ARLoop: fully-open 1.2B looped transformer beats Huginn-3.5B and Qwen3.5-2B on all 8 benchmarks
hhsun1 · x · 2026-11-10
A team led by Yuekun Yao (with OSU NLP) releases ARLoop, the strongest fully-open pretrained looped LLM family. The key idea: attention residuals to selectively reuse information across repeated depth loops.
- ARLoop-1.2B, pretrained on just 0.5T tokens, beats both Huginn-3.5B (prior best fully-open looped LLM) and Qwen3.5-2B Base on all eight benchmarks spanning commonsense reasoning, math, and coding
- Builds on the team's prior "Loop, Think, & Generalize" work showing recurrent depth boosts multi-hop generalization
- Everything is open: weights, training code, data processing, intermediate checkpoints, and eval code
A complete, reproducible recipe for a direction where training recipes were previously hard to find.
More from Models
- Argon Model Identifies a Russian Accent From a 3-Minute Clip With Phoneme-Level Analysis — soumitrashukla9 · 2026-11-10
- Claude Fable 5.2 tops VIALS-AA science visual reasoning benchmark at 67% with heavy tool use — ArtificialAnlys · 2026-11-10
- Dev builds a mobile game with Fable 5.2 as a Claude Artifact in under an hour — ChrisGPT · 2026-11-10
- SpaceX says Grok detects aerospace foreign object debris with 99.9% accuracy vs humans' 94% — elonmusk · 2026-11-10
- Fable 5.2's first ASCII backrooms art is so good people suspect it's programmatic — repligate · 2026-11-10
- Community pushes AI labs to publish system cards as machine-readable jsonl data — scaling01 · 2026-11-10