This year, the loop variables have become common place, so much work has appeared in them, especially in the last six months, that you can write about it continuously several times a day. Here is another paper that tried different looping strategies, and interestingly, found that they apply differently depending on the constraints (tables or parameters). The theme of LoRA is also fun. 149 pages is absolutely a mockery. As if LLM is really written for LLM, manual reading is no longer possible. How to Loop Your Transformer: Loop Scaling, Specialization, and Parallel Inference. the layer) only for Backdladgondgondladladladladladladgondela: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: What: The Writers (Backdators prefer dense medium sub-bass) to Backers (Back they also suggested adapters Common Loop (Loop Matrix) and Normalized Shifted PLT, a variant of parallel Loop Transformers (where loop states switch neighboring tokens for parallel encoding), where the current token embedding and the state of the former are separately normed using RMS Norm before mixing. What you have: with a fixed number of parameters, a full overhaul works better, and with a limitation on FLOP - Middle-Stack Looping. In DCLM with 200M parameters (40 tokens per parameter, TPP), the 4-recycled normal Shifted PLT reduces bits per BPB (BPB) to 1.060 compared to -1.087 for the original PLT and 1.088 for a basis without cycles. In 1B parameters, the two cycles recycled PLT improves the BPB (-0.86 vs. -0.90 in 80 TPP) and accuracy on 8 tasks (~47.5% vs. -46.1% in 160 TPP).