ARLoop: Fully Open-Source Looped Transformers, 1.2B Beats Huginn-3.5B

A UPenn team (Yuekun Yang et al., with OSU NLP involvement) has released ARLoop: the strongest fully open-source pretrained looped LLM family to date, along with all resources needed to replicate the research, including the paper, code, training recipe, and checkpoints.

Confirmed

• ARLoop is built on the Attention Residuals architecture, with the core research question being "how to help looped Transformers make better use of repeated computation"
• The strongest model, ARLoop-1.2B, outperforms both the previous state-of-the-art fully open-source looped model Huginn-3.5B and Qwen3.5-2B / Qwen 2B Base across commonsense reasoning, math, and coding benchmarks
• The team has open-sourced a complete training recipe covering everything from controlled experiments to full LLM pretraining
• The 11-11 team announced further release of the full replication package: paper, code, and checkpoints

Why it matters

• Looped Transformers boost model capability through repeated computation; ARLoop sets the strongest and fully open-source pretraining baseline in this direction, surpassing the larger Huginn-3.5B and demonstrating the payoff of thoughtful architecture design
• The lab's fully open-source approach (paper + code + recipe + checkpoints) lowers the barrier for follow-up research, replication, and improvement

2026-11-10 ~ 2026-11-11 · 5 related posts

• [source] ARLoop: fully-open 1.2B looped transformer beats Huginn-3.5B and Qwen3.5-2B on all 8 benchmarks — hhsun1 · 2026-11-10
• [source] ARLoop team releases full artifacts to reproduce looped-transformer LLM pretraining: paper, code, checkpoints — hhsun1 · 2026-11-11

3 near-duplicate retellings: hhsun1 · burny_tech · Lianhuiq