Score the Update, Not the Token: Descent-Aligned Routing for Combinatorial 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

Descent-aligned routing over combinatorial LoRA reader-writer pairs

published

2026-09-30

The gains are real but small and demonstrated only on 3B-8B fine-tuning benchmarks, so a constrained team should prototype it against standard LoRA on their own workload before committing to the added routing complexity.

1. claim 01Paperraft inference
   Read the brief interpretation

   VANE replaces token-similarity MoE-LoRA routers with a router that scores each reader-writer LoRA pair by the inner product between its implied weight update and a low-rank predicted descent direction, enabling all N_A x N_B pairs to be ranked from N_A + N_B vectors and giving every pair an exact first-order router gradient. The cost is modest inference-side compute for the compass and pair scoring, implementation complexity well above standard LoRA or off-the-shelf MoE-LoRA, and current validation limited to 3B-8B models on a small set of commonsense and multi-task benchmarks. What can fail: the first-order loss-reduction approximation can mislead under large learning rates or noisy gradients, the compass may mispredict descent directions out of distribution, gains of roughly one accuracy point may not transfer to other domains, and no mature open-source implementation or serving integration is guaranteed.

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

   Inspect the source
2. claim 02Paperraft inference

   Best average among twelve PEFT and MoE-LoRA baselines: +0.9-1.1 points on single-domain commonsense and +1.3-1.5 points on a four-domain multi-task mixture (Llama-3.2-3B, Llama-3.1-8B), with under half the trainable parameters of an 8-expert MoE-LoRA.

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

   Inspect the source