MuPACE: Learning to Pace Muon for LoRA Fine-Tuning

You already need task-specific training and want to reduce its memory or compute cost.

What you would change

Change only the training technique on a small controlled run after checking prerequisites.

Change only the training technique on a small controlled run after checking prerequisites.

Compare against

Your current training recipe with the same model, data split, and evaluation.

1. Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.
2. Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.
3. Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.
4. Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test.

What to measure

Held-out quality, peak memory, training time, and full compute cost to the target quality.

When to keep it

Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation.

recommendation

test

research area

efficient training

technical change

Meta-learned learning rate pacing for Muon+LoRA

published

2026-10-20

Eliminating per-task LR sweeps directly saves GPU-hours on a constrained budget, but the method requires either a released pretrained policy or an amortized meta-training investment whose payoff depends on how often the team fine-tunes new tasks.

1. claim 01Paperraft inference
   Read the brief interpretation

   MuPACE replaces per-problem learning rate schedule tuning for LoRA fine-tuning with a small recurrent policy that adjusts the step size of an otherwise untouched Muon optimizer on top of an interpretable analytic schedule. The cost is the meta-training phase across a distribution of fine-tuning runs plus an extra recurrent controller in the training loop, though inference and the deployed adapter are unaffected. It can fail when the production task or model distribution differs sharply from the meta-training distribution, when the meta-training budget is unavailable to a small team, or when the claimed transfer does not hold for non-Qwen architectures or non-LoRA configurations.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source
2. claim 02Paperraft inference

   Outperforms strong shared-default fixed schedules on 64-78% of 64 fine-tuning problems and matches per-task tuned-peak schedules, with one-shot transfer to unseen tasks on models up to 64 layers (Qwen3-32B, Qwen3.8-27B).

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source