Gradient sliding is a technique for composite optimization problems — minimizing f(x) + g(x) where f is smooth and g is structured (a sum of many components, or a regularizer). The “sliding” refers to efficiently reusing gradient information: rather than computing a full gradient of g at every step, you slide along the structure, computing only the minimum necessary.
The authors (arXiv:2603.23492) make the method parameter-free. Standard gradient sliding requires knowing the smoothness constant of f and the structure of g to set step sizes. These parameters are often unknown in practice and estimated conservatively, leading to slower convergence. The universal version adapts to the problem's actual difficulty by adjusting step sizes online.
The method achieves optimal oracle complexity — the theoretical minimum number of gradient evaluations needed to reach a given accuracy — without any tuning. No grid search over learning rates, no warmup schedule, no problem-specific calibration.
The through-claim: the parameter tuning in first-order optimization isn't necessary information for the algorithm — it's a design limitation. The optimal step size at each iteration depends on local curvature that the algorithm already measures (through its gradient evaluations). The “universal” version uses information it already has, rather than requiring the user to supply information the algorithm could infer.