Reproduce “Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language 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
Xiao-Wen Yang, Ziyu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, Yu-Feng Li
Researchers at Nanjing University developed the STARS framework to enable reliable test-time scaling in Looped Language Models by stabilizing their recurrent latent dynamics. This method, which integrates Jacobian Spectral Radius Regularization with random loop sampling, prevented performance collapse and enhanced reasoning quality across various mathematical tasks, achieving consistent 100% accuracy on multi-digit addition.
Key insights
- Normalization layer placement (internal vs. external) dictates a fundamental trade-off in LoopLMs between maintaining effective information flow for deep reasoning and ensuring stable, bounded latent states.
- Standard remedies like Prelude/Coda layers, random loop sampling, and L2 regularization largely fail to address the core stability issues that cause performance collapse in recurrent LLMs.
- Effective latent reasoning in iterative systems corresponds to the convergence of hidden states towards stable and effective fixed points in a dynamical system.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.