How Many Ranks Does a Token Need? A Noise-Floor Squelch for Rank-One LoRA Experts

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

Router-free per-token LoRA expert selection (SQUELCH)

published

2026-10-13

The method is compatible with a single 24 GB GPU since it adds no trainable parameters and builds on standard LoRA fine-tuning, but as a single recent paper its per-token adaptive routing gains require validation on the reader's own tasks and serving latency constraints before production adoption.

1. claim 01Paperraft inference
   Read the brief interpretation

   Replaces the fixed-rank LoRA update and learned top-k routers of mixture-of-LoRA methods with a router-free soft-thresholding of each token's rank-one responses, with the threshold chosen per module by Stein's unbiased risk estimate and a Gram-matrix explaining-away step, adding no trainable parameters. Cost is an extra per-token response computation, robust noise estimation, and a value Gram solve at inference, plus implementation complexity beyond a standard LoRA stack; training cost is comparable to LoRA since thresholds are set post hoc per module. It can fail if the Gaussian noise-floor assumption does not hold for a given adapter or domain, if coherence among co-active components is high enough to degrade the explaining-away step, or if the reported gains do not transfer beyond the evaluated models and benchmarks.

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

   Inspect the source
2. claim 02Paperraft inference

   Improves over LoRA by 2.2-3.2 points and over the strongest mixture-of-LoRA baseline by 0.8-1.2 points on commonsense, knowledge, math, code, and GLUE benchmarks, activating 8.6 of 32 rank-one components per token on average.

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

   Inspect the source