Reproduce “SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruizhi Qian, Lei Shi, Tianyu Hu, Bo Yan, Song Wang, Yu Tian
SOMA monitors a reasoning model’s text for specified behaviors, first offering a reminder and taking over only if the behavior persists. For Qwen3-8B, it reduces S2 tokens by 36% while accuracy changes from 91.0% to 91.3%; abstention rises from 70.5% to 87.2%, with false abstentions rising from 2.4% to 8.0%. Tests cover two 8B models and mathematical scenarios, not runtime.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.