Marginal utility, matrix factorization, and the Key-Value (KV) cache: a unified information-economic framework for sovereign geo-mining inference

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

Eigenvalue-gated small classifier and calibrated TIES merging

published

2026-09-17

The latency and cost gains are real but demonstrated only on a narrow uranium-mining corpus with the main extraction benchmark merely projected, so the reader should validate the small-classifier-plus-calibrated-merging recipe on their own workload before production adoption.

1. claim 01Paperraft inference
   Read the brief interpretation

   Replaces per-document calls to a proprietary extraction API with a small locally trained hierarchical classifier, and replaces uniform-density TIES model merging with layer-wise calibrated densities. Costs are modest (about five minutes of single-GPU training, 11.2M parameters), but the approach requires a labeled domain corpus and the theoretical marginal-utility framework adds complexity without a measured full-scale extraction benchmark. A documented failure mode is degenerate merging: uniform-density TIES produced token-identical, high-confidence outputs across five distinct districts, and the fix is validated only on a diagnostic sample.

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

   Inspect the source
2. claim 02Paperraft inference

   An 11.2M-parameter hierarchical classifier runs at 2.62 ms per document card versus approximately 2,000 ms for a proprietary API, with 90.0% level-1 accuracy on a 973-document test set against 92.0% for the API on a 50-document human audit; the full extraction benchmark is reported as projected, not measured.

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

   Inspect the source