Fine-tune without keeping the backward graph.

Forward-pass model adaptation

Zeroth-order optimizers estimate gradients from loss differences under parameter noise, so adaptation needs only forward passes and inference-level memory — no backprop activation store.

Approach

LOZO and related methods use low-rank perturbation structure to keep the search tractable on-device. We’re applying that family so models can keep improving as they are used, without datacenter-sized fine-tuning loops.

The practical bar is simple: if inference fits, adaptation should too.

What we’re measuring

Sample efficiency versus quality under tight memory budgets.

Stability when the same device both serves and adapts.