A profile predicts which model abilities weaken first
Language model capabilities fail in a shared, predictable order under compression and perturbation · arXiv:2611.11695
In plain words
A language model can keep a good average score after its abilities have weakened, because some tasks may fail much earlier than others. This study introduces a fragility profile: it measures how quickly different short-answer tasks and answer formats lose accuracy as models are damaged.
Across the tested short-answer tasks, different kinds of damage usually produced similar failure orders. The profile also helped predict accuracy after compression using four calibration tasks, but the tests do not show that the same pattern holds for longer answers or every kind of use.
The research
Average accuracy can hide which particular abilities a compressed model is likely to lose.
Knowing that short-answer tasks tend to weaken in a shared order could make it easier to anticipate those losses, while the GPTQ and long-answer exceptions show why that expectation needs to stay within the tested settings.
The fragility profile measures which short-answer capabilities fail first under damage, then helps forecast how much accuracy remains after compression.
Researchers applied random perturbations and compression methods to the same short-answer tasks, then compared the order in which task accuracy fell.
The orders were generally similar within models, and the profile’s capability-specific estimates predicted compression accuracy better than assuming every capability declined uniformly.
For a forecast of remaining accuracy, four tasks calibrated the amount of damage; all formats of twelve other tasks were held out. Profiles from other model families performed about as well as each target model’s own profile, supporting forecasts in these tested settings.
The shared ordering and accuracy forecasts concern the tested short-answer tasks and models. Long-form task positions did not meet the study’s preregistered threshold for consistency across damage methods.
What the evidence shows
- A shared order of losses
The fragility profile is a method for tracking how quickly short-answer capabilities weaken as damage increases. It addresses a problem with average accuracy: a model can lose only a little overall while particular tasks fail much earlier.
The key question is whether those early failures depend on the damage method, or whether task vulnerabilities have a shared order.
- The deciding comparison
The comparison shows that changing from random perturbations to compression usually left the short-answer task-failure order similar within a model. GPTQ was an exception across models, so the shared order did not reliably transfer between models for that method.
- What predicts early failure
Two properties helped explain which capabilities failed earlier. Tasks with smaller answer margins in intact models tended to fail sooner; across eight models, the correlations ranged from 0.58 to 0.73.
In two recall tasks, more frequent answers in an open text corpus tended to survive damage better. The relationship was much weaker in a key-lookup control, where the answer code was already supplied, linking the frequency pattern to recall in these tasks.
- From ordering to forecasts
Across 23 model-and-compression cells, capability-specific profiles predicted held-out accuracy more accurately than a uniform-loss estimate.
A profile averaged from other model families performed about as well as the target model’s own profile, with forecasts calibrated on four fixed tasks and tested on all formats of twelve other tasks.
- Where the pattern stops
The evidence does not support carrying the shared ordering to every setting. GPTQ’s task order agreed with random damage within models but had a median cross-model correlation of 0.095 across six model pairs.
Long-form tasks also fell short of the study’s preregistered consistency threshold. The findings therefore describe short-answer capabilities under the tested damage methods and do not establish behavior in untested domains.
Deep
- Original paper authors
- Zhixiang Wang, Ulas Bagci
- Submitted to arXiv
- License
- arXiv non-exclusive