Model soups: averaging independently fine-tuned models from a common initialization

What changed

Studied uniform and validation-greedy weight averaging across independently fine-tuned models from a shared pretrained initialization, retaining a single-model inference footprint.

Scientific scope

Studied uniform and validation-greedy weight averaging across independently fine-tuned models from a shared pretrained initialization, retaining a single-model inference footprint. Inherits weight averaging and low-error-basin observations. Fine-tuning sweeps and validation selection still cost compute; no extra inference cost does not mean free training. No guarantee for arbitrary independently initialized models, architectures, basins or every downstream task.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits weight averaging and low-error-basin observations. Fine-tuning sweeps and validation selection still cost compute; no extra inference cost does not mean free training. No guarantee for arbitrary independently initialized models, architectures, basins or every downstream task.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction