouraging the model to output the stop signal earlier.
3. Learn-to-Stop: A lightweight supervised module trained on internal activations.