ROM: Real-time Overthinking Mitigation

This dataset contains the Counterfactual Self-Correction (CSC) training data used in the paper ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention. It consists of 1,533 samples (740 efficient + 793 overthinking) derived from MATH500, used to train a lightweight hidden-state detector that identifies when a large reasoning model has reached the first correct solution and should stop generating.

Links

• Paper: arXiv | Hugging Face Paper
• Code: GitHub Repository
• Project Page: https://xinyan-wang-stat.github.io/ROM-LRM/

Usage

Download the dataset using the Hugging Face CLI:

huggingface-cli download xinyan-wang/ROM --repo-type dataset --local-dir data

Place the data under the data/ directory of the ROM codebase and follow the training instructions in the repository.