Papers worth knowing this week

Multi-Token Prediction Done Right:Soft-MTP removes a noisy gradient component from multi-token-prediction training with a one-line change, improving perplexity and speculative decoding from 167M to 1.2B parameters at no extra cost, which makes it the same kind of cheap pretraining win as MSAR. Recurrent Looped Transformer:it uses a recurrent decoder with feedback to grow computation depth per token, so it is a different answer to the depth-scaling question that MSAR addresses through residual retrieval.