Column generation folklore holds that initial columns hurt: a warm start may bias the dual and slow the root of branch-price-and-cut. A new arXiv paper (2610.21001) argues the opposite, but with a crucial caveat—the columns must be imperfect in the right way.
The authors show that preloading the support of the column set can be beneficial, rather than detrimental. The key is not avoiding warm starts, but choosing the right kind of imperfection in the initial columns. This reframes a long-standing assumption in the field.
The paper introduces a learned column prefilling method, suggesting that machine learning can guide which imperfect columns to preload. If the approach holds up, it could turn a known pitfall into a practical speed-up for branch-price-and-cut solvers.