{"id":"arxiv/2603.07955","url":"https://arxiv.org/abs/2603.07955","tags":["math.GT","cs.LG","stat.ML"],"score":0,"title":"RL unknotter, hard unknots and unknotting number","domain":"math.GT","source":"arxiv","authors":["Anne Dranowski","Yura Kabkov","Daniel Tubbenhauer"],"reviews":0,"subject":"math.GT","summary":"We develop a reinforcement learning pipeline for simplifying knot diagrams. A trained agent learns move proposals and a value heuristic for navigating Reidemeister moves. The pipeline applies to arbitrary knots and links; we test it on ``very hard'' unknot diagrams and, using diagram inflation, on $4_1\\#9_{10}$ where we investigate the recently established and surprising upper bound of three for the unknotting number. In addition, we explain a self-improving workbook-driven extension of the pipeline that systematically improves unknotting number upper bounds on the prime knots.","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-03-09","reactions":{}}