LoCA Enables Forward-Only Tuning via One-Shot Calibration
August 5, 2026
LoCA eliminates the need for repeated backward passes during parameter-efficient post-training by using a one-time calibration step. It fits low-rank adapters using closed-form ridge solves from forward activations, matching or exceeding Qwen2.5 performance on five benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can reduce training hardware requirements and memory overhead by avoiding repeated backpropagation.
●
researcherThis provides a method for small-shift adaptation without full backbone gradient computation.