Enabling Dynamic Computation in Looped LMs

Aayush Mishra, Arnau Padrés Masdemont, Victor Conchello Vendrell, Jordi Ros Giralt, Arash Behboodi, Fabio Valerio Massoli

Lets tokens leave a looped language model at different depths by reusing their final key-value cache entries in later loops. Adds an exit-depth training objective and ragged-cache training to make variable-depth computation practical.

First public: October 6, 2026 · Added to this list: October 22, 2026