Language model capabilities fail in a shared, predictable order under compression and perturbation
Organizations: Department of Radiology, Feinberg School of Medicine, Northwestern University, Chicago, IL, USA
Abstract
Compression can preserve a language model's average accuracy while selectively degrading individual capabilities. Whether these losses follow method-specific patterns or a transferable order is unclear. We applied seven perturbation and compression methods to eight instruction-tuned models, scoring 16 tasks in four answer formats against preregistered predictions. Short-answer capabilities failed in nearly the same order under four random perturbations with different mechanisms (median correlation 0.90) and, within each model, under pruning and quantization (0.875). Intact answer margin, answer frequency and answer format predicted this order. Fragility profiles measured under random perturbation predicted which tasks and facts compression removes first and, after four-task calibration, held-out task accuracy (median absolute error 0.11-0.12, against 0.20-0.21 for uniform decline); profiles borrowed from other model families did almost as well. The order did not transfer across models under GPTQ quantization or hold for long-form tasks. Within short answers, what compression removes is predictable before deployment.
Figures & tables
| Level | Predicted | From | Result | Against | Cells |
|---|---|---|---|---|---|
| 1 | Order of capabilities under a held-out random damage | task and format effects of the other three random damages | task effects alone 0.597; format effects alone 0.512 | better in 20 of 20 | |
| 2 | Order of capabilities under Wanda, SparseGPT and GPTQ | task and format effects of the four random damages | task effects alone 0.525; format effects alone 0.575 | better in 18 of 19 | |
| 3 | Which facts compression removes (capitals, SciQ) | survival of each fact under the four random damages | Spearman 0.69, 0.71, 0.58 (Wanda, SparseGPT, GPTQ) | answer frequency 0.36, 0.37, 0.23; intact item margin 0.09–0.13; key-lookup control 0.12–0.13 | positive in 48 of 48 |
| 4 | Accuracy left in each capability after compression | the model’s own random-damage exponents; strength from four calibration tasks | absolute error 0.110–0.118 | uniform decline 0.200–0.213 | better in 23 of 23; bound of 0.10 not met |
| 5 | The same, without the target model’s own fragilities | exponents of the same task and format in other model families; same calibration | absolute error 0.117–0.125 | uniform 0.200–0.213; intact margins 0.133–0.142; intact accuracy 0.185–0.205; format alone 0.171–0.175 | better than uniform in 23 of 23 |
| Study | Damage | Published ranking | On a common footing |
|---|---|---|---|
| Reverses | |||
| Men et al. 7 | layer removal | summarization (generated) collapses in the 7B models | 13B models: summarization loses 25–39%, BoolQ 55–61% |
| Du et al. 31 | longer context | Mistral-7B: GSM8K loses more raw points than MMLU ( vs ) | corrected: MMLU loses 70%, GSM8K 48% |
| Huang et al. 5 | bit flips | code and BoolQ lose the most raw points ( , ) | corrected: binary multiple choice loses 12–24%, GSM8K 3.6% |
| Sun et al. 6 | hardware faults | generation ( ) worse than option tasks ( ) | on the same items, chain of thought is more robust than direct answers |
| Shrinks or vanishes | |||
| Model | Noise | Zeroing | Neuron | Layers | Wanda | SparseGPT | GPTQ | Joint (separate) |
|---|---|---|---|---|---|---|---|---|
| Qwen2.5-1.5B | 41, 0.95 | 41, 0.94 | 41, 0.94 | 33, 0.84 | 39, 0.85 | 38, 0.78 | 32, 0.85 | 0.82 (0.85) |
| Qwen2.5-7B | 54, 0.94 | 55, 0.92 | 56, 0.93 | 52, 0.86 | 47, 0.90 | 50, 0.89 | 43, 0.76 | 0.78 (0.90) |
| Qwen2.5-14B | 59, 0.92 | 61, 0.87 | 61, 0.89 | 61, 0.74 | 55, 0.88 | 55, 0.84 | 54, 0.70 | 0.74 (0.87) |
| Qwen2.5-32B | 61, 0.91 | 62, 0.91 | 63, 0.91 | 64, 0.81 | 58, 0.91 | 56, 0.90 | 46, 0.91 | 0.74 (0.91) |
| Llama-3.2-3B | 44, 0.90 | 44, 0.88 | 45, 0.90 | 41, 0.79 | 43, 0.84 | 40, 0.77 | 37, 0.77 | 0.72 (0.84) |
| Mistral-7B | 43, 0.93 | 44, 0.89 | 45, 0.92 | 43, 0.77 | 39, 0.83 | 38, 0.90 | 21, 0.93 | 0.77 (0.90) |
| Task | Code | Source | Bare answer | Items | |
|---|---|---|---|---|---|
| Add 2 | ARITH1 | generated | number | 4 | 300 |
| Add 3 | ARITH2 | generated | number | 4 | 300 |
| Add 4 | ARITH3 | generated | number | 4 | 300 |
| Capitals | FACT | generated (78 countries) | city name | 4 | 300 |
| Key lookup | RETR | generated | three-digit code | 4 | 300 |
| Copy word | COPY | generated | nonsense word | 4 | 300 |
Language Model Robustness
Language model robustness studies how reliably language models preserve correct, safe behavior under input changes, distribution shifts, and adversarial pressure.
See all 518 papers