Tag
This paper systematically studies when recurrence helps in looped language models (LoopLMs), finding that extra recurrence can improve reasoning beyond the training horizon but degrade knowledge retention, and proposes channel-wise history-state injection with timestep conditioning as a more robust design for variable inference budgets.
The Stream Recursion Model (SRM) is a modification of the Hierarchical Reasoning Model that organizes computation into recursive latent streams to improve mechanistic interpretability for large language models, achieving performance comparable to GPT-2 per parameter.