Prune it, quantize it or fill it with noise. What a language model loses first is decided before the damage, by three properties of what you ask it. So it can be forecast before a compressed model is deployed.
Noise, zeroing, neuron ablation and layer removal break a model in very different ways. Pruning and quantization choose what to remove on purpose and could have rewritten the order. They did not: damage sets how fast capabilities fall, not which fall first.
At strength 1.00× the lines are measured: accuracy kept at the dose that halves each model's overall ability, mean of eight models, short-answer format. Other strengths follow the paper's rank-one model, in which a damage turns one dial and each capability fails at its own rate: kept = exp(−n·h). ρ, rank correlation of each damage with the mean order. Median agreement over all model pairs: 0.90 between random damages, 0.875 between compression and random damage.
They act separately and add up: a task effect plus a format effect explains a median of 89% of how fast each capability fails.
Tasks the intact model holds with a smaller margin between the right answer and the best wrong one break first.
Within a task, facts whose answers are rare in pretraining text vanish before common ones. The long tail goes, the core stays.
The same question, answered by picking, by writing the answer, or inside a template. The more the model must produce exactly, the sooner it breaks.
Measure a model's fragility profile once, or borrow it from other model families. After compressing, test four fixed tasks. The profile then predicts the accuracy every other capability keeps.
One example: Qwen2.5-7B after Wanda pruning to 63% sparsity, answer format; calibration on WinoGrande, OpenBookQA, PIQA and copying a word. Across all 23 model and method combinations, the profile had 44–46% less error than a uniform decline in every one (median error 0.110–0.118 vs 0.200–0.213), close to the difference between two runs of the same damage (0.117). All preregistered.
A profile measured once under random damage predicts the order of losses under Wanda, SparseGPT and GPTQ, down to individual facts.
Four calibration tasks give the accuracy every other capability keeps, each held out completely.
A profile borrowed from other model families, excluding the model's own, predicts about as well as the model's own.
In semantic dementia, rare knowledge and naming fail first. Lesioned networks of a few hundred units reproduced this decades ago. We find it in networks of billions of parameters, under random lesions and under compression.
The resemblance concerns what is lost and in what order, not shared mechanisms. It suggests that how robust a capability is depends on what a network has learned rather than on how it is damaged.
Every main test was written into dated protocols before the data it tests were collected.
Analysis scripts were fixed by checksum before they were run.
Including the predictions that failed.
Code, protocols, per-item model outputs and data on GitHub.
@article{wang2026damageorder,
title = {The predictable order in which damaged language models
lose their capabilities},
author = {Wang, Zhixiang and Bagci, Ulas},
year = {2026},
note = {Manuscript}
}