Machine Learning Submissions Surge 168% in a Week
The section logged 1,475 new papers against a usual 550, outpacing every other corner of arXiv.
Machine Learning (cs.LG) absorbed 1,475 new papers over the past week, well over double the usual 550, and no other field came close to that swing. Among the arrivals: 'Decoding Looped Transformers Better for (Almost) Free', which promises efficiency gains without the usual asterisks, and 'On Language Drift during RLVR Post-Training', a reminder that reinforcement learning from verifiable rewards keeps finding new ways to misbehave.
The counters are lighting up around 'looped', 'rlvr', 'proposer' and 'teachers', suggesting a cohort of papers all circling the same ideas from different angles. Even the audio people have caught the looping bug: 'LAST: Looped Audio Spectrogram Transformer' turned up filed under Sound.
Whether this is deadline pressure, a genuine shift in how the field iterates on itself, or both, the desk isn't saying. It rarely does.