Preprint halves measurements needed to fit an AI pruning predictor

Researchers report in a preprint that sharing a model's response across pruning levels halves the configuration measurements needed to predict capability losses when removing model parameters. On tested Pythia and OLMo-2 model states and under Wanda pruning, the reduced-measurement predictor matched a regression using all measurements on mathematics and code within 0.020 nats per token, with coefficients refitted for each setting. The study also compared predictions for choosing among compression methods. Across two model families, selection captured most available benefits for question answering within the tested candidates, but a fixed method priority achieved the same regret. Opportunities were smaller for mathematics and code.