Reproduce “Asymmetric Looped Transformers: Fixed Global Representations for an Iterative Decoder” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cheng Liang, Tengfei Zhang, Wenjie DU, Jianyu Wu, Changlian Ma, Shuchen Pu, Ya Zhang, Yanfeng Wang and 1 more
AsymLoop separates one-pass encoding from recurrent decoding in a looped Transformer. A causal encoder supplies fixed Global key-value representations, while a weight-tied decoder repeatedly updates its state using attention to that context. At 1B scale, it reaches 50.85% seven-task zero-shot average accuracy versus 49.58% for Full Loop at matched effective depth and training-token budget, though not on every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.