Verify scalability and task generalization of the learning‑rate‑tuned LoRA findings

Establish whether the reported finding that vanilla LoRA and advanced LoRA variants achieve similar peak performance after proper learning rate tuning extends to larger foundation models beyond 7B parameters and generalizes to downstream tasks beyond mathematical reasoning and code generation.

Background

The study systematically tunes learning rates for vanilla LoRA and several variants on decoder‑only models up to 7B parameters and two task families (mathematical reasoning and code generation), concluding that once learning rates are properly tuned, methods reach similar peak performance with rank‑dependent nuances.

The authors explicitly acknowledge computational limitations and state that they have not verified whether these findings hold for larger models and broader task suites, identifying this as an unresolved question.