LT2

Summary

LT2, short for Linear-Time Looped Transformers, is the architecture family introduced by LT2: Linear-Time Looped Transformers. It replaces the full-attention bottleneck inside looped Transformers with linear, sparse, or hybrid token mixers.

Role In The Wiki

Use this page as the object card for the LT2 method family. The source page carries the evidence details, limitations, and agenda mapping.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in lt2-2026 rather than duplicating verdict rows here. At the entity level, LT2 is relevant as an adjacent long-context and dynamic-compute mechanism: it makes recurrent-depth Transformers cheaper by changing the token mixer, but current evidence is language-modeling and synthetic token-state evidence rather than numeric time-series or action-conditioned world-model evidence.

Evidence

• LT2: Linear-Time Looped Transformers

Official Artifacts

• Preprint: arXiv 2605.20670
• Official code: chili-lab/LT2 — GitHub API reported BSD-3-Clause license at ingest time.
• Official model: chili-lab/Ouro-hybrid-1.4B — research checkpoint distilled from ByteDance/Ouro-1.4B; Hugging Face API reported license: other at ingest time.
• Official X announcement: Chunyuan Deng thread

Related Pages

• Universal Transformers
• Huginn
• LoopFormer
• Parcae
• Sparse Layers are Critical to Scaling Looped Language Models
• Hyperloop Transformers
• Mamba-2
• Mamba-3
• RWKV-TS
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda