CUHK study maps when recurrence helps in looped language models, proposes history-state injection

CUHK-CSE · hf · 2026-10-01

A CUHK study systematically examines when recurrence is effective in looped language models (LoopLMs), which add computational depth via parameter sharing without extra parameters.

Key findings

The paper offers practical design guidelines for LoopLMs under variable inference budgets.

Original post →

More from Models

Models channel →